You are on page 1of 123

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/299560759

Maintenance Management of Wind Power Systems by means of Reliability-


Centred Maintenance and Condition Monitoring Systems

Technical Report · February 2012


DOI: 10.13140/RG.2.1.1874.3446

CITATIONS READS

3 558

1 author:

Katharina Fischer
Fraunhofer Institute for Wind Energy Systems IWES
23 PUBLICATIONS   387 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Investigation of converter failure in wind turbines (CONFAIL) View project

Reliable Power Electronics for Wind Turbines View project

All content following this page was uploaded by Katharina Fischer on 01 April 2016.

The user has requested enhancement of the downloaded file.


Maintenance Management of Wind Power Systems by
means of Reliability-Centred Maintenance
and Condition Monitoring Systems

Final project report


February 2012

Dr.-Ing. Katharina Fischer

Division of Electric Power Engineering


Department of Energy and Environment
Chalmers University of Technology
2(122)

Abstract
Wind power plays a central role for the development of a sustainable electric power supply
system. In view of climate change and limited primary energy resources, ambitious goals
have been set to promote a strong increase of wind energy utilization. While the cost of on-
shore wind power is already today comparable to that from conventional power generation
sources, the cost is higher for wind power from offshore and remote onshore wind parks. A
considerable portion of 20-30% of the life-cycle cost of wind turbines is constituted by the
cost for operation and maintenance (O&M), required to ensure the technical availability of
the turbines.
To reach cost-efficient maintenance for wind power plants by means of data-based, quanti-
tative methods is the main objective for research in the Wind Power Asset Management
(WindAM) group at the Division of Electric Power Engineering, Chalmers.
In the WindAM-RCM project, funded by the research foundation of Göteborg Energi AB, the
concept of Reliability-Centred Maintenance (RCM) has been applied to wind turbines. In a
workgroup involving Göteborg Energi AB as a wind-turbine owner and operator, Triventus
Service AB as a maintenance service provider, SKF as a provider of condition-monitoring ser-
vices and a wind-turbine component supplier, as well as the WindAM group at Chalmers, an
RCM analysis of the wind turbines Vestas V44-600kW and V90-2MW has been carried out.
The study has identified the most critical subsystems, their dominant failures modes, the
underlying failure causes and suitable preventive measures. In this way, it provides the basis
for the development of quantitative models for the comparison and optimization of mainte-
nance strategies for wind-turbine components, e.g. in the context of the Reliability-Centred
Asset Maintenance (RCAM) method. In the second part of the project, approaches for opti-
mally using information from condition-monitoring systems (CMS) in the maintenance pro-
cess have been investigated. The main focus has been set on developing a methodology to
predict the residual life of wind-turbine components based on their age and data from
online-CMS. A lifetime-prognosis model has been implemented for the case of generator
bearings in wind turbines using vibration data in order to investigate the applicability of the
approach.
3(122)

Sammanfattning
Vindkraft spelar en central roll för utvecklingen av ett hållbart elförsörjningssystem.
Ambitiösa mål har satts med tanke på klimatförändringar och begränsade primära
energiresurser för att främja en kraftig ökning av vindkraftsanvändning. Medan kostnaden
för landbaserad vindkraft redan idag är jämförbar med den från konventionella kraftverk, är
kostnaden högre för vindkraft från havsbaserade och avlägsna landbaserade
vindkraftsparker. En betydande del av livscykelkostnaden för vindkraftverk, 20-30%, utgörs
av kostnaden för drift och underhåll (O&M) som krävs för att säkerställa teknisk
tillgänglighet för vindturbinerna.
Målet för forskningen inom Wind Power Asset Management (WindAM) gruppen vid
Avdelningen för elteknik, Chalmers, är att nå kostnadseffektivt underhåll av vindkraftverk
med hjälp av databaserade, kvantitativa metoder är
I WindAM-RCM-projektet, som finansieras av Göteborgs Energis forskningsstiftelse, har
konceptet Reliability-Centred Maintenance (RCM) tillämpats på vindkraftverk. I en
arbetsgrupp med Göteborg Energi AB som ägare och operatör av vindkraftverk, Triventus
Service AB som underhålls-serviceleverantör, SKF som leverantör av tillståndsövervaknings-
tjänster och vindturbinkomponenter, liksom WindAM-gruppen på Chalmers, har en RCM-
analys av vindturbinerna Vestas V44-600kW och V90-2 MW utförts. Studien har identifierat
de mest kritiska delsystemen, deras dominerande felmoder, de bakomliggande orsakerna
och lämpliga förebyggande åtgärder. På detta sätt ger studien en grund för utveckling av
kvantitativa modeller för jämförelse och optimering av underhållsstrategier för
vindturbinkomponenter, t.ex. i samband med Reliability-Centred Asset Maintenance (RCAM)
metoden. I den andra delen av projektet har metoder för optimal användning av information
från tillståndsövervakning system (Condition-Monitoring Systems, CMS) i underhålls-
processen undersökts. Huvudfokus har satts på att utveckla en metod för att förutsäga den
återstående livslängden för vindturbinkomponenter baserade på deras ålder och data från
online-CMS. En livstidsprognosmodell har skapats för fallet av generatorlager i vindkraftverk,
baserande på vibrationsdata från CMS, för att undersöka tillämpbarheten av metoden för
vindturbiner.
4(122)

Table of contents

Abstract ........................................................................................................................................................... 2

Sammanfattning .............................................................................................................................................. 3

Table of contents ............................................................................................................................................. 4

Preface ............................................................................................................................................................. 7

List of abbreviations ......................................................................................................................................... 8

1 Introduction ............................................................................................................................................ 9

1.1 Background ............................................................................................................................................... 9

1.2 Report outline ......................................................................................................................................... 10

2 Methodology and state of the art ......................................................................................................... 12

2.1 Reliability-Centred Maintenance (RCM) ................................................................................................. 12

2.2 Quantitative maintenance optimisation techniques .............................................................................. 12

2.3 Reliability-Centred Asset Maintenance (RCAM) ..................................................................................... 14

2.4 RCM, QMO, and RCAM – State of the art with respect to application in wind power ........................... 16

2.5 Condition-Monitoring Systems (CMS) – Facilitators for condition based maintenance......................... 19

2.6 CMS for wind turbines - State of the art ................................................................................................. 20


2.6.1 Data acquisition, processing and analysis techniques .................................................................. 22
2.6.2 CMS utilisation in the maintenance process ................................................................................. 27
2.6.3 Failure prognosis ........................................................................................................................... 28

3 Reliability-Centred Maintenance study ................................................................................................. 30

3.1 Implemented RCM process..................................................................................................................... 30

3.2 Results and discussion: V44-600kW wind turbine .................................................................................. 32


3.2.1 System description ........................................................................................................................ 32
3.2.2 Subsystem selection based on statistical analysis and practical experience ................................ 34
3.2.3 Electrical system ............................................................................................................................ 36
3.2.4 Generator ...................................................................................................................................... 40
3.2.5 Gearbox ......................................................................................................................................... 43
3.2.6 Hydraulic system ........................................................................................................................... 48

3.3 Results and discussion: V90-2MW wind turbine .................................................................................... 50


3.3.1 System description ........................................................................................................................ 50
3.3.2 Subsystem selection ...................................................................................................................... 51
5(122)

3.3.3 Gearbox ......................................................................................................................................... 53


3.3.4 Generator ...................................................................................................................................... 54
3.3.5 Converter ...................................................................................................................................... 55
3.3.6 Yaw system .................................................................................................................................... 57
3.3.7 Pitch system .................................................................................................................................. 57
3.3.8 Rotor.............................................................................................................................................. 58

3.4 Wind turbine service and maintenance practices today ........................................................................ 59


3.4.1 Service maintenance ..................................................................................................................... 59
3.4.2 Corrective maintenance ................................................................................................................ 59
3.4.3 Condition-based maintenance ...................................................................................................... 60
3.4.4 Inspections .................................................................................................................................... 60
3.4.5 Identified challenges ..................................................................................................................... 60

3.5 A quantitative model for maintenance strategy assessment ................................................................. 62

4 CMS-data based prognosis of the residual life of wind-turbine components......................................... 63

4.1 Data acquisition and signal processing ................................................................................................... 63

4.2 Component selection and analysed data ............................................................................................... 65

4.3 Main assumptions................................................................................................................................... 68

4.4 Factors influencing the vibration level.................................................................................................... 69


4.4.1 Ambient temperature ................................................................................................................... 70
4.4.2 Rotational speed ........................................................................................................................... 70

4.5 Data reduction ........................................................................................................................................ 71

4.6 Mathematical reliability modelling ......................................................................................................... 74


4.6.1 Basic equations and terminology .................................................................................................. 74
4.6.2 Univariate models ......................................................................................................................... 75
4.6.3 Multivariate models ...................................................................................................................... 77
4.6.4 Proportional Hazards Model ......................................................................................................... 77
4.6.5 Maximum-likelihood based parameter estimation....................................................................... 79
4.6.6 Goodness-of-fit test for the PHM .................................................................................................. 80
4.6.7 Prediction of covariate behaviour ................................................................................................. 81
4.6.8 Residual-life prognosis .................................................................................................................. 85
4.6.9 Quality measures for the prognosis model ................................................................................... 86

4.7 Results and discussion ............................................................................................................................ 88


4.7.1 Lifetime prognosis results in case of a turbine with bearing failure ............................................. 88
4.7.2 Aspects investigated during model development ......................................................................... 90
4.7.3 Potential and limitations of the final prognosis model ................................................................. 95

5 Summary and conclusions ................................................................................................................... 100

6 Related work....................................................................................................................................... 103


6(122)

7 Publications and communication ........................................................................................................ 105

7.1 Publications and reports ....................................................................................................................... 105

7.2 Conferences and seminars.................................................................................................................... 106

7.3 Study visits ............................................................................................................................................ 107

Appendix ...................................................................................................................................................... 108

A.1 Relevant terminology ............................................................................................................................... 108

A.2 Example of an event report ..................................................................................................................... 110

A.3 Failure report sheet ................................................................................................................................. 111

A.4 Test for equality of life distributions by means of the proportional-hazards model ............................... 112

References ................................................................................................................................................... 114


7(122)

Preface
The post-doc project “Maintenance management of wind power systems by means of
Reliability-Centred Maintenance (RCM) and Condition-Monitoring Systems” (Swedish:
“Underhållsmanagement för vindkraftverk med RCM och CMS”, project number 09-11,
acronym WindAM-RCM) has been carried out by Katharina Fischer at the Division of Electric
Power Engineering, Chalmers, during February 2010 to January 2012. It has been conducted
within the Wind Power Asset Management (WindAM) group led by Lina Bertling Tjernberg
and in close cooperation with Göteborg Energi AB, SKF and Triventus Service AB.
The project has been funded by the Research Foundation of Göteborg Energi AB, which the
author gratefully acknowledges.
The author would like to thank the project partners for their active and valuable involvement
in the project. In particular, these are
• the members of the RCM workgroup Inge Aasheim (SKF Renewable Energy), Ulf Halldén
(Triventus Service AB), Björn Mathiasson (SKF Sweden), Peter Schmidt (SKF Denmark)
and Thomas Svensson (Göteborg Energi AB), for their comprehensive contributions to
the RCM study; and
• Michael Wika and Erwin Weis (SKF WISC), for supporting this project with CMS data and
valuable discussions.
Within the RCM study, failure data of the V44-600kW system has been provided by courtesy
of Swedpower / Vattenfall Power Consultant. Failure data of the V90-2MW has been used
for statistical analysis with kind permission of Vattenfall.
The project has been followed by a reference group, the support of which is gratefully
acknowledged. The reference group involved: Inge Aasheim and Olle Bankeström (SKF);
Thomas Svensson and Per Carlson (Göteborg Energi AB); Maria Danestig, Linus Palmblad and
Angelica Pettersson (STEM); Solgun Furnes (EBL); Jørn Heggset (SINTEF); Gunnar Olsson
Sörell (Fortum / Triventus); Thomas Stalin and Fredrik Carlsson (Vattenfall).
Finally, thanks to Lina Bertling for the opportunity to carry out this work, and to the
colleagues in the Optimization group (Mathematics Department, Chalmers) and the WindAM
group, in particular to François Besnard, for the valuable discussions.
8(122)

List of abbreviations
ALCA Asset Life-Cycle Analysis
CBM Condition-Based Maintenance
CMS Condition-Monitoring System
DS Drive side of a generator
DTMM Delay-Time Maintenance Model
EMS Engine Management System
HUMS Health and Usage Management System
MSF Modelling System Failures
FMECA Failure Mode, Effects and Criticality Analysis
NDS Non-drive side of a generator
O&M Operations and Maintenance
PHM Proportional-Hazards Model
RCAM Reliability-Centred Asset Management
RCC Rotor Current Control
RCM Reliability-Centred Maintenance
RPN Risk Priority Number
QMO Quantitative Maintenance Optimization
WindAM Wind Power Asset Management
hc high rotational-speed class
lc low rotational-speed class
9(122)

1 Introduction

1.1 Background
Wind power plays a central role for the development of a sustainable power supply system.
Ambitious goals have been set to promote the strong increase of wind energy utilisation
which is required in view of climate change and limited primary energy resources: From
84 GW installed wind power capacity in Europe at the end of 2010, an increase to 230 GW by
2020 is targeted, of which a minimum of 40 GW is supposed to be installed offshore.
[EWEA (2010), EWEA (2011)]
The cost for operations and maintenance (O&M) of wind turbines presently constitutes a
considerable portion of the life-cycle cost (LCC) and thus of the cost of wind energy:
approximately 20-30% onshore and up to 30% of the – in this case considerably higher – LCC
in offshore installations.
A reduction of maintenance cost and therefore a reduction of the cost of wind energy can be
achieved by further improvements in wind turbine design leading to improvement of its
inherent reliability, but also by systematic solutions for maintenance management. Previous
research has shown that the present maintenance, in both on- and offshore installations, is
not optimised and that there are large potential savings by reducing (a) the cost for
maintenance activities and component failure, and (b) cost due to production losses.
The selection of the most suitable maintenance strategies for the wind turbine components
and their optimal implementation are important steps towards achieving cost-effective
maintenance of wind turbines.
Figure 1 gives an overview over the different types of maintenance strategies. In contrast to
the conventional maintenance strategies of (a) running to failure and subsequently
performing repair (so-called corrective maintenance) or (b) doing preventive maintenance in
a predetermined way, e.g. in fixed time intervals, (c) condition-based maintenance (CBM) is
based on the information collected by means of condition monitoring.

Figure 1: Maintenance strategies, adopted from [SS-EN 13306]


10(122)

Figure 2 illustrates the influence of the three different maintenance strategies on the condi-
tion of a system.
Among the advantages of CBM over the former approaches are:
• Developing faults can be identified before they reach a critical level. Unexpected failures,
which are typically afflicted with a long downtime, can be avoided
• Unnecessary maintenance is avoided by carrying out maintenance only when there is
evidence of abnormal behaviour of a component.

Figure 2: Impact of maintenance strategies on component condition,


adopted from Faulstich et al. (2009)

In case of gradually deteriorating systems for which the cost of corrective replacement
exceeds the cost for preventive replacement (e.g. due to secondary damage caused by
component failure or due to longer downtime in case of unexpected failure), CBM is often
the most cost-effective maintenance strategy.

1.2 Report outline


The WindAM-RCM project has covered two main topic areas: the application of the concept
of Reliability-Centred Maintenance (RCM) to wind turbines, and the optimal utilisation of
data from condition-monitoring systems (CMS) in the wind-turbine maintenance process.
The project report is therefore structured in the following way:
Chapter 2 provides an introduction to the two topic areas RCM and CMS and the related
methodology. It summarises the results of a literature study on the key topics and focuses in
particular on the respective state of the art in wind power.
Chapter 3 covers the work on first main topic area of this project: It presents the RCM
analysis of two wind turbines, including the implemented analysis process, the system
descriptions, the comprehensive results obtained and their utilisation for the development
of quantitative models for maintenance strategy selection and optimisation.
Subsequently, Chapter 4 covers the second main topic area. It presents a methodology for
CMS-data based prognosis of the residual life of wind-turbine components and its
11(122)

implemented for the case of generator bearings monitored by means of a vibration-based


CMS.
Chapter 5 provides a summary of the overall performed work, the key results and the
conclusions.
Three thesis projects, which have been carried out within the WindAM-RCM project and are
complementing the described work, are summarised in Chapter 6.
Chapter 7 provides a compilation of the publications and reports as well as the conference
presentations through which the results from the present project have been disseminated.
The Appendix contains a list of the relevant terminology as well as material complementing
the main part of the report.
A bibliography covering all references given in the text constitutes the final part of the
report.
12(122)

2 Methodology and state of the art

2.1 Reliability-Centred Maintenance (RCM)


The Reliability-Centred Maintenance method is a risk-based qualitative approach for
maintenance strategy selection, which aims at preserving the functions of a system or, using
another common term, of a physical asset. Rausand and Høyland (2004) summarise RCM as
a systematic analysis of system functions and the way these functions can fail as well as a
priority-based consideration of safety and economics that identifies applicable and effective
preventive maintenance tasks. A comprehensive introduction to the RCM method is given in
a book by Moubray (1991) with the same title. In the SAE standard JA1011 (“Evaluation
Criteria for Reliability-Centered Maintenance (RCM) Processes”), the key attributes of RCM
are summarised using seven questions:
1. What are the functions and associated desired standards of performance of the asset in
its present operating context?
2. In what ways can it fail to fulfil its functions?
3. What causes each functional failure?
4. What happens when each failure occurs?
5. In what way does each failure matter?
6. What should be done to predict or prevent each failure?
7. What should be done if a suitable proactive task cannot be found?
The working process incorporated in the RCM method originates in the civil aircraft industry
where it was developed and applied to the Boeing 747 in the late 1960s. A first full
description and publication under the present name Reliability-Centred Maintenance was
provided by Nowlan and Heap in 1978 on behalf of the US Department of Defense [Quinlan
(1987)]. Since that time, the approach has been applied in several industrial sectors with
considerable success, including e.g. the railway, offshore oil & gas and manufacturing sectors
[Andrawus (2008)].
However, a significant limitation of RCM as a purely qualitative method is its capability of
determining which maintenance strategies are the most cost effective options available.

2.2 Quantitative maintenance optimisation techniques


Quantitative maintenance optimisation (QMO) techniques are characterized by the
utilization of mathematical models which quantify both the cost and the benefit of
maintenance and determine an optimum balance between these [Dekker (1996)]. As
illustrated in Figure 3, the optimisation objective is often to find the minimum total cost con-
sisting of
• the direct maintenance costs, e.g. for spare parts, technicians, transportation and
equipment, which increase with the maintenance effort, and
13(122)

• the indirect costs of not performing maintenance as required, i.e. due to loss of
production and due to additional labour and material after component breakdowns
[Andrawus (2008)].

Figure 3: Basic concept of quantitative maintenance optimisation,


adopted from Andrawus (2008)

The main purpose of quantitative maintenance optimisation is to assist the management in


decision making, by utilising available data and thus reducing reliance on subjective
judgement of experts [Jardine and Tsang (2006)].
In the context of maintenance optimisation, it is important to note that a maintenance
strategy being optimal at one point in time might no longer be appropriate in the near
future. Due to the variable nature of the input variables to the optimisation problem
(including e.g. interest rate, component cost or failure behaviour), maintenance optimisation
is not a one-time procedure but a continuous process which requires periodic evaluation
[Andrawus (2008)].
A central limitation of quantitative maintenance optimisation techniques to be aware of is
that these alone do not ensure that the maintenance efforts are targeted on the critical
components and the practically most relevant issues of a system.
14(122)

2.3 Reliability-Centred Asset Maintenance (RCAM)


The Reliability-Centred Asset Maintenance approach (RCAM) is a quantitative approach of
RCM relating preventive maintenance on the component level to system reliability and total
cost. As illustrated in Figure 4, it merges the concepts of RCM and QMO and in this way
overcomes the drawbacks of the two separate approaches.

Figure 4: Interrelation of RCM, QMO and RCAM

The method was originally proposed for power distribution systems by Bertling (2002) and
published under the name RCAM in Bertling et al. (2005). Figure 5 gives a logic diagram of
the RCAM method. The figure illustrates the different stages and steps in the method as well
as the systematic process for analysing the system components and their failure causes. The
three main stages of the RCAM approach are the following [Bertling et al. (2005)]:
Stage 1: System reliability analysis; definition of the system and identification of critical
components
Stage 2: Component reliability modelling; detailed analysis of the components and, based
on appropriate input data, definition of the quantitative relationship between
their reliability and preventive maintenance measures
Stage 3: System reliability and cost/benefit analysis; places the results of the component-
level analysis (Stage 2) in a system perspective and evaluates the effect of
component maintenance on system reliability and cost
By merging the concept of RCM with quantitative methods, the RCAM method provides an
instrument for the quantitative assessment and comparison of maintenance strategies.
15(122)

Figure 5: Reliability-Centred Asset Maintenance method (RCAM),


adopted from Bertling et al. (2005)
16(122)

2.4 RCM, QMO, and RCAM – State of the art with respect to
application in wind power
A literature study concerning the application of RCM, QMO, and the more recent RCAM
method within the field of wind power has revealed that there are only few publications
documenting work on this topic. Relevant contributions, some of them being limited to a
partial application of the methods of interest, are summarised in the following.
An early paper presenting the logical process of RCM and its application to an example wind
park was published by P.J. Quinlan in the USA in 1987 [Quinlan (1987)]. Based on statistical
studies of failures in wind parks by Lynette [Lynette (1985), Lynette (1986)] as well as
operational experience gathered at Southern California Edison [Scheffler et al. (1986)], the
RCM logic was applied and examined for each component of a horizontal-axis based wind
power system. For a relatively low level of differentiation between components (e.g.
foundation, gearbox, generator), failure consequences are classified (hidden, safety,
operational, and non-operational) and the maintenance strategies resulting from the RCM
process are listed. Due to the considerable growth and further development of wind power
technology during the past decades, the outcomes of this early work are of limited validity
for the O&M of wind turbines today.
A pre-study on reliability-centred maintenance for wind power systems with focus on the
effect of CMS on the maintenance process was carried out at KTH in 2005-2006 [Bertling et
al. (2006)]. Within this work, two master’s theses were accomplished as initial steps towards
the application of the RCAM method to wind power systems:
In the master’s thesis of J. Ribrant, titled “Reliability performance and maintenance – A
survey of failures in wind power systems” [Ribrant (2006)] and a resulting publication
[Ribrant and Bertling (2007)], failure data from >95% of all Swedish wind turbines (1997-
2005) was analysed and compared with German and Finnish failure statistics. By studying
system and component reliability and answering the question which components were most
critical in wind turbines regarding the number of failures and the resulting downtime caused
by these failures, the thesis contributed to the stages 1 and 2 of the RCAM method. The
results revealed that the gearbox was most critical to the availability of wind turbines and
that the majority of gearbox failures were caused by wear, indicating that this component
was suitable for the application of condition monitoring systems. Among the results of the
statistical analysis was the observation that turbines with rated capacities >1MW, being
state of the art technology today, suffer from higher failure rates than smaller turbines
(<1MW) installed ealier which still represent the majority of the wind turbines installed
worldwide. The same conclusion of failure frequency increasing with turbine size was drawn
from failure data collected within the comprehensive German WMEP programme [Faulstich
et al. (2008)], making the necessity of both, increased inherent reliability and improved
maintenance particularly for large offshore wind turbines even more evident.
A master’s thesis by J. Nilsson titled “Maintenance management of wind power systems –
Cost effect analysis of condition monitoring systems“ [Nilsson (2006), published in Nilsson
17(122)

and Bertling (2007)], a life-cycle cost analysis was carried out to investigate if CMS were
profitable for wind turbines onshore and offshore, comparing different alternative
maintenance strategies. The work included case studies based on data from the onshore
wind park Olsvenne2 at Näsudden (Gotland, Sweden) and the offshore wind park Kentish
Flats (U.K.). It was shown that in the considered case wind turbine availability had to be
increased by not more than 0.43% to cover the cost of a CMS with the increased profit from
electricity generation. With its economic focus on the cost and benefit of different
maintenance strategies involving CMS, this work is a contribution to stage 3 in the RCAM
process.
Quantitative maintenance optimization techniques have to date been applied to wind power
systems in the work of Andrawus at the University of Aberdeen (UK) as well as in the work of
Besnard at KTH /Chalmers:
In his dissertation “Maintenance optimisation for wind turbines” [Andrawus (2008)] and the
related publications [Andrawus et al. (2006a-c, 2007a-b, 2008)], Andrawus applied a
combination of RCM and QMO in order to investigate and optimize the reliability, availability
and maintenance of 26 wind turbines with a rated capacity of 600 kW. For this purpose, he
combined the analysis of alarm and failure information from the SCADA systems with that of
the related maintenance work orders. RCM was applied to identify failure modes, failure
causes and resulting effects on the system operation. An approach similar to RCAM, called
Asset Life-Cycle Analysis (ALCA), was used for the economic assessment of different
maintenance activities. A Failure Mode, Effects, and Criticality Analysis (FMECA) was carried
out and is included in the dissertation. The published details of the FMECA do not include
any quantitative information about risk priority numbers (RPN), but provide a first
compilation of failure modes and causes in horizontal-axis wind turbines.
The QMO approaches used by Andrawus are Modelling System Failures (MSF) and the Delay-
Time Maintenance Model (DTMM). Using these methods, (a) an age replacement strategy
for several drivetrain components is optimised for which replacement intervals of 3-6 years
are determined, and (b) optimal inspection intervals are calculated for scheduled inspections
of the drivetrain, resulting in an optimum interval of only 30 days. Considering the work and
results of Andrawus, it is important to note that the application of CMS was at a very early
state at the time and in the wind farms investigated. In addition, CMS for continuous
monitoring of wind turbine equipment was not considered in this work. In spite of the
frequent inspection visits required for condition monitoring without CMS, Andrawus states
that condition-based maintenance (CBM) was the most cost-effective option over an 18 year
wind turbine life-cycle. However, the relevance of quantitative results obtained in this work
remains limited to wind turbines without any permanent CMS.
In his licentiate thesis “On Optimal Maintenance Management for Wind Power Systems”
[Besnard (2009)] and the related publications [Besnard et al. (2009, 2010, 2011a), Besnard
and Bertling (2010)], Besnard presents different types of decision models for cost efficient
maintenance planning and maintenance strategies for wind power systems. A so-called
opportunistic maintenance model is proposed for the maintenance planning of service
18(122)

maintenance activities. This takes into account the opportunities for preventive mainte-
maintenance arising from required corrective maintenance after failures, but also weather
and production forecasts in order to reduce transportation costs and production losses. Two
models are developed to analyse the economic benefits of condition based maintenance for
the drive train and for the blades of wind turbines, respectively. In addition, a model is
proposed to optimise the inspection interval for blades, and maintenance strategies for
small components are presented using simple models for component redundancy and age
replacement. The models are tested in case studies and sensitivity analyses are performed
for different parameters of interests. In subsequent work, Besnard has carried out
comprehensive life-cycle cost and profit analysis for offshore wind farms, and investigated
the impact of a variety of factors on O&M cost, among them e.g. the location of
maintenance-team accommodations, the duration of maintenance work-shifts, the
transportation of maintenance teams via boat vs. helicopter, and the optimal number of
maintenance teams [Besnard et al. (2011b)]. An important overall conclusion from the
described work is that the total O&M costs of wind turbines (incl. revenue losses due to
downtime) can be significantly reduced through the optimisation of the maintenance
strategies, the maintenance planning and the maintenance support organisation.
19(122)

2.5 Condition-Monitoring Systems (CMS) – Facilitators for


condition based maintenance
Condition monitoring of physical assets aims at detecting changes in their condition that
represent deviations from the normal operational behaviour and indicate a developing fault
and therefore impending failure. Condition monitoring techniques fall into two categories:
(a) the so-called offline monitoring with portable equipment, and (b) the online monitoring
using permanently mounted sensors. By means of these sensors, the machine is
continuously monitored during operation and warnings or alarms are triggered when
predefined threshold values are exceeded. The term “Condition-Monitoring System” usually
refers to the second type, i.e. it includes only online monitoring systems.
A central task within the present project has been to determine how information from CMS
in wind turbines can be used in an optimal way for maintenance planning. As shortly
explained in the introduction, the utilisation of CMS data enables the implementation of
condition-based maintenance (CBM). As proposed by Jardine et al. (2006), the utilisation of
CMS data for CBM can be considered a three-step process, as illustrated in Figure 6.

Figure 6: Three-step process of utilising CMS data for condition-based maintenance

Data acquisition describes the collection and storage of information, e.g. the measurement
of the vibration amplitude of a component. In order to make use of the large amount of data
collected in this way, it is necessary to extract so-called features from it, which are suitable
indicators for the condition of the monitored component. This usually requires processing
techniques like e.g. a transformation of a vibration signal to the frequency domain and the
calculation of characteristic values. A diagnostic step is required in order to relate the
observations in the data to certain components and their faults, e.g. to link the vibration
amplitude in a certain frequency range to the condition of a particular bearing.
An important step for the utilisation of CMS data for maintenance planning and optimisation
is the third, prognosis. Having identified that a fault is developing and where it is located,
20(122)

this step aims at estimating how soon and how likely failure will occur. At present, the step
of prognosis is usually based on the subjective judgement of the monitoring expert and does
not make use of quantitative methods.

2.6 CMS for wind turbines - State of the art


Condition-monitoring systems are well established in a many industrial sectors. However,
wind turbines differ from the former areas of application with respect to strongly varying
loads and rotational speeds. The monitoring process of several drive-train components in
wind turbines is in addition complicated by their low rotational speed. On the one hand, this
complicates the utilization of recognised condition monitoring and analysis techniques from
other industries. On the other hand, this implies also an advantage since the resulting large
variety of excitation frequencies can also reveal particularly comprehensive information
about the health condition of the wind turbine components of interest.
The history of CMS for wind power applications is still a short one: According to Lucente
(2006), no CMS for wind turbines were available in 2001. At that time, a project for
examining the possibility of using a CMS was carried out by Elforsk in cooperation with
Göteborg Energi and SKF Nova, and a first system by SKF Nova was tested on a Vestas V44
turbine [Jonasson (2001)].
In contrast, more than 40 CMS for application in wind turbines are commercially available
today [SKF (2010)], of which 20 are included in a recent survey by Crabtree [Crabtree
(2010)].
Historically, CMS for wind turbines have mainly been vibration-based systems monitoring
the rotating drive-train components, i.e. the gearbox, the main bearing and the generator.
On the one hand, this is a natural consequence of the existing experience with vibration-
monitoring of rotating machinery from other industries. On the other hand, this focus is
justified by the high value of these components, the considerable effort and cost for repair
after catastrophic failures (usually requiring a crane), and the high average downtime per
wind turbine and year due to failures of these components (see Figure 7).
21(122)

Figure 7: Distribution of downtime over the affected wind-turbine subsystems;


Source: Ribrant and Bertling (2007)

According to Hyers et al. (2006), the focus of CMS is in the process of shifting (or rather
being extended) to blade monitoring as the cost of rotors strongly increases with turbine size
and already at the time of publication in 2006 almost equalled the total cost of the nacelle
and the drive-train, see Figure 8.

Figure 8: Cost of wind turbine components; Source: US DOE (2005)

In the following section, the state-of-the-art condition monitoring and analysis techniques
for wind turbines are summarised.
22(122)

2.6.1 Data acquisition, processing and analysis techniques

2.6.1.1 Drive train


The most established technology for condition monitoring of rotating equipment is vibration
monitoring. Vibration-based CMS make use of the fact that most damages in rotating
machinery lead to excessive vibration and that each mechanical unbalance or defect
generates a unique vibration pattern. Figure 9 illustrates how the time waveform of a vibra-
tion signal is decomposed into its spectral components, often using Fast Fourier Transform
(FFT), and how characteristic frequencies in the resulting spectrum can be related to ma-
chine components. Taking into account the specific geometries of e.g. bearings or gear-
wheels and the related defect frequencies, the content of the vibration spectrum can reveal
detailed diagnostic information about the location of a fault.

Figure 9: Principle of vibration monitoring and signal analysis,


adopted from SKF Vibration Diagnostics Guide (2000)

In variable-speed wind turbines, the permanently changing rotational speed of the wind
turbine rotor causes a shift of the kinematic excitation frequencies and thus a continuously
moving frequency spectrum. In order to still enable analysis in the frequency domain, the so-
called order analysis is used in many CMS for wind turbines today. Order analysis implies
that the time signals are recorded with a speed-dependent sampling frequency or that
recorded time data is re-sampled with a respective speed-dependent frequency, as shown in
Figure 10. In the resulting “static” order spectra, the frequency axis is scaled in accordance
with the harmonics of the rotational frequency.
23(122)

Figure 10: Fixed-frequency vs. moving frequency (or order) analysis; Source: Becker (2008)

Compared with the raw acceleration spectrum analysis, a significantly higher sensitivity to
bearing and gear-tooth defects is provided by so-called frequency enveloping or envelope
analysis: The low-amplitude, repetitive shock impulses induced by initiating defects in bear-
ings and gearwheels are often hidden by the dominating machine vibrations. The enveloping
procedure filters out low-frequency signals and enhances the frequency components in the
range of defect frequencies, which enables a particularly early detection of developing faults
in these components. [SKF Reliability Systems (2000)]
Signal analysis in the time domain can be advantageous in case of low rotational speeds, at
which frequency analysis would require long measuring periods. In addition to waveform
analysis in the time domain and spectrum or order analysis in the frequency domain, an im-
portant procedure in condition monitoring is trend analysis. In that case, the development of
characteristic values (of the time signal or the frequency / order spectrum) are visualised
over time in order to enable an easy visual check of the evolution of a feature and the com-
parison with historic “known-good” values of the same machine. The interpretation of trend
curves of CMS signals is most valuable in combination with operational parameters like e.g.
generated power or wind speed, i.e. parameters which have an impact on the vibration am-
plitude and should therefore be included in the assessment of vibration trends. [Germa-
nischer Lloyd (2007)]
As a complement to vibration monitoring, the continuous monitoring of gearbox lubricant oil
has become a recognised condition-monitoring technique. It can be distinguished between
two systems: “Offline” oil analysis systems sample only a small portion of the oil that is circu-
lated, whereas the less common “inline” monitoring systems monitor all of the circulating oil
[Hyers et al. (2006)]. To maintain oil that is free of hard particles, with controlled properties
such as viscosity, acidity, temperature and moisture content, is essential for the health of
the bearings and gears in the drive train. While these properties are measured with oil con-
dition sensors, the size and concentration of metallic particles in the lubricant are monitored
by means of optical or electromagnetic oil particle counters. Metallic particles are usually
generated by wear and their concentration can, according to Raadnui (2009), be used as an
indicator for gear and bearing condition, see also Figure 11.
24(122)

Figure 11: Wear particle size distribution vs. machinery condition. Source: Raadnui (2009)

Oil monitoring techniques that reveal less detailed diagnostic information but indicate that
“something is wrong” include a measurement of the difference between inlet and outlet
lubricant oil temperatures as well as measurement of pressure drops across oil filters in the
gearbox lubricant or the hydraulic oil. In the first case, an increased temperature difference
is an indication for higher heat dissipation due to wear in the gearbox, while in the second
case, a sudden increase in pressure drop over the oil filters gives evidence of large number
of particles being released, a potential sign of damage. [Hyers et al. (2006)]
A project in which various monitoring techniques using e.g. accelerometers, strain gauges,
thermocouples, and laser-based techniques were evaluated with respect to their application
in wind turbines showed that the earliest warning indications originated from bending and
azimuth measuring devices attached to the main rotor shaft. The monitoring technique de-
veloped on this basis evaluates main-shaft loads as a precursor to equipment degradation
and an indicator of insufficient wind-to-rotor alignment or mass imbalances in the rotor. In
this way, it supports the optimisation of power output as well as a reduction of stress in the
mechanical system. [Schmidt (2010)]
Monitoring techniques which utilise electric signals (voltage, current or power) measured at
the generator terminal - an established method for condition monitoring of induction ma-
chines in other industries - is currently being investigated for application in wind turbines
(see e.g. [Watson and Xiang (2006), Yang et al. (2010), Crabtree (2011)]). The approach has
the advantage that the generator terminal quantities are often easily accessible and that
usually no additional sensors are required for signal acquisition. It may be applied to detect
turn faults, broken rotor bars, bearing failures and air gap eccentricities in generators, but is
also subject to investigation for gearbox and roller bearing failure diagnosis as well as the
detection of imbalance and certain defects in wind turbine blades [Amirat et al. (2009)].
25(122)

2.6.1.2 Blades
Imbalances and aerodynamic asymmetry in the rotor of wind turbines can be detected
through variations in the electric output power or by monitoring of transverse oscillations of
the nacelle. Similarly, increased surface roughness due to erosion, icing or insects can be
identified through comparison of the actual performance with the machine-specific func-
tional relation between wind speed and power output of the wind turbine. [Hyers et al.
(2006)]

The detection of initiating mechanical damage like delamination or crack formation in the
wind turbine blades requires more advanced monitoring techniques. A common but time-
intensive method for offline inspection of the blade structure is the investigation with ultra-
sound waves.
As blades have become an important cost factor in wind turbines, a variety of online condi-
tion-monitoring techniques has been developed for blades, including stress, strain, acoustic
emissions and modal analysis [Hyers et al. (2006)]. Since electrical strain gauges are prone to
failure by disbonding, creep or fatigue and the measurement of electronics drift with time,
they are not suitable for the application in wind turbine blades. Instead, fibre optical sensors
are applied, which are long-term stable, lightning-safe, not susceptible to electromagnetic
interference and embeddable into the rotor-blade laminate. The application of fibre-optical
sensor based techniques like e.g. the Fibre Bragg Grating measurement system shown in
Figure 12 aim at monitoring the spatial load distribution and the detection of initiating cracks
in the blades [Hyers et al. (2006)]. The survey of wind-turbine CMS provided by Crabtree
(2011) lists two commercially available blade-monitoring systems based on fibre optic strain
measurement.

Figure 12: Fibre Bragg Grating measurement system for monitoring of wind turbine blades.
Source: Timmerman (2007)

An approach called Two Wave Mixing (TWM) technique for online blade monitoring is pro-
posed by Kung (2009). It uses surface waves of two different frequencies mechanically cou-
pled onto each blade by means of small actuators and analyses their interaction in order to
detect delamination and fiber breaks in the blade composites. The sensitivity of this tech-
nique is claimed to be two orders of magnitude higher than ultrasound condition monitoring
of blades.
An alternative approach makes use of the fact that every wind turbine blade has characteris-
tic modal dynamics which will be affected by any structural damage in the blade. Inertial
26(122)

sensors (acceleration sensors and rotational rate sensors) located at several positions on the
blade surface are used to measure the linear acceleration or angular motion. In order to
provide diagnostic information about the occurrence and location of damages in the blade,
the signal analysis is carried out in combination with a structural model of the blade and
takes into account the operating conditions of the wind turbine [Lading et al. (2002),
Hameed et al. (2009)]. At least one blade-monitoring system using vibration data from ac-
celerometers is commercially available [Crabtree (2011)].

2.6.1.3 Controls and power electronic components

The high failure frequency in both power electronics and electronic controls in wind turbines
gives reasons to concern, especially for offshore installations, in which the downtime per
failure is usually considerably higher due to accessibility constraints. A common technique
for detecting damage and overheating of electronic and electric components in other indus-
tries is thermography. In wind turbines, manual thermography is occasionally being used for
inspections of electrical-system components. In a recent feasibility study by Anjar et al.
(2011), the use of permanently mounted thermal imaging cameras for condition-monitoring
of wind turbine components has been investigated and identified to be a promising option.
An online-CMS for wind turbines utilising thermal imaging is, however, not available yet.
A basic problem concerning any condition monitoring of the electronic components in wind
turbines is that, particularly for power electronics, the timescales for fault development are
often too fast to take preventive actions before catastrophic failure occurs. Therefore, Hyers
et al. (2006) suggest online protection functions or hardware redundancy for controls and
power electronics instead, as used in aerospace applications, to be implemented in large
turbines and in offshore installations.

2.6.1.4 Tower and foundation

According to Ribrant and Bertling (2007), failures of the structural components of wind tur-
bines stand for only a portion of ca. 1.5% of all failures as well as for a minor mean down-
time per wind turbine and year. Presumably due to that reason, condition monitoring
systems focusing on the integrity of the tower and the foundation, also called structural
health monitoring (SHM) systems, are not commercially available today. For offshore wind
turbines, Kraemer and Fritzen (2008) propose a damage detection approach based on
changes in the dynamic properties of the structural components. For the localisation of
damage in the structure, the authors suggest a mode-based algorithm, solving an inverse
problem using eigenfrequencies and mode shapes from a finite element model in combina-
tion with the measured data.
In contrast to this, the measurement of tower vibrations being part of presently available
CMS aims predominantly at the detection of mass imbalances in the rotor instead of damage
detection in the wind turbine structure itself.
27(122)

2.6.1.5 Further wind turbine components

Among the wind turbine components not discussed in the preceding sections, the yaw sys-
tem stands out due to high failure rates and significant resulting downtime [Ribrant and
Bertling (2007), Khan et al. (2005)]. However, as stated by Verbruggen (2003), the monitor-
ing of this component is difficult due to its intermittent usage. Condition monitoring of the
yaw system is not included in presently available CMS.

2.6.2 CMS utilisation in the maintenance process


At present, the analysis of CMS data is carried out by condition monitoring specialists at
monitoring centres, i.e. at monitoring bodies that are in charge of continuously monitoring a
large number of wind turbines. Based on detailed technical documentation of all turbine
components, including information about relevant natural frequencies of e.g. the tower,
rotor blades and drive train as well as about excitation frequencies of separate units like
pumps or yaw drives (see [Germanischer Lloyd (2007)]), characteristic quantities from the
CMS data which are related to component condition are chosen for the monitoring process.
Based on the experience of the condition monitoring specialists and on reference values of
these quantities measured e.g. during commissioning of the wind turbine, threshold values
are defined for each quantity, which can be more than 200 for a single wind turbine. Subse-
quently during the operational life of the wind turbine, the monitoring expert is automatical-
ly informed with an alarm as soon as the CMS detects the transgression of a threshold value.
In the case of an alert, the possible reactions range from ignoring the fault indication to
watching the evolution of the damage signature, ordering inspections or maintenance to
recommending pre-emptive shutdown. Notifications are sent from the monitoring body to
the wind turbine operator about received alarms and recommended maintenance measures,
e.g. in form of an event report as included in the appendix of this report.
Among the identified challenges in the condition-monitoring process and the utilisation of
CMS for maintenance planning are:
• the limited efficiency of current CMS, leading to unforeseen failures in spite of installed
CMS (see e.g. Bellens and Chemweno (2010)), and
• the large number of alarms some CMS on the market overwhelm the user with, several
of those being related to the same fault in some cases or being caused by transients in-
stead of deteriorated machine condition in other cases
On this background, Hyers et al. (2006) point out that the further development of CMS
should aim at raising the fraction of faults detected, at detecting these faults as early as pos-
sible, and at correctly identifying the faulty component, all while reducing false positive indi-
cations.
A challenge and time-consuming task in the present condition monitoring process is the
manual identification and regular update of suitable threshold values. Since every machine is
manufactured and installed differently, has a unique maintenance history and is operated
28(122)

under specific conditions, it is not appropriate to judge a machine’s condition by comparing


its current measurement value against a wide classification ISO Standard or other general
threshold levels [SKF Reliability Systems (2000)]. Therefore, Hyers et al. (2006) recommend a
combined analysis of CMS data and service history both of the fleet and of the particular
machine, the use of databases of material properties and models of damage progression as
well as forecasts of future service conditions to allow prognosis of the wind turbine condi-
tion. An overview over the state of the art in prognosis related research is subject of the fol-
lowing section.

2.6.3 Failure prognosis


Prognosis aims at providing a probabilistic forecast of the future condition of a machine un-
der projected usage conditions. As explained above, condition monitoring data revealing the
current state of a component is combined with event data and a mathematical model de-
scribing the underlying failure mechanism to develop a forecast of the remaining useful life.
Each piece of new information available reduces the uncertainty in the updated forecast, as
illustrated in Figure 13.

Figure 13: Principle of failure prognosis; Source: Hyers et al. (2006)

Compared to condition-monitoring related data-acquisition, data-processing and diagnosis


techniques, the literature dealing with prognosis is sparse:
A comprehensive review on machinery diagnosis and prognosis for implementing condition-
based maintenance, not being limited to wind power applications, has been published by
Jardine et al. (2006). The authors summarise different methods suitable for combined data
analysis of event data and condition monitoring for the purpose of failure prognosis, among
29(122)

them the Proportional Hazards Model, Hidden Markov Model, Delay-Time Concept, and sto-
chastic process models.
Early efforts in applying prognosis to wind turbines have been described by Seeliger et al.
(2006). These studied the use of condition monitoring together with physics based simula-
tion models to assist in the prognosis of wind turbine failures.
Hyers et al. (2006) give a comprehensive overview over condition monitoring and failure
prognosis for wind turbines, and include in their review the related technologies of Health
and Usage Monitoring Systems (HUMS) and Engine Management Systems (EMS) in civil and
military aviation. According to these authors, failure prognosis could significantly increase
the value of condition monitoring in the drive train and the rotor of wind turbines, while
specific research would be required to apply the principles of prognosis to wind turbines.
From the available literature and the comprehensive discussions with the project partners in
the present project, it may be concluded that prognosis has a large potential, in particular
for offshore turbines. The resulting probabilistic forecast that would be specific to each tur-
bine could allow balancing the risk of running a machine with damage indications against the
lost revenue while waiting for maintenance and could in this way become a valuable tool for
decision support.
30(122)

3 Reliability-Centred Maintenance study


As one of the central elements in this project, a limited-scope RCM analysis of two wind tur-
bines has been carried out. Its purpose in the context of the project is to reveal those com-
ponents, the failure modes as well as the major underlying failure causes in the wind
turbines that are most relevant for the system reliability, availability and O&M cost, and to
identify suitable preventive maintenance measures. In this way, the RCM study forms the
basis of RCAM; it ensures to focus the subsequent development of mathematical models on
the practically relevant items and failures.
The RCM study has been carried out in a workgroup with representatives from Göteborg
Energi AB as a wind turbine owner and operator, Triventus Service AB as a maintenance ser-
vice provider, SKF as both a provider of condition-monitoring services and a wind-turbine
component supplier, as well as the WindAM research group at Chalmers.
The RCM analysis has been carried out for two onshore wind turbines, namely
1) Vestas V44-600kW and
2) Vestas V90-2MW.
In the following, the turbine description and results of the RCM analysis for these two tur-
bines are presented in subsequent order. It is important to note that the turbines have been
selected for analysis not due to any abnormal occurrence of failures but because of the fact
that these are of particular interest to the project partners and, particularly in case of the
V44-600kW, because of the available experience with operation and maintenance of this
turbine type in the RCM workgroup.
Parts of the results described in the following sections have also been published in [Fischer
et al. (2011a)] and [Fischer et al. (2011b)].

3.1 Implemented RCM process


A general introduction to the concept of RCM has been given in Section 2.1 of this report. In
this section, the method is explained in more detail and with a focus on the way of imple-
mentation in the WindAM-RCM project.
According to Rausand and Høyland (2004), a full RCM analysis consists of the following steps:
1. Study preparation
2. System selection and definition
3. Functional failure analysis
4. Critical item selection
5. Data collection and analysis
6. Failure Mode, Effect and Criticality Analysis (FMECA)
7. Selection of maintenance actions
8. Determination of maintenance intervals
31(122)

9. Preventive maintenance comparison analysis


10. Treatment of non-critical items
11. Implementation
12. In-service data collection
Within the limited-scope RCM analysis described here, the steps 1-7 have been implement-
ed. As stated above, this is the basis for maintenance strategy selection and optimization
(covering steps 8 and 9 in the list above) by means of mathematical models, in accordance
with the method of RCAM introduced in Section 2.3. The remaining steps 10-12 do not lie
within the scope of the WindAM-RCM project.
The level of analysis moves from the system level (whole wind turbine) to the subsystem
level (e.g. electrical system, hydraulic system, generator) for which failure data is available,
and further on to selected critical components (e.g. rotor current control unit, hydraulic
pump) of these subsystems. The width of analysis is limited to the most relevant subsystems
of each turbine with respect to failure frequency and resulting downtime as well as their
dominant failure modes. The focus lies on providing an in-depth understanding of the main
failure modes, failure causes and the underlying failure mechanisms on the one hand and
suitable maintenance measures to prevent these on the other hand.
Figure 14 shows the table template used during the RCM process. For each subsystem se-
lected for in-depth analysis and its component specified in the first column, it compiles the
function, the dominant failure modes, their effects or consequences, their causes and under-
lying mechanisms, their failure characteristic, proposed preventive measures and, where this
is appropriate.

Figure 14: Table template used for the RCM analysis (adopted from Rausand and Hoyland (2004))

The consequences of failure have been assessed for the four criteria

S – Safety of personnel,
E – Environmental impact (in a wind turbine e.g. discharge of oil or glycol),
A – Production availability (i.e. the impact on electricity generation),
M – Material loss (including primary damage to the component itself, but also second-
ary damage to other equipment).
32(122)

3.2 Results and discussion: V44-600kW wind turbine

3.2.1 System description


The structure of the V44-600kW system, including its main components located in the na-
celle, is shown in Figure 15. Launched in 1996, it has a design which is mostly identical to the
Vestas V42 turbines and which was state-of-the art during the mid 1990s:
The V44-600kW is an upwind turbine with an electrically driven yaw system and three
blades. Its rotor has a diameter of 44m and a weight of 8400 kg, the rated rotational speed
of this is 28 rpm. The hydraulic pitch system is used for speed control, optimization of power
production, for start-up and for stop (aerodynamic braking) of the turbine. Additional break-
ing functionality is provided by a disc brake located on the high-speed side of the gearbox.
During operation, the main shaft transmits the mechanical power from the rotor to the
gearbox, which has either a combined planetary-parallel design or, as in case of the Göte-
borg Energi-owned turbines analysed here, a parallel-shaft design. Gearbox and generator
are connected with a Cardan shaft. The generator is an asynchronous 4-pole generator with
integrated electronically controllable resistance of the rotor windings (so-called OptiSlip
technology), which requires neither brushes nor slip rings.
The main function of the V44 system (and the only one being of interest in the scope of this
study) is the conversion of kinetic wind energy to electric energy which is provided to the
electric power grid. The generator stator is, via a transformer, directly connected to the grid.
More specifically, the system function is to provide up to 600kW electric power at 690V and
50 Hz to the electric power grid, in an operating temperature range of -20…+40°C and at
wind speeds of 4-20 m/s.
33(122)

1. Base frame 9. Generator


2. Main shaft 10. Cardan shaft
3. Hub 11. Hydraulic system
4. Blade 12. Yaw gear
5. Blade bearing 13. Yaw ring
6. Gearbox 14. Yaw control
7. Gear tie rod 15. VMP top control unit
8. Disc brake

Figure 15: Structure of the V44-600kW system; Source: [Vestas General Specifications
V42/V44-600kW (1995)]

Failures on the system level which are relevant in this study are (1) the complete loss of en-
ergy conversion capability or (2) the partial loss of energy conversion capability of the tur-
bine.
The wind turbine system has four operating states:
1) RUN (DRIFT),
2) PAUSE (PAUS),
3) STOP (STOPP), and
4) EMERGENCY (NÖDSTOPP).
34(122)

Only in the operating state RUN, the turbine can be connected to the grid and fulfil the sys-
tem function defined above.
In 2010, a total number of 35 V44-600kW systems were in operation in Sweden. Five of the-
se are installed in the harbour of Göteborg since 1996. Four of them are owned by Göteborg
Energi, while one is jointly owned by Göteborg Energi and Shell.

3.2.2 Subsystem selection based on statistical analysis and practical


experience
In the RCM study, failure statistics of the investigated wind turbine have been used in com-
bination with expert judgement in order to prioritize the wind turbine subsystems for subse-
quent detailed analysis.
The failure data used for statistical analysis covers the failures of 32 V44-600kW turbines in
the period 1996-2005 which are part of the database “Felanalys, Database of failures for
Swedish wind power turbines 1989-2005” [Swedpower database]. This database includes
failures that caused a downtime of at least 1h. The sheet used for failure reporting to this
database, which shows the classification of components into subsystems, is provided in Fig-
ure 48 in the appendix of this report. Downtime due to external grid failures, inspections,
service maintenance and other external causes are not reported in the database and have
therefore not been taken into account in the analysis.
The V44 subsystems with the highest failure frequency are, in the given order, the electrical
system, the hydraulic system, the control system, and the generator. The longest downtime
per failure has been found to occur due to failures of the gearbox, the yaw system, and the
generator. As both the failure frequency and the downtime resulting from a failure are rele-
vant for the criticality assessment of components, it is advantageous to combine these two
measures by multiplication, resulting in the average downtime per wind turbine and year
due to failures of a specific subsystem or component (see also e.g. Koutoulakos (2008)):
I

∑d
i =1
i
t lost = I (1)
∑X
i=1
i ⋅ Ti

with
di Downtime due to failures of a subsystem in the time interval Ti
Xi Number of wind turbines reporting to the database in time interval Ti
Ti Duration of time interval i
Figure 16 shows the distribution of this quantity tlost, the average downtime per turbine and
year, over the causing subsystems.
35(122)

Figure 16: Results of the V44 failure data analysis; Data source: [Swedpower database]

In order to include the practical experience of the RCM workgroup members in the subsys-
tem selection, questionnaires were filled in by all group members having professional expe-
rience with wind turbine operations and maintenance.
The different priority lists resulting from the statistical analysis and the expert judgement are
summarised in Table 1. Note that the subsystems are listed in the order of descending priori-
ty, i.e. the most upper row contains the subsystem related to the highest failure frequency,
downtime per failure and average downtime per turbine and year, respectively.

Table 1: Criticality of wind turbine subsystems with respect to failure frequency and downtime,
according to expert judgment and statistical data analysis (in order of descending priority)

V44-600kW
Expert judgment Data analysis

Failure frequency Downtime per failure Downtime per year and turbine

1. Gearbox Gearbox Electrical system


2. Generator Generator Generator
3. Hydraulic system Yaw system Control system
4. Rotor Rotor Gearbox
36(122)

Based on the results of both the failure data analysis and the questionnaire assessment, the
following subsystems of the V44-600kW system were selected for in-depth analysis in the
RCM study:
• electrical system
• generator,
• gearbox,
• hydraulic system
In spite of the significant contribution of the control system to the average downtime per
wind turbine and year (see Fig. 16), it was decided not to include this system in the RCM
analysis because failures in this subsystem can hardly be influenced by means of preventive
maintenance.
The main results of the RCM analysis are summarized in the following sections. It should be
noted that these reflect to a large extent the subjective experience of the RCM workgroup
members. Where it was possible, the obtained results have been compared and backed up
with information from the open literature, as will be indicated by references to the respec-
tive source. Information provided without any specified reference has been obtained from
the RCM workgroup discussions.

3.2.3 Electrical system


Out of the numerous items falling into the category Electrical system (see the failure report-
ing sheet in Figure 48, the following two components have been discussed in the RCM analy-
sis: the Rotor Current Control (RCC) and the power factor correction unit.

3.2.3.1 Rotor Current Control (RCC)

The most frequently failing part of the electrical system according to the failure statistics is
the RCC unit. This is in agreement with the experience of the RCM group and also supported
by Carlson et al. (2001). In average, the RCC unit has to be exchanged at least once during
the 20-25 years lifetime of the wind turbine. The RCC unit is mounted to the non-drive end
of the generator rotor and thus permanently rotates during wind turbine operation. In spite
of its position inside the generator, the RCC unit is discussed as a part of the electrical sub-
system here in accordance with the categorization of failure data in [Swedpower database].
Depending on the generator model, two different models of rotor current control units are
found in V44 turbines: VRCC (Vestas Rotor Current Control) in ABB generators and RCC in the
generators manufactured by the Vestas-owned Weier. Of the five V44 turbines owned by
Göteborg Energi, four contain Weier generators, so that the discussion mainly focused on
this generator model which is predominantly installed in V42 and early V44 turbines.
The function of the RCC unit is to control the generated electric power by regulating the ro-
tor current, which in turn is realised through variation of the rotor resistance. Failure of the
37(122)

RCC unit results in a loss of this current control capability. However, the consequence for the
system function, i.e. if this results in a full or only a partial loss of the energy conversion ca-
pability, depends on which part of the unit has failed.
The RCC unit consists of three components, see also Figure 17, which are discussed in subse-
quent order: the power electronics unit, the microprocessor unit, and the resistor unit.

Figure 17: Components of the Rotor Current Control (RCC) unit; connection panel at the generator
backside (left); resistor unit (middle); power electronics unit (right). (Source: Triventus Service AB)

3.2.3.1.1 RCC - Power electronics unit

Out of the three parts of the RCC, the power electronics unit has the shortest lifetime of ap-
prox. 10-12 years. Its failure rate increases with age. In some cases, the failure is caused by
loose contacts or cable twist and thus by mechanical impact, often due to vibrations. More
frequently, however, it is observed that the power electronics components fail. This happens
predominantly in the mornings and evenings, at times during which there are frequent cou-
pling activities in the power grid. Due to this observation, it is probable that the failures are
related to grid disturbances, e.g. voltage dips, harmonics or frequency variations. The fre-
quent failure of this component in connection with grid disturbances might be explained by
the change of grid characteristics since mid of the 1990es when the V44 system was de-
signed, thus with the design being no longer adapted to today’s operating conditions in the
grid. Measuring and analysing the grid voltage and frequency at the moment of RCC failures
could help to clarify a possible relationship between these phenomena.
A further cause of failure of the RCC power electronics unit is overheating due to high tem-
perature inside the nacelle which might exceed the design operating-conditions of the unit.
This has for example once been observed as a secondary damage after a defect in the cool-
ing system of the gearbox (in more detail: the hose leading the heat out of the nacelle).
In case of failure of the power electronics unit, the large majority of turbines can be operat-
ed at reduced power of 300kW (with a fixed rotor resistance resulting in 10% slip in the gen-
erator). Failure of this unit thus causes a partial, but usually not a full loss of power
generation capability.
38(122)

The different failure causes explained here require different measures for prevention: Vibra-
tion-related failures of the power electronics unit can be avoided through the reduction of
vibrations, e.g. by monitoring of the bearing condition and a preventive bearing replacement
before the occurrence of strong vibrations. Due to the age-dependency of failures, a preven-
tive replacement of the RCC power electronics unit (which can only be exchanged together
with the microprocessor unit) at the age of 10 years could be taken into account. However,
due to the high cost and the possibility of continued operation at reduced power, this is not
considered to be an economic solution.

3.2.3.1.2 RCC - Microprocessor unit

The microprocessor unit of the RCC provides the target value for the rotor current control.
The control signal is transmitted optically. Failure of the microprocessor unit has seldom
been found in V44 turbines with Weier generators. As in case of a defect in the RCC power
electronics unit, the wind turbine can be operated at reduced power of 300kW when a fail-
ure of the microprocessor unit has occurred.

3.2.3.1.3 RCC - Resistor bundle

The resistor bundle, which is bolted to the non-drive end of the rotor, is electrically connect-
ed to the rotor windings. The rotor resistance is varied in the way that the resistor bundle is
short-circuited at varying frequency by means of IGBTs, see also Figure 18.
Failure of the resistor bundle occurs seldom in V44 turbines with Weier generators. It fails
due to overheating of the resistors when the generator operates with a slip ≥3% for a longer
time which leads to high power in the rotor circuit and thus strong heat dissipation in the
resistors. The RCC system is designed to control the turbine power during wind gusts. Over-
heating can occur when the turbine is hit by wind gusts too frequently so that the heat gen-
erated in the resistor bundle is not sufficiently compensated by a heat flow out of the
generator. If overheating of the resistor unit is detected by the SCADA system, the wind tur-
bine is set to PAUSE state in order to avoid material damage. Thermal damage can occur
when the temperature measurement is not functioning.
Another failure cause of the resistor unit is cracking of welds due to the impacts of rotation
and vibrations during operation of the generator.
In case of a failure of the resistor bundle, it is observed that the wind turbine starts to swing,
a process which involves a periodical pitching of the blades as well as oscillation of the rota-
tional speed and electric power. In that case, the turbine is stopped by the control system in
order to avoid excessive loading of the gearbox.
39(122)

Figure 18: Structure of the wound rotor asynchronous generator with “OptiSlip”
technology. Source: Lin et al. (2009)

3.2.3.2 Power factor correction unit

A frequently failing component of the V44 is the capacitor bundle that is located at the bot-
tom of the tower and provides reactive power to the generator. It is therefore also called
phase compensation or power factor correction unit. Failure of these capacitors is a common
cause of fire in wind turbines. A lack in manufacturing quality is considered to be the main
cause of failure. No suitable preventive measures to be applied in the operational phase
have been identified.

3.2.3.3 Current maintenance practices

Maintenance of the electrical system is today typically dominated by corrective mainte-


nance. A previously done preventive exchange of selected components in a 2-year interval is
no longer carried out. In this way, it is tried to reduce the risk of introducing additional fail-
ures through preventive replacements. In particular for the RCC unit, preventive replace-
ment is not considered to be economic due to its high cost and the limited advantage of
planned versus corrective replacement.
The fact that the fault search process in the RCC requires a wind speed sufficient to start the
wind turbine can lead to additional downtime in case of failure of this component.
Repair of the RCC unit can only be carried out at the manufacturer since it requires tools and
test equipment not available otherwise. Depending on the failed component, either the sin-
gle resistor unit or the bundle of the power electronics and the microprocessor unit have to
be replaced. RCC units are no longer being produced. Spare units are usually bought second
hand, i.e. in a used, possibly overhauled condition.
40(122)

3.2.4 Generator
The function of the asynchronous generator of the V44 turbine is to provide 600kW of elec-
tric power at a voltage of 690V and a frequency of 50 Hz or 60 Hz (this analysis is limited to
operation in a 50 Hz grid) at a rotational speed of 1500 rpm (idle) to 1650 rpm. Enabled by
the OptiSlip technology, i.e. the control of the rotor resistance according to Figure 18, the
slip

ns − n
s= (2)
ns

describing the relative deviation from the synchronous rotational speed ns may vary in the
range -1…-10%.
According to the failure statistics, the main components causing generator failure are the
windings and the bearings. This agrees with the experience of the RCM workgroup.

3.2.4.1 Windings

The generator incorporates the windings both in the 4-pole stator and in the wound rotor.
The functions of the windings are to lead electric currents that generate magnetic fields;
their interaction leads to the generation of electric power which is, from the terminal of the
stator windings, fed into the electric power grid.
The consequences of winding failure range from a limited electric unbalance to a standstill of
the wind turbine. It requires de-installation from the wind turbine nacelle and a re-winding
of the generator, which takes approximately 1 week. The failure consequences also include
potential secondary damage of the RCC unit as well as of the gearbox. The latter is subject to
excessive mechanical loads (oscillations up to 200-250% of the rated torque) occurring dur-
ing emergency stop in case of failure of the generator windings.
Failures of the generator windings can be a short-circuit due to a failed insulation, or the loss
of conduction, i.e. the occurrence of an open circuit. Short-circuit failure occurs usually due
to melting of the isolation as a result of overheating, due to material degradation (aging) of
the isolation or due to mechanical impact, predominantly caused by vibrations (e.g. due to
bearing damage). The probability of short-circuit failure increases with time according to the
experience of the workgroup. Measures to prevent failures of the isolation are: (1) tempera-
ture measurements in the windings as they are commonly implemented today as part of the
SCADA system and (2) the early detection of bearing failures by means of vibration-
monitoring with a condition-monitoring system (CMS) or, as CMS are seldom installed on
V44 turbines, by manually checking the temperature of the casing as a possibly indicator of
bearing failures.
Open-circuit failure of the generator windings, i.e. a loss of the conduction capability, is
caused by broken windings or failing contacts at connections. The underlying failure cause
41(122)

can be fatigue as a consequence of excessive vibrations or material defects in the copper


conductors. Thus it is also a preventive measure against open-circuit failures to avoid opera-
tion at strong vibrations e.g. by using a CMS. Thermography can be applied to detect initiat-
ing damage in the conductor or contacts, where a reduced cross section gives rise to local
hot spots. An additional method to identify impending failure of this type is the application
of a so-called motor tester during standstill of the generator which enables diagnosis based
on an injected test signal.

3.2.4.2 Bearings

The function of the generator bearings is to keep the generator rotor shaft in its position
with a minimum amount of friction for rotary motion. The occurrence of high friction as a
result of spalling and cracks in the bearings surfaces is considered to be the main failure
mode. The consequences of bearing failure range from partial (due to friction losses) or
complete production losses (when severe bearing damage prohibits further operation) to
comprehensive secondary damage of other components due to excessive vibration and
strong heat generation. The latter implies even a risk of fire.
For the analysed wind turbine model, two main failure causes of generator bearing failure
have been formulated: inappropriate lubrication and bearing currents.
Inappropriate lubrication covers a lack of grease due to e.g. too long lubrication intervals,
but also the use of wrong types of lubricant. Both failure causes result in increased friction
and thus high contact temperature which leads to damage of the inner bearing surfaces. In
addition, inappropriate lubrication also includes over-greasing: too much grease between
the rolling elements and the races will cause the elements to slide instead of rolling, which
results in overheating of the bearing.
Bearing damage accumulates over time and is accelerated by insufficient maintenance. The
corresponding failure characteristic is thus an increasing failure rate. Preventive measures
include a strict compliance with the lubrication scheme which prescribes lubrication of the
generator bearings every six months (part of service today) or the installation of an automat-
ic lubrication system which has been shown to be an effective measure. According to Carlson
et al. (2001), more comprehensive preventive maintenance of generator bearings including a
change of lubricant effectively reduced the frequency of generator bearing failures in Vestas
turbines in the past.
Vibration-based CMS and complementary temperature measurement detect approx. 90% of
impending bearing failures. In this way, secondary damage resulting from excessive vibration
and heat generation can effectively be prevented, but not the bearing damage itself. A major
difference between vibration and temperature monitoring is that vibration-CMS usually pro-
vide a pre-warning time (P-F interval) in the range of several weeks to months while this is
only in the range of hours to days in case of temperature-based detection (see also Figure
19).
42(122)

Figure 19: Intervals between detection of a potential failure P and functional failure F (P-F
intervals) for vibration-based vs. temperature-based monitoring of bearings

The causes of bearing currents in wind turbines are not fully understood today and are sub-
ject of ongoing investigations. Among the suspected causes are winding damage in the gen-
erator rotor, ground currents and unbalanced electrical characteristics in the power
electronics connected to the generator electrical circuits. The current flow through the bear-
ings involves the formation of sparks, which at the contact surfaces have a similar effect as
electric-arc welding; locally, the material is tempered, re-hardened or melted, in the latter
resulting in the formation of small craters [SKF (1994)]. Again, the failure characteristic is
increasing with time due to the accumulating nature of the damage, but also with the cur-
rent amplitude.
Today, a run-to-failure strategy is usually applied in the context of bearing currents. Sched-
uled replacement of bearings after ca. 10 years of operation is a potential preventive meas-
ure against bearing failure due to this cause. Vibration-CMS and temperature measurements
can be applied as discussed above. In order to assess if bearing currents are present in a
wind turbine generator, it is possible to measure the high-frequency electromagnetic emis-
sions originating from the sparks by means of a portable device. In case of detected bearing
currents, a replacement with hybrid bearings can be taken into consideration. In hybrid
bearings, the rolling elements are made from an insulating engineering ceramic material.

3.2.4.3 Other

A component being subject to regular visual inspection during service maintenance is the
terminal panel. This includes mainly a check of the connectors. With this procedure being
common practice, failures of this part has not been observed in the V44 turbine.
43(122)

3.2.5 Gearbox
Wind turbines of the model V44-600kW have been manufactured with either combined
planetary/parallel or solely parallel gearboxes. The turbines owned by Göteborg Energi con-
tain solely the second type of gearbox, so that only the parallel design has been analysed in
the present study.
The function of the gearbox in a wind turbine is the transmission of torque from the rotor to
the generator shaft, providing the desired conversion ratio for speed and torque. The failure
mode of interest is thus the loss of torque transmission capability. Failure of the gearbox can
have severe consequences: in case of complete demolition, parts of the gearbox can fall
down in the tower and constitute a risk for personnel. Oil spill of up to 120l of lubrication oil
contained in the gearbox of the V44 can have an environmental impact. Gearbox failure is
among the failures being afflicted with the longest downtime and thus has a strong impact
on production availability, and it can cause severe secondary damage, e.g. in the main bear-
ing or the rotor shaft.
According to the experience and knowledge of the RCM workgroup, an important failure
cause which is considered valid for all gearbox components (e.g. bearings, gearwheels, and
shafts) during different periods are design problems, i.e. that the components are under-
dimensioned with respect to the loads they are exposed to in reality. This lack of inherent
reliability cannot be improved by means of maintenance but would require design changes.
An additional failure cause of general validity for the different gearbox components are
manufacturing and installation deficiencies, which lead to increased friction or inappropriate
high cyclic loading which leads to damage. Training of technicians for improved quality in
manufacturing, installation, and repair as well as alignment checks for the wind turbine drive
train are thus considered to be important preventive measures in order to avoid premature
failure of the gearbox.
The bearings, gearwheels and the lubrication system are the components being most rele-
vant for gearbox failure and have thus been analysed in detail. Failure of shafts in the gear-
box is considered to occur only as a secondary failure and has thus not been included in the
RCM analysis.

3.2.5.1 Bearings

The function of the gearbox bearings is to keep the shafts in their position while allowing
rotary motion with a minimum amount of friction. Again in this case, the failure mode of
interest is the presence of high friction in a bearing. In the gearbox of the V44-600kW sys-
tem, bearing problems occur mainly in intermediate-speed shaft and high-speed shaft bear-
ings. The main failure causes identified during the analysis are overloading, inappropriate
lubrication including lubricant contamination as well as corrosion.
Overloading of bearings, which is often a consequence of design or installation deficiencies
as discussed above, can lead to material fatigue and plastic deformation in a bearing. The
resulting increased friction and heat evolution lead to additional deterioration, resulting in
44(122)

an accelerating development of the damage. Utilisation of vibration-CMS and temperature-


monitoring are suitable measures to detect damages at an early stage, to prevent secondary
damage and to enable planned replacements, but are often not effective to prevent the
bearing damage from occurring, as discussed earlier.
Inappropriate lubrication in the form of insufficient or over-lubrication or use of a wrong
type of lubricant leads to bearing damage through the mechanisms already discussed in the
context of generator bearings. The same failure mechanism results from a high concentra-
tion of particles in the lubricant to which bearings are more sensitive than the other parts of
the gearbox. The failure characteristic is an increasing failure rate over time.
Suitable preventive measures to prevent these failure causes are the measurement of the oil
pressure and temperature, as well as oil analysis. Oil samples are presently taken once a year
and analysed with respect to the content of additives, particles, and moisture as well as signs
of corrosion. Online particle counters which are available as a complement to existing CMS
are an option to monitor the oil quality continuously and enable earlier countermeasures in
case of observed irregularities. In order to avoid the ingress of particles into the lubrication
circuit, it is recommendable to filter even new oil during an oil change.
Another important cause of gearbox bearing failure is the presence of water or moisture in
the gearbox oil, a factor which is usually strongly dependent on the location of the wind tur-
bine. Moisture can on the one hand lead to corrosion through oxidation of the metallic com-
ponents. In addition, hydrogen ingression into the steel reducing the material strength has
been put forward as a possible explanation for surface failures (see also Ruo and Olver
(2008)). However, there are alternative theories which include interaction with the additives
present in the grease, tribo-chemical effects having influence on the localized crack behav-
iour.
As mentioned above, oil analysis is a suitable measure to identify moisture in the gearbox
oil. An alternative is online moisture-detection with an automatic system, but this is not con-
sidered to be economic on machines in the kW range. A both cheap and effective possibility
to prevent the ingress of humidity into the gearbox is the installation of a filter dryer in the
top casing of the gearbox, which dehumidifies the air flowing into the gearbox during ther-
mal cycles.
The cost-effectiveness of measures for prevention and early detection of bearing damages is
considered to be high as it serves to prevent severe secondary damage to the gearbox. In
this way, the lifetime of the gearbox can be extended, up-tower repair may be possible in-
stead of a significantly more expensive workshop-repair of the gearbox. In case of a neces-
sary replacement, the loss of residual value of the gearbox (e.g. due to the fracture of shafts)
can be avoided.
45(122)

3.2.5.2 Gearwheels

The function of the gearwheels is to transmit mechanical power while converting speed and
torque at the desired ratio. The failure modes that are considered most relevant by the RCM
group and have been chosen for analysis are pitting, scuffing, and tooth breakage.
Micro-pitting (or grey staining) as the onset of pitting is very common on gearwheels and is a
normal sign of usage, see Fig. 20. Pitting describes the failure of a material due to surface
fatigue, being a result of continuous surface stress. It is characterized by removal of material
and the formation of cavities and cracks which can progressively increase in size and destroy
the tooth shape which results in significantly increased friction or even tooth breakage [Lu-
cente (2008)].

Figure 20: Grey staining on an intermediate-speed shaft pinion.


Source: [Wind Turbine Gearbox Inspection Report (2009)]

Pitting can have different causes: in addition to (i) overloading and (ii) metal particles in the
lubricant leading to high local stress levels, also (iii) false brinelling on the teeth evolving dur-
ing standstill of the wind turbine leads to pitting. The failure characteristic of this failure
mode is an increasing failure rate over time. Suitable preventive measures include regular oil
analysis or particle counting to avoid operation with particles in the oil. In addition to oil
analysis in a 1-year interval, it is common practice today to use a magnet installed in the
gearbox to indicate the presence of metal particles. Inline magnetic filtering as an alternative
preventive measure does not only indicate, but also remove metal particles.
Scuffing describes damage on the tooth surface in the form of scratch lines resulting from
welding and tearing of the tooth surface by the flank of the mating tooth. This is caused by
an insufficient (or wrong type of) lubricant film leading to metal-to-metal contact of the in-
teracting gear teeth. In addition, scuffing occurs when the load exceeds the scuffing load
capacity of the oil, a quantity which is usually determined in a standard lubricant test on so-
46(122)

called FZG test rigs and which is strongly dependent on the extreme-pressure additives in
the oil (see also Hargreaves and Planitz (2009)).
Scuffing leads to higher friction and heat generation and the release of metal particles into
the oil, which are a potential source of secondary damage as discussed above.
Preventive measures are lubrication in accordance to the lubrication scheme and with the
right type of oil as well as oil sampling to ensure high quality of the lubricant. In order to
avoid excessive load on the gearbox, it is desirable to avoid emergency stops of the wind
turbine. In general, it is beneficial to minimise the utilisation of the mechanical brake on the
high-speed shaft in order to avoid the resulting high load on the gearbox.
Tooth breakage in wind turbine gearboxes can occur as a secondary damage of pitting or
scuffing when the excessive removal of material has decreased the strength of a gear tooth.
It results in high friction, strong heat evolution, and potentially secondary damage caused by
the metal particles.
Further causes of tooth breakage are overload due to misalignment or design deficiencies,
and fatigue which can in particular result from frequent emergency stops. Again, alignment
check, vibration-CMS for early detection of cracks in a tooth, and securing good lubricant
quality by means of regular oil analyses are effective preventive measures against tooth
breakage.

3.2.5.3 Lubrication system

The function of the gearbox lubrication system is to supply lubricant to the contact areas
between the gearwheels and to the bearings, at right temperature and viscosity, to filter the
lubricant and to provide cooling for the gearbox. The failure mode considered here is that
the system does not fulfil this function and thus does not provide sufficient lubrication.
According to the RCM group, the dominant failure in the lubrication system is too high oil
temperature which results in an insufficiently thin oil film. In the large majority of cases, this
is caused by a defect in the temperature control system, in a few cases by a defect in the
cooling circuit. Preventive measures to counteract this type of failure in the lubrication sys-
tem have not been suggested. Instead, a modification of the control system was proposed:
by introducing lower alarm levels for the oil temperature, a warning alarm could be sent to
the maintenance service provider in good time before the wind turbine is stopped automati-
cally due to a critical level of the oil temperature; in that case, the operation of the wind tur-
bine could be continued at reduced power until the maintenance personnel arrives at the
turbine.
A second important failure of the lubrication system is a too low temperature of the oil,
which in that case has a too high viscosity to ensure appropriate lubrication. This can be
caused by failure of the gearbox oil heating system, but can even occur in case of a function-
ing heating system when the oil is heated too inhomogeneously and mixing of the lubricant
is insufficient. A possible measure to prevent secondary damage resulting from low oil tem-
47(122)

perature would be to measure the temperature at relevant positions and not to permit start-
up of the turbine when these are below a threshold value.
A third relevant failure of the lubrication system is insufficient oil filtration. In the V44-
600kW turbine, the oil is filtered by means of an inline filter through which the complete oil
flow is passed. A blockage of this filter causes a bypass to open unless the blocked filter has
been replaced. In that case, the oil is circulated without filtration.
Today, oil filters are exchanged every 12 months, which is a suitable interval according to the
experience of the RCM group. In order to prevent insufficient filtration, the installation of
offline filters (see Figure 21) or, more generally speaking, of additional stages of filtration is
proposed. Suitable measures to detect insufficient filtration are the installation of an online
particle-counter or the measurement of differential pressure over the inline filter which
could indicate and provide an alert when the filter is blocked. An interesting observation is
that the filter surface is often only partially utilized because air is enclosed in the upper part
of the filter. The introduction of a bleeding point for air in the filter enclosure as it was pro-
posed by the RCM group would enable a better utilization of the available filter surface and
a longer life of the filter.

Figure 21: Inline vs. offline filtration of gearbox oil

A fourth, but less frequent failure of the lubrication system is leakage or loss of oil leading to
low oil pressure. Leakage can occur at a slow rate e.g. at the shaft seals or at badly installed
filters while sudden loss of oil from the lubrication system can e.g. be caused by a rupture of
an oil hose. The only preventive measures in this context would be an improved training of
maintenance technicians to ensure the correct installation of oil filters or the development
of filter designs for foolproof mounting. In order to detect oil leakage or loss, the gearbox is
visually checked at service, i.e. in intervals of 6 months. Suitable measures to detect this fail-
ure at an earlier stage could be the introduction of a pressure measurement in the SCADA
system.
48(122)

3.2.6 Hydraulic system


The hydraulic system comprises hydraulic pumps, accumulators, control valves and oil hoses.
The function of the hydraulic system is (a) to enable braking by means of the mechanical disc
brake on the high-speed shaft as well as (b) pitching of the rotor blades for pitch control and
aerodynamic braking. Since both the mechanical and the aerodynamic brake are activated
through the hydraulic system, a worst-case consequence of hydraulic system failure could be
a loss of braking capability and over-speeding of the turbine, with considerable risk for per-
sons and installations near the wind turbine, risk of environmental impact due to oil spill,
and severe potential damage to the wind turbine. However, this is mainly a theoretical sce-
nario which is prevented by two different safety systems for the mechanical brake and the
pitch system: One of these, a safety valve, releases the pressure from the hydraulic pitch
actuators in case of emergency and in this way allows the blades to be pitched out of the
wind solely by the aerodynamic forces. Secondly, emergency operation of the mechanical
brake is ensured by a hydraulic accumulator with stored pressure.
The consequences of less severe failures of the hydraulic system discussed in the following
include mainly the partial or full loss of electricity production capability, but no risk for per-
sonnel or the environment. Secondary damage to the drive train can results from pitch er-
rors of single blades, which lead to unbalances and thus impose high cyclic load on the drive
train components.
An important failure mode of the hydraulic system is a loss of the blade pitching capability or
a too slow pitching. The most common cause of this is a failure of a control valve due to
wear-out of its mechanical parts. The failure characteristic is thus an increasing failure rate
over the age of the control valve. At present, hydraulic valves are maintained with a run-to-
failure strategy. A possible preventive measure could be a scheduled replacement of the
valves at the age of ca. 10 years, i.e. once during the wind turbine lifetime.
A second failure cause of too slow pitch is the contamination of the hydraulic oil. This is
caused by particles, which result from component wear-out and accumulate in the hydraulic
system. The level of contamination rises over time, so that an increasing failure rate is to be
assumed for this failure cause. It is common practice today to exchange the hydraulic oil
every 4-5 years (dependent on the maintenance service provider). This is considered by the
RCM group to be an appropriate interval.
Another cause of pitch failure is low pressure in the hydraulic system. This can result from
failure of a hydraulic pump, failure of a control valve or of the 24V transformer feeding it, or
failure of an accumulator. Accumulators store energy by means of pressurized gas which is
contained in a rubber bladder. As the rubber ages, the probability of cracks or rupture lead-
ing to a pressure loss in the accumulator increases. After frequent failures of hydraulic
pumps in the V44 system, this component has been preventively replaced with an upgraded
design in the past which has effectively reduced the occurrence of pump failures since then.
As mentioned above, low pressure in the hydraulic system can also be a result of oil leakage,
e.g. due to rupture of a hose.
49(122)

An additional failure mode of the hydraulic system considered in the analysis was the case of
the mechanical brake being continuously activated. This is mostly caused by an error in the
control system and can thus hardly be influenced by maintenance. In a few cases it is caused
by low pressure in the hydraulic system, the possible causes and countermeasures of which
have been discussed above.
The present maintenance practices for the hydraulic system cover a visual check of its com-
ponents and a test of the control valves at service every 6 months, as well as the scheduled
exchange of hydraulic oil every 4-5 years as explained above. More frequent oil changes and
functional tests would be a possible measure to prevent failure, but are not expected to be
cost-effective for the V44 system. Condition-monitoring systems suitable for the hydraulic
systems of wind turbines are not available to date.
50(122)

3.3 Results and discussion: V90-2MW wind turbine

3.3.1 System description

The structure of the Vestas V90-2MW system is shown in Figure 22. The first turbines of this
type were installed in 2004. The V90-2MW series is based on the nacelle of the V80 and the
blades of the V90-3MW high-wind turbine. As the V44, the V90-2MW is an upwind turbine
with three blades and active, electrically driven yaw. The rotor has a diameter of 90m and a
weight of 38t. The nominal rotor speed of 14.9rpm is about half of the rotor speed of the
V44.

Figure 22: Structure of the V90-2MW system;


Source: Vestas product brochure V90-1.8MW & 2.0MW (2007)
51(122)

The Vestas “OptiTip” pitch control system with individual pitching capability for each blade
continuously adapts the blade angle to the wind conditions and in this way provides opti-
mum power output and noise levels. In addition, it serves for speed control, turbine start-up
and stop by aerodynamic braking. Similarly to the V44, an additional disc brake is located on
the high-speed shaft.
In contrary to the V44 turbine, all V90-2MW systems contain gearboxes with one planetary
and two parallel stages, from which the torque is transmitted to the generator through a
composite coupling. A major difference to the V44 system is the generator concept: The
V90-2MW contains a 4-pole doubly-fed asynchronous generator (DFIG) with wound rotor. A
partially rated converter controls the current in the rotor circuit of the generator, which al-
lows control of the reactive power and serves for smooth connection to the grid. In particu-
lar, the applied DFIG concept (Vestas’ so-called “OptiSpeed” technology) allows the rotor
speed to vary by 30 per cent above and below synchronous speed. The electrical connection
between the power electronic converter and generator rotor requires slip rings and carbon
brushes. The generator stator is directly connected to the power grid. (Sources: Vestas Gen-
eral Specification V90-1.8/2.0MW (2005), Vestas product brochure V90-2MW (2007, 2009,
2010))
The system function of the V90-2MW is to provide up to 2MW electric power at 690V and
50Hz to the grid, in a standard operating temperature range of -20…+30°C and at wind
speeds of 4-25 m/s. Again in this case, system level failures of interest in this study are (1)
the complete loss of energy conversion capability or (2) the partial loss of energy conversion
capability of the turbine. As the V44, the V90-2MW wind turbine system has four operating
states, among which RUN is the only one allowing connection to the grid.
For the RCM analysis of the V90-2MW, it is important to note that the series is not fully con-
sistent, i.e. that small changes in design have been implemented in every year of production
[Lindqvist and Lundin (2010)]. Where this is known, it will be mentioned in the context of the
RCM analysis results.
End of November 2010, a number of 124 V90-2MW wind turbines were installed in Sweden
([www.vindstat.nu]), while Vestas reports a total number of 2754 turbines of this type deliv-
ered worldwide until mid of 2010 [www.vestas.com]. A V90-2MW located in Mariedamm
and commissioned in January 2010 is owned by Göteborg Energi. Triventus supervises the
operation of this turbine, the service is performed by Vestas.

3.3.2 Subsystem selection

As in case of the V44-600kW system discussed before, also in case of the V90-2MW system
the subsystem selection for in-depth analysis has been based on statistical analysis and ex-
pert opinion.
Statistical data analysis for the V90-2MW system has been carried out based on data from
[Balschuweit (2009)], which was used with kind permission of Vattenfall. It includes failures
of 57 V90-2MW turbines located in Germany, from the period 2004-2008. Downtime due to
52(122)

external grid failures, inspections, service maintenance and other external causes has not
been taken into account in this analysis. Figure 23 shows the distribution of average down-
time per wind turbine and year over the concerned subsystems.

Figure 23: Results of failure data analysis for the V90-2MW system;
Data source: Balschuweit (2009)

Table 2 summarises the results of the statistical analysis and of the assessment of expert
opinion by means of questionnaires.

Table 2: Criticality of wind turbine subsystems with respect to failure frequency and downtime,
according to expert judgment and statistical data analysis (in order of descending priority)

V90-2MW

Expert judgment Data analysis

Failure frequency Downtime per failure Downtime per year and turbine

1. Gearbox Gearbox Generator incl. converter

2. Generator Generator Rotor

3. Converter Converter Drivetrain incl. gearbox

4. Hydraulic system Rotor Control and protection system


53(122)

Based on these results, the gearbox, the generator, the converter, the yaw system, the pitch
system as well as the rotor blades of the V90-2MW were selected for analysis of their domi-
nant failures within the RCM study. In addition to the frequency and criticality of failures in
terms of resulting downtime or secondary damage, an important criterion for the selection
of components and failures for deeper analysis was the question if these could be effectively
influenced by means of preventive measures.
The main results of the RCM analysis are presented in the following. Due to the fact that a
large number of failure modes, failure causes, and suitable preventive measures identified in
case of the V90-2MW are identical with those discussed before for the V44 system, the re-
sults will be limited to those being different for the two turbines or going beyond the analy-
sis of the V44.

3.3.3 Gearbox
In V90-2MW wind turbines, gearboxes from several different manufacturers are applied. In
addition to two parallel stages, all of these include a planetary stage which allows a signifi-
cantly more compact design than a parallel-stage gearbox of the same rated power.

3.3.3.1 Bearings

Similarly to the V44, most gearbox bearing failures in the V90-2MW occur in bearings on the
high-speed shaft (HSS) and intermediate-speed shaft (IMS). In some cases, up-tower repair
of these bearings is possible. However, any failure in the planetary stage requires removal of
the gearbox from the nacelle for external repair. In contrary to the V44, in which only a sin-
gle bearing temperature (HSS bearing) is measured, there are four temperature measure-
ment positions for the gearbox bearings of the V90-2MW.
One of the preventive measures suggested in case of the V44 to prevent a lack of lubricant is
implemented as standard in newer models of the V90-2MW series: It now provides an ana-
logue measurement of the oil pressure in the lubrication system and in this way facilitates
both an early detection of problems as well as the fault diagnosis itself.
Endoscopy is a technique which is increasingly applied on wind turbines to assess the condi-
tion of the larger gearbox bearings as well as of the gearwheels. Combined with digital imag-
ing technology, it allows even a remote assessment of the gearbox condition by gearbox
experts. Endoscopy is not only applied in case impending failure is detected by vibration or
temperature measurements, but also without any indication of irregularities. The V90-2MW
as well as the majority of other contemporary wind turbines provides an opening in the
gearbox casing, which allows easy access for endoscopy. Due to the compact design of gear-
boxes, not all components can be inspected with this technology, e.g. on the V90-2MW tur-
bine, endoscopy inspection of the of HSS bearings is not possible. The application of
vibration-CMS is considered to be state of the art on MW turbines. However, vibration moni-
toring and vibration-based diagnosis of planetary gearbox stages is still challenging at pre-
54(122)

sent and an improvement of condition-monitoring technology for this purpose is subject to


intensive development activities today.

3.3.3.2 Gearwheels

In spite of the different gearbox designs applied in V44-600kW and V90-2MW, the dominant
failure modes, failure causes and preventive measures were found to be identical with those
identified for the V44 system (see Section 3.2.5.2).

3.3.3.3 Lubrication system

In contrary to the V44-600kW, the lubrication system of the V90-2MW consists of three sep-
arate oil circuits: (a) a main cooling and lubrication circuit with a mechanical pump, (b) an
additional cooling circuit with an electric pump, which starts operation when the oil temper-
ature reaches a threshold value, and (c) an offline filter circuit. The lubrication system of the
V90-2MW contains 330-405l of oil. The oil filtration occurs by means of a coarse (25 μm)
inline filter in the main lubrication oil circuit as well as with the finer offline filter (3 μm). In
addition to particles, the offline filter even removes moisture contained in the lubricant.
Magnetic filters are not applied in the V90-2MW system. If a particle counter is installed, this
is located in the offline circuit.
In contrary to the V44-600kW, an analogue pressure measurement in the main lubricant
circuit is standard on the V90-2MW. The result of this measurement is accessible online for
the maintenance service provider.
A comparison with the results of the RCM analysis of the V44 shows that several preventive
measures suggested for the V44 gearbox have been implemented in the V90-2MW system.

3.3.4 Generator
The concept of the asynchronous generator applied in the V90-2MW differs considerably
from that in the V44 system. Apart from its significantly higher rated power of 2MW, the
rotor current is no longer controlled by means of variable resistance in an RCC unit but by a
partially rated converter (Vestas Converter System, VCS). The RCM analysis results for the
converter unit will be presented in a separate section following those for the generator.

3.3.4.1 Windings

The RCM analysis of the V90-2MW has provided identical results as those presented above
for the V44-600kW system. With respect to winding failures of the generators, a difference
between the two wind turbines is that in a V90-2MW generator, the rotor windings are ac-
cessible for diagnosis measurements e.g. using a motor tester. This facilitates both the de-
tection of impending winding failure as well as the fault search for the rotor windings.
55(122)

3.3.4.2 Bearings

A major difference between the analysed turbines is the type of generator bearings: In the
V90-2MW, hybrid bearings are applied [Vestas product brochure V90-2MW (2010)]. In this
way, any electric current through the bearing being a potential source of bearing failure is
prevented. A second difference is that, in contrary to the V44, the temperatures of the gen-
erator bearings are measured in the V90. Vibration-CMS has been found to be an effective
measure to detect bearing damage in the V90 at an early stage.
Apart from this, the RCM analysis of V90 generator bearings has provided results identical to
those presented for the V44 system.

3.3.4.3 Other

In the past, the slip rings and carbon brushes providing the electrical contact to the genera-
tor rotor were a source of frequent failure. Carbon dust originating from the wear of the
brushes caused electric spark-over which lead to damage of the slip ring. This problem has
been solved successfully by means of a suction system that continuously removes the carbon
dust during operation of the wind turbine. In addition, the quality of the brushes could be
improved significantly. In this way, the lifetime of the slip rings could be significantly en-
hanced to approx. 2 years today. Both the slip rings as well as the carbon brushes are re-
placed depending on their condition, which is checked during service every 6 months.
Two novel approaches which are, to the knowledge of the RCM group, not applied in V90
generators yet provide the possibility to utilise the full useful life of the rather expensive
carbon brushes while preventing damage from over-worn brushes: electrical contacts in
brushes can provide an alert signal when the brush thickness reaches a threshold value; an
alternative approach is the use of cameras for continuous monitoring of the brush thickness.
The service maintenance of the V90 generator today usually includes greasing of the bear-
ings as well as condition check and possibly replacement of carbon brushes and slip rings.
The maintenance of the other generator components follows a run-to-failure strategy.

3.3.5 Converter
In a variable-speed wind turbine with DFIG like the V90-2MW, the function of the converter
is to feed the generator rotor circuit with an electric current of desired amplitude and fre-
quency in order to ensure the desired output of active and reactive power from the stator to
the grid. Depending on the operating speed, the converter either transmits electric power
from the generator rotor circuit to the grid (in super-synchronous operation) or vice versa (in
sub-synchronous operation), see e.g. Manwell et al. (2009).
56(122)

Figure 24: Partially-rated converter in a variable-speed wind turbine with DFIG [Petersson (2005)]

As shown in Figure 24, the converter consists of a rotor-side and a grid-side converter in
back-to-back configuration. The rotor-side converter controls the current to the rotor
whereas the grid-side converter controls the DC link voltage [Nilsson (2010)]. In addition to
these two power electronics units with IGBT switches, the converter subsystem includes
corresponding microprocessor control units.
The converter is water-cooled by means of a pump-driven cooling circuit. High failure rates
of the converter systems in V80 and early V90-2MW turbines could be reduced but are still
an important source of failure in newer turbines from the V90-2MW series. Salty environ-
ments, grid disturbances and insufficient cooling are considered by the RCM workgroup to
negatively impact the failure frequency of the converter system. The consequences of con-
verter failure are mainly limited to production loss. In case of failure, the turbine is stopped
smoothly by means of the aerodynamic brake so that excessive loading of the drive train
components does usually not occur in this case and potential secondary damage of this is
avoided.
The failure modes of the converter identified in the present RCM study are (a) open circuit,
i.e. an interruption of the electrical connection, and (b) a malfunction of the rotor current
control.
A common cause is the failure of the microprocessor units, a problem which in some cases
could be addressed with software updates, but usually not by maintenance measures. An-
other cause is failure of the power electronic components with potential causes ranging
from high temperature, vibration, moisture or condensation problems and component aging
to disturbance from the electric power grid (i.e. voltage interruptions, voltage dips or fre-
quency variations). Based on that, possible preventive measures are those reducing or moni-
toring the given failure causes, such as reduction of vibration in the power electronic unit by
design, vibration-CMS or temperature monitoring.
57(122)

3.3.6 Yaw system


The function of the yaw system is to continuously align the wind turbine rotor with the wind
direction and to hold this position. The yaw bearing system is a plain-bearing system with
built-in friction; the friction is provided by disc springs moving on a slew ring. The four elec-
trical yaw drives are 2.2 kW asynchronous motors with electrical brakes. These rotate the
nacelle on the top of the tower by means of planetary gears. Lubrication of these gears is
provided by an automatic lubrication greasing system. (Sources: Vestas General Specification
V90-1.8/2.0MW (2005), Vestas product brochure V90-2MW (2007, 2009, 2010))
Yawing of the wind turbine is possible as long as at least two of the four yaw motors are
functioning. The failure mode of interest here is the case in which no rotation of the nacelle
can be provided by the yaw system. The consequences of yaw-system failure are limited to
production loss.
According to the experience of the RCM group, the by far most common failure cause in the
V90-2MW yaw system is fracture of the gears on the yaw motors. This concerns mainly the
shafts, but also the teeth of the gearwheels. Design deficiencies are considered to be the
major underlying cause of this failure: Insufficient friction against rotational movement on
the slew ring causes a large portion of the torque to be transmitted to the gears and the
electrically-braked motors. This leads to overloading of the yaw gear. The problem is aggra-
vated when one or several motors or gears are out of operation. In this case, the full torque
acts on the remaining gears and motors, which significantly increases the risk of overloading.
Due to this failure mechanism, a suitable preventive measure would be testing the disc
springs to ensure that these provide sufficient friction on the slew ring. To avoid fatigue of
the yaw gears, it would be effective to regularly measure the clearing between the gear
teeth. Alternatively, it is suggested to measure the amplitude of rotational movements when
the turbine is supposed to keep a fixed position in order to detect too large clearances in the
yaw gears. At present, an alarm is generated only when the turbine is supposed to yaw and
does not. Condition-monitoring systems for the yaw system are not available today and are
not considered to be cost-effective.

3.3.7 Pitch system


The function of the pitch system is to control the pitch angle of the rotor blades in order to
enable the conversion of kinetic energy of the wind to mechanical energy. A failure of the
pitch system causes the wind turbine to stop and thus results in production loss. In addition,
pitch errors of single blades can cause an unbalance in the rotor and cause high loading of
the drive train components. In the worst case, it is even possible that a wind turbine blade
hits the tower if the pitch system does not work correctly, as it has occurred in early
V90-2MW turbines.
Failure of the pitch system can be caused by failure of the blade bearings or failure of the
hydraulic system. The latter has been analysed in detail for the V44 turbine and the failure
58(122)

mechanisms are mostly identical in case of the V90-2MW. An additional cause of failure of
the hydraulic system not yet given in the analysis of the V44 system is water or other con-
tamination in the hydraulic oil, which is assumed to accelerate the wear-out of the hydraulic
control valves. A possible measure to counteract this could be the offline filtration of the
hydraulic oil. Alternatively, a measurement of the moisture content in the hydraulic oil
would make it possible to take action when the presence of water is detected. The hydraulic
oil in V90-2MW turbines can usually be used only for three years today instead of the five
years it would be supposed to last.
Finally, apart from failures in the hydraulic system, a wrong pitch angle can also be caused by
a malfunctioning anemometer.

3.3.8 Rotor
The function of the rotor is to convert kinetic energy of the wind to mechanical energy and
to transmit this to the main shaft. Possible failure modes are that the rotor can either not
convert the right amount of energy, that it structurally fails or that it collides with the tower.
A very common cause of reduced energy conversion capability is the formation of ice on the
rotor blades. De-icing systems (e.g. using a thin film covering the whole blade surface) are
currently under development, but are not part of the standard equipment of V90-2MW tur-
bines. Other systems that do not prevent but can detect ice on wind turbine blades (e.g. la-
ser-based systems) are available on the market today. These allow a turbine shutdown for
safety purposes or can enable pitch angle adaption to improve the blade aerodynamics in
the presence of ice.
The structural failure of wind-turbine blades often originates in cracks. In Sweden, cracks in
rotor blades are predominantly caused by lightning strike. In case of the V90-2MW, a light-
ning detection system belongs to the standard equipment of the turbine. In this way, a blade
can be inspected after it has been hit by a lightning strike, possible cracks can be discovered
at an early stage and can be repaired before they reach a critical size. In addition, it is possi-
ble to monitor the structural health of wind turbine blades with suitable condition-
monitoring system, as discussed in Section 2.6.1.2.
A known problem of wind-turbine rotors, which is not limited to V90-2MW turbines, is the
loosening of bolts connecting the blades with the hub. At present, it is common practice to
check 10% of all bolts at service. However, taking into account the large threat for life and
property and the severe damage to the wind turbine that results from blade tear-off, it
would be desirable to apply a system that can automatically detect loose bolts at an early
stage. Suitable technology for this purpose is available on the market already today. An ex-
ample given in this context is a microphone installed in the hub of Enercon turbines which
acoustically detects loose bolts.
59(122)

3.4 Wind turbine service and maintenance practices today


In addition to the subsystem- and component-specific results described in the previous sec-
tions, the RCM workgroup discussions have provided comprehensive information about the
present common practices in wind turbine maintenance, which is summarized in the follow-
ing.
Wind turbine maintenance consists of service maintenance in regular intervals of usually 6
months (i.e. time-based preventive maintenance), of corrective maintenance after failure, as
well as of condition-based maintenance. The analysed V44 system is in average visited twice
per year for service maintenance and approximately additional two times per year for repair.

3.4.1 Service maintenance


Service is usually carried out according to the turbine-specific service plan provided by the
wind-turbine manufacturer. Experience has shown that the better the schedules and plans
for service maintenance are kept, the better is the reliability and availability of the wind tur-
bine; already an extension by 2-3 months has been found to cause a noticeable increase in
the frequency of failures. This confirms the adequacy of the present service intervals and
shows that an extension of the service intervals is not a suitable approach to reduce total
costs. A lack or bad execution of maintenance has been found to lead to low availability as
well as to secondary damage, allowing the conclusion that not to perform maintenance in
the right way can result in high consequence costs.

3.4.2 Corrective maintenance


In case of failure in the wind turbine, the maintenance service provider is usually remotely
informed and provided with a descriptive alarm or error text. In case of Triventus, SCADA
data of the operated turbines is available remotely. The alarm text supports in fault localiza-
tion, but requires a lot of experience to allow a diagnosis of the failure lying beyond the
message. Remote diagnosis is in practice often complicated by a large number of error mes-
sages being triggered at the same time. If applicable, remote test starts are carried out. The
objective is to obtain a precise remote diagnosis, which facilitates the choice of the right
spare parts and tools for the technician and in this way allows a fast repair. In many cases,
however, it is necessary to visit the turbine for local fault search, and the cost for the often
lengthy fault diagnosis many times exceeds the cost for the required spare part. The fact
that some types of faults can only be detected when sufficient wind for test operation is
available (as in case of the RCC failures in V44 turbines discussed earlier) can additionally
prolong the time to repair.
60(122)

3.4.3 Condition-based maintenance


Condition-based maintenance is carried out based on visual inspections and tests of compo-
nents carried out during service maintenance, on the analysis of oil samples taken at these
occasions, and also based on CMS where such systems are installed. On the relatively old
V44 turbines, CMS are not commonly installed; the “Elida” turbine of Göteborg Energi, which
has been equipped with a CMS prototype from SKF since 2001, is an exception in this re-
spect. On V90-2MW turbines, CMS are not part of the standard equipment provided by the
wind turbine manufacturer, but are in practice installed in virtually all turbines of this type.
Due to the fact that the interpretation of CMS data for fault detection and diagnosis purpos-
es requires specific expert knowledge, CMS data is usually monitored from dedicated moni-
toring centers, in the case of the V44 “Elida”, the V90-2MW in Mariedamm and other SKF
WindCon equipped turbines this is carried out from the SKF Wind Industry Service Centre
(WISC) in Hamburg. Although to some extent accessible through an observer software, CMS
data is neither directly used by Triventus as the maintenance service provider nor by Göte-
borg Energi as the owner of the monitored wind turbines. The common procedure is that the
maintenance company is informed by the monitoring centre in case the experts discover
irregularities indicating damage or impending failure in the wind turbine, see also Section
2.6.2.

3.4.4 Inspections
Inspections can be distinguished into status inspections and regulatory inspections.
Status inspections are normally ordered by the wind turbine owner. They are usually carried
out either when a wind turbine is sold or with the purpose of assessment of the status quo
when a new maintenance service provider takes over the responsibility for a wind turbine.
The comprehensive condition-assessment of wind-turbine components includes thermogra-
phy, a technology which is increasingly applied in wind turbines today (see also Section
2.6.1.3) and which is internally used at Göteborg Energi even independently from the com-
prehensive inspections named above. Thermography is applied for assessment of the elec-
trical system, the gearbox and the generator.
In contrary to status inspections, regulatory inspections are mandatory. These concern in
particular safety-critical equipment like e.g. cranes or pressure vessels. Regulatory inspec-
tions are either performed within a Service and Availability Agreement (SAA) with a mainte-
nance service provider or carried out in the regime of the wind turbine owner.

3.4.5 Identified challenges


Based on the discussions within the RCM workgroup, several main difficulties and fields for
potential improvement with respect to operation and maintenance of wind turbines have
been identified:
61(122)

A central challenge is the lack of information flow between the many different parties in-
volved in wind turbine O&M – the wind-turbine manufacturer, the owner, the maintenance
service provider and, in case of turbines with CMS, the monitoring centre. On the one hand
this is often due to a reluctance against sharing information and documentation with the
other parties; on the other hand, it is caused by the variety and the incompatibility of the
present systems used for the reporting of failures and maintenance actions. An important
step to solve this problem would be the standardization (and partial automation) of wind-
turbine failure and maintenance reporting. Such standards have e.g. been developed by the
German VGB based on the RDS-PP designation system. The broad introduction of Computer-
ized Maintenance Management Systems (CMMS, e.g. Maximo, SAP PM, EAM etc.) already
applied in other industries is an important step to merge all relevant information in a single
database, including the information which component failed, which type of failure occurred,
possibly due to which cause, and which maintenance action was taken.
Another difficulty identified concerns the manufacturer-independent companies providing
wind turbine maintenance services: Due to the high prices (for both man-hours and material
or spare parts) that are requested by some manufacturers for their maintenance services,
there is a market for alternative wind turbine service providers like e.g. Triventus and SKF.
However, for these companies the access to detailed information about the wind turbines is
difficult since this is often kept confidential by the manufacturer. In addition, there are limi-
tations in the access to spare parts and special tools, which may be bought only by the man-
ufacturer or which are extremely expensive for parties requesting only small numbers of
pieces. The latter problem could be overcome in the long run by standardising wind-turbine
components, which would introduce a higher degree of competition in this market and facili-
tate the access to spare parts at appropriate prices.
A third identified challenge is the large number of new personnel in wind turbine mainte-
nance with limited experience in this field, being a side-effect of the strong growth of wind
turbine installations. During the RCM analysis it was pointed out that correct installation and
de-installation routines and in particular the proper alignment of components have a strong
impact on the reliability of wind turbines and that there was a need of enhanced training
and education of maintenance personnel in this respect.
A fundamental problem identified during the RCM study is that maintenance decisions are
usually made with the aim of a short-term cost minimisation, not with a focus on long-term
minimisation of total life cycle cost. At present, the RCM group perceives a difficulty in prac-
tically justifying the installation of additional equipment for prevention of failures in wind
turbines (i.e. to prove that they lead to additional profit/less total cost) since a quantification
of the benefit of such installations is challenging. This issue is addressed by means of the
data-based, quantitative methods being further developed and applied in the WindAM
group and for which the RCM study described here provides a comprehensive knowledge
basis.
62(122)

3.5 A quantitative model for maintenance strategy assess-


ment
Based on the results of the above described RCM study, a quantitative wind-turbine reliabil-
ity and maintenance model has been developed in a thesis project, which was carried out in
close cooperation with Göteborg Energi and supervised within the WindAM-RCM project: In
his thesis “A comparison of wind turbine life-cycle costs and profits resulting from different
maintenance strategies”, B. Kerres developed a MATLAB model capable of simulating differ-
ent maintenance scenarios by means of Monte-Carlo simulation. For this purpose, the wind-
turbine has modelled as an - in terms of reliability - serial connection of the most critical tur-
bine subsystem identified in the RCM study, see Figure 25. Two different deterioration pro-
cesses, namely the binary and the delay-time deterioration process, have been used to
model these subsystems.

Figure 25: Serial structure of the wind-turbine reliability and maintenance model developed for
maintenance strategy assessment [Kerres (2011)]

The model has been implemented for the case of the Vestas V44-600kW. The component
reliability models required for this purpose have been derived from turbine specific data
from the previously named Swedpower database as well as from data available at Göteborg
Energi.
Different maintenance strategies and scenarios, covering run-to-failure strategies, regular
inspections, preventive component replacement and online condition-monitoring, have
been simulated and the resulting probability distributions of the Net Present Value and
overall turbine availability have been compared. As one aspect of particular interest, the
cost-effectiveness of CMS and its dependency on factors like the turbine size, the energy
price has been investigated.
A detailed description of the model and the results obtained for the case of the V44-600kW
can be found in [Kerres (2011)].
63(122)

4 CMS-data based prognosis of the residual life of


wind-turbine components
In Chapter 2 of the present report, the state of the art in wind-turbine condition-monitoring
and the current practices regarding the utilisation of CMS data in the maintenance have
been summarised. Based on the results of the literature study and on the discussions with
the project partners, CMS-data based prognosis of the residual life of wind-turbine compo-
nents has been identified as main focus in the CMS-related part of this project. For this pur-
pose, the tools and the comprehensive experience required for the first steps of data
acquisition, processing and diagnosis have been provided by experts at project partner SKF’s
condition-monitoring centre WISC (Wind Industry Service Centre) in Hamburg. Building on
top of this, the research work within the WindAM-RCM project has focused fully on the
prognostic step, the development of a model that provides a quantitative prediction of the
residual component life based on the age and the CMS-data history of this component.

4.1 Data acquisition and signal processing


The data analysed in this project has been measured by means of SKF’s condition-monitoring
system WindCon. This is a commercially available and widely applied CMS for vibration-
based monitoring of wind-turbine drive-train components (main bearing, gearbox, genera-
tor). Figure 26 shows the positions of the eight piezoelectric, usually glue-bonded accel-
erometers and the tachometer used for data acquisition. Optionally, additional input signals
can be included by WindCon, e.g. the generated electric power or signals provided by a par-
ticle counter monitoring the gearbox oil. Via an Ethernet connection, the data is transferred
to a central data server.

Sensor configuration in WindCon

Accelerometers:
1 Main bearing, radial
2 Main bearing, axial
3 Gearbox, main shaft
4 Gearbox, intermed.-speed stage (2x)
5 Gearbox, high-speed shaft
6 Generator, drive-side (DS)
7 Generator, non-drive side (NDS)

8 Tachometer

Figure 26: Sensor positions, input and output signals of the CMS WindCon
(adopted from [Heintz (2010)])
64(122)

By processing the vibration signals measured by the 8 accelerometers, the WindCon system
provides in total 38 different quantities for monitoring and diagnosis, which cover both
velocity, acceleration and envelope signals.
Through the summation over specific, damage-indicative parts of the measured frequency
spectrum, so-called trend values ares calculated, see Figure 27.

Figure 27: Typical trend development of a machine (adopted from [Kolerus (1995)]

Together with diagnosis values, which are based on a summation of single, failure-indicative
frequency components, these offer the possibility to analyse and visualise selected
properties of a component’s vibration spectrum over a longer period of time. Due to the
fact that machines can have very different basic noise levels even in absence of a damage,
not only the momentary absolute value of a CMS signal but in particular its progress over
time are of importance for the successful detection of an evolving fault.
Upon the detection of deviating machine behaviour based on the sum signals, stored full
frequency spectra are usually used for a more detailed diagnosis. This can provide
information about the location of the damage.
Due to the fact that the vibration levels vary to some extent with the rotational speed of the
high-speed shaft, usually two different operating states of a wind turbine are defined: The
range between the minimum and the maximum rotational speed is divided into two classes.
The one with the lower rotational speeds is denoted “low-class” (“lc”), while the one
describing the higher rotational-speed range is denoted “high-class” (“hc”).
65(122)

4.2 Component selection and analysed data


While the methodology for CMS-data based lifetime prognosis to be investigated in this pro-
ject is of generic nature, it is implemented for the specific case of one wind-turbine compo-
nent in order to demonstrate its feasibility. The following reasoning was applied for the
selection of the component to be modelled:
The RCM analysis described in Chapter 3 has revealed that bearing failure dominates among
the failures in both wind-turbine gearboxes and generators. In addition, vibration, which
often originates from damaged bearings, has been identified to be a common cause of sec-
ondary damage. This shows the particular importance of condition-monitoring of the bear-
ings in wind-turbine drive trains.
As a second criterion for the component selection, it was ensured to choose a proven field of
vibration-monitoring, with a known particularly strong correlation between the component
condition and the corresponding CMS signals. By selecting a case with little uncertainty in
the fault detection, it was intended to keep the uncertainty in the resulting mathematical
model as small as possible. Finally, the available CMS should cover a considerable number of
failures in order to provide the broadest possible statistical basis for the model develop-
ment.
As a result, the generator bearings have been selected for detailed analysis and the devel-
opment of a prognosis model in the scope of this project. Four CMS signals have been
provided for analysis, two from the drive-side generator bearing (DS, sensor position 6 in
Figure 26) and two from the non-drive side generator bearing (NDS, see sensor position 7):
1) trend DS hc
2) trend DS lc
3) trend NDS hc
4) trend NDS lc
All of these signals are sum values resulting from envelope analysis (frequency range:
500-10000Hz), which allow a particularly early detection of damage in bearings and
gearwheels (see Section 2.6.1.1 about envelope analysis).
Figure 28 shows an example of a wind turbine trend-data history: After the start of online
condition-monitoring in June 2007, low trend levels (<1.5gE1) are observed over a period of a
quartal. From the end of 2007, a gradual increase of the DS trend value indicates an evolving
damage in the DS generator bearing. Following a notification to the operator issued by the
monitoring centre in the form of an event report, the DS bearing is replaced in January 2010.
As observed in Figure 28, the stiff coupling between the DS and NDS sensors leads to a slight
increase of the trend level also at the undamaged NDS bearing.

1
All trend values used in the present work have the unit gE. Herein, g denotes the acceleration of gravity, while
E indicates that the value results from enveloping analysis.
66(122)

trendWT220 (green: hc, red: lc)

trend DS
4

0
Jan07 Jan08 Jan09 Jan10 Jan11 Jan12

6
trend NDS

0
Jan07 Jan08 Jan09 Jan10 Jan11 Jan12

Figure 28: CMS trend data history from the drive-side bearing (DS, upper graph) and the non-drive
side bearing (NDS, lower graph) of a wind-turbine generator

The analysed sample covers data from a fleet of 2022 wind turbines of identical model (iden-
tical turbine manufacturer and rated capacity), with the only exception that they comprise
two different rotor diameters (referred to as wind-turbine type A and B in the analysis). The
turbines are equipped with generators by two different manufacturers, i.e. each turbine
contains a generator by either manufacturer C or D. All generators are equipped with deep-
groove ball bearings (see also Figure 29) of identical make but with a slight difference in di-
mensions, referred to as bearing types E and F in the following.

Figure 29: Deep-groove ball bearing applied in the investigated wind-turbine generators.
Sources: left [Kugellagerzentrale (2012)], right [Agrolager (2012)]

2
From the original sample covering 217 wind turbines, 14 with a monitoring period <0.5 month were excluded.
In addition, a single turbine with hybrid instead of standard generator bearings was excluded from the analysis
due to its expected deviating reliability characteristics resulting from the different (ceramic) material used in
hybrid bearings.
67(122)

Table 3 summarises the main properties of the data sample analysed in this project. In addi-
tion, Figure 30 shows the distribution of wind-turbine age in the fleet, indicating that the
majority of turbines has been in operation for 4-5 years at the time the CMS-data was ob-
tained. It should be noted that online condition-monitoring is usually started several weeks
to months after the commissioning of a wind turbine. In the present sample, the monitoring
started within the first year of operation in 157 cases (77%), while it started >3 years after
commissioning in only 10 cases (5%). In total, CMS-data is available from a portion of 77% of
the operational years of the turbine fleet.

Table 3: Properties of the analysed data sample

Number of wind turbines 202

Wind-turbine age 8 months – 8.1 years (on average 3.6 years)

CMS-data histories 3 months – 4.5 years

Total CMS-monitored time 568 operational years (average 2.8 yrs/turb.)

Total number of DS bearing replacements 9

Total number of NDS bearing replacements 2

Resulting number of generator-bearing 202+9 = 211 (DS)


vibration histories 202+2 = 204 (NDS)

Due to the fact that the data contains only two cases of bearing replacement on the non-
drive side (NDS) of a generator during the whole monitored time, the analysis will be limited
to data concerning drive-side (DS) bearings in the following.

80
No. of wind turbines [-]

60

40

20

0
1 2 3 4 5 6 7 8
Wind-turbine age [years]

Figure 30: Distribution of turbine age in the analysed sample of 202 turbines
68(122)

In the context of reliability analysis, bearings are classified as non-repairable systems, i.e.
after a failure, they are replaced with a new component. The term “failure” in the context of
bearings does not necessarily mean a decomposition of the bearing, but also includes an
unacceptably bad technical condition which e.g. involves the generation of secondary dam-
age.
The classification of bearings as non-repairable systems implies that their reliability analysis
is to be based on the bearing age (i.e. the component age, also called “local age” in reliability
analysis), not on wind-turbine age (“global age”). In spite of the fact, that wind turbines do
not operate continuously, calendar age is used in the present analysis instead of working
age, as the latter could not have been reconstructed with sufficient accuracy from the avail-
able data.
As indicated in Table 3, a total number of 211 DS-bearing vibration histories is provided in
the data sample. Of these, 9 histories end with a failure while 202 histories are right-
censored, i.e. 202 DS-bearings in the considered turbine fleet had not yet failed at the end of
observation (when the vibration data was made available for analysis).
Due to the probabilistic nature of failure, the lifetime of components can be described with a
random variable T. The sample thus provides:
• exactly observed lifetimes Ti = Ci, with Ci: bearing age at failure, and
• 202 right-censored observations with Tj > Cj with Cj: bearing age at the end of observa-
tion

4.3 Main assumptions


The following assumptions are made for the subsequent data analysis and model develop-
ment:
(1) All DS bearings in the sample have identical lifetime distribution and can thus be
described by a single reliability model. This assumption will be quantitatively justified
in a later section of this report, see Section A.4 in the Appendix of this report.
(2) Only damaged bearings are replaced. In the present work this means in particular
that no opportunistic replacement of DS bearing is performed in case of a corrective
NDS bearing replacement. According to discussions with the CMS analysts at SKF
WISC, this is a reasonable assumption due to the fact that the replacement of a DS
bearing together with a NDS bearing would considerably increases the amount and
duration of the maintenance work compared to the sole replacement of a NDS bear-
ing.
(3) All bearing replacements are considered to be preceded by failure of the concerned
bearing. The last moment of monitored operation before the shutdown of a turbine
for the replacement work is regarded as the time of failure. It should be noted that
this is a strong assumption. It is, however, an unavoidable assumption if the de-
scribed data is to be used as the basis for the development of a reliability model, be-
69(122)

cause there is no information available about the real condition of bearings at the
time of replacement.
(4) All bearing replacements correspond to a significant step-decrease in the CMS
trend level. This assumption is justified by the high detection efficiency of generator-
bearing condition monitoring process already stated in Chapter 4.2.
(5) No generator bearing was replaced between the commissioning of each turbine
and the start of the online condition-monitoring. At the start of observation, the
component age of the bearings can therefore be equated with the turbine age. Due
to the limited age of the turbine fleet and the above stated start of condition-
monitoring during the first year of operation, this assumption is considered a reason-
able approximation of the real situation.

4.4 Factors influencing the vibration level


In condition-monitoring, one of the key objectives is to draw, from the CMS signals, conclu-
sions about the condition of the monitored component. In wind-turbine monitoring, this task
is significantly complicated by the interrelation of both the component condition and the
CMS signals with the highly variable operating environment (e.g. with respect to the ambient
temperature) and operating parameters like rotational speed or generated power, see also
Figure 31: The CMS signal does not only depend on the component condition but is to some
extent affected by the operating environment and operating point; at the same time, also
the component condition itself is influenced by the operating environment and the operat-
ing-point dependent stress on the turbine.

Figure 31: Interrelation of the CMS signals, component condition and external factors complicating
the monitoring task

From the data analysed in the present project, two factors influencing the vibration level can
be identified: the ambient temperature and the rotational speed.
70(122)

4.4.1 Ambient temperature


Figure 32 shows a cyclic increase of the CMS trend values during the winter months, which is
observed in the data from several wind turbines in the sample. This cyclic increase can most
likely be explained with the impact of low ambient temperatures. The temporary increase in
vibration level could be due to e.g. an enhanced vibration transfer from the location of the
damage to the sensor at low temperatures, or it could be caused by a lower viscosity of the
grease having a negative impact on the bearing lubrication. Due to the fact that the vibration
level returns to the previous levels during spring, summer and autumn indicates that the
impact of ambient temperature on bearing condition is low or even negligible. The impact of
ambient temperature has thus not been investigated further within this project.

Figure 32: Examples of cyclic seasonal variations of the CMS trend signals

4.4.2 Rotational speed


Together with every vibration data point and its timestamp, the WindCon system provides
the rotational speed in the corresponding moment of measurement (see Figure 26). This
allows a quantitative investigation of the relation between the rotational speed of the gen-
erator shaft and the CMS trend data. If it was possible to identify any systematic relation
between the two quantities, this would offer the opportunity to perform a speed-dependent
“correction” or normalisation of the vibration data and to reduce its undesired affection by
the rotational speed (as one of the key operating parameters referred to in Figure 31).
Plots of the trend level over the corresponding rotational speed have been generated for the
whole fleet of turbines. Figure 33 shows these for two turbines from the fleet. For the sake
of conciseness, data from the two classes of rotational speed, low-class (lc) and high-class
(hc), have been combined in the plots.
71(122)

While the turbine WT40 in Figure 33 (left) displays an almost linear relationship between the
rotational speed and the trend level, the case of WT28 in Figure 33 (right) as well as the ma-
jority of other plots (covering a wide variety of shapes) prove this to be an exception: From
the analysed data sample, no common systematic relation between these two quantities
could be identified. The initial proposal of data normalisation by means of the rotational
speed signal therefore has to be abandoned.

Figure 33: Plot of the CMS trend level over the rotational speed of the generator for two exempla-
rily chosen turbines

4.5 Data reduction


Due to the vast amount of vibration CMS data accumulated over the life of a monitored tur-
bine, it is necessary to suitably reduce the data in order to use it in the described mathemat-
ical models described in the next section. However, as described in the following, the step of
data processing and reduction provides also an opportunity to improve the quality of the
data for the chosen purpose.
It has been found that the (undesired) impact of the rotational speed on the vibration level
described in the previous section can be significantly diminished by a reduction to maximum
values within chosen time intervals: The whole CMS-monitored period of each turbine is
divided into a sequence of intervals of e.g. 1 week or 1 month. From each time interval, only
the maximum trend value together with its time stamp and the corresponding rotational-
speed value are kept. Figure 34 demonstrates the effect of this data-reduction procedure for
the two turbines already considered in Figure 33. It compares the case of the unreduced
original data with the cases of reduction using intervals of 1 week and 1 month, respectively.
Besides the strong reduction in the number of data points, which is a prerequisite for subse-
quent step of model development, Figure 34 shows the resulting concentration of remaining
data points in similar ranges of rotational speed. (Note that the data-reduction procedure is
applied to lc and hc trend data separately.) In this way, the comparability of the trend values
measured at different points in time is increased and the CMS data can be considered a bet-
ter indicator of the bearing condition.
72(122)

Figure 34: Impact of data reduction to maximum values on the representation of


different rotational speeds

Two alternative options of data reduction have been investigated:


(1) The reduction to vibration data measured in a selected range of rotational speed, e.g.
in the maximum 10% of the generator’s overall rotational-speed range. While prom-
ising an even better comparability of the resulting data due to the significantly re-
duced rotational-speed range, a problem arising from reducing the data to selected
ranges of rotational speed is that the data can become too sparse for the subsequent
analysis. This undesirable effect makes it the approach unsuitable for the data reduc-
tion in the present context.
(2) The reduction to moving-average values calculated for averaging intervals of 1 week
to 3 months. As in the case of reduction to maximum values described above, only
one value is kept from each time interval in the CMS history. A significant drawback
of this approach is the loss of peak-value information in the averaging procedure, as
it is illustrated for the case of turbine WT145 in Figure 35.
A comparison of the investigated approaches shows that the trend-data reduction to maxi-
mum values is the most suitable one, as it combines the advantages of maintaining the peak-
value information, of keeping data from the whole CMS-monitored period and of enhancing
the comparability of trend values measured at different points in time. A certain drawback of
this approach to be aware of is, however, its sensitivity to measurement errors or disturbing
effects (e.g. operating yaw motors in the moment of measurement) which cause an overes-
timation of the real vibration level.
On the background of the above reasoning, the reduction to maximum values has been cho-
sen and applied in the present work. Figure 36 gives an example of a CMS trend data history
before and after reduction to maximum values, using reduction intervals of 1 week, 2 weeks
and 1 month, respectively.
73(122)

trendWT145, averaging interval: 7 days trendWT145, averaging interval: 14 days


40 30
DS data DS data
report 25 report
30
20
trend

trend
20 15

10
10
5

0 0
Jul07 Jan10 Jul07 Jan10

trendWT145, averaging interval: 30 days trendWT145, averaging interval: 90 days


25 10
DS data DS data
20 report 8 report

15 6
trend

10 trend 4

5 2

0 0
Jul07 Jan10 Jul07 Jan10

Figure 35: A trend-data reduction to moving-average values (here over intervals of 7, 14, 30 and 90
days, respectively) is afflicted with a loss of peak-value information and thus unsuitable
in the present context.

Figure 36: CMS trend data history before and after reduction to maximum values using reduction
intervals of 1 week, 2 weeks and 1 month
74(122)

4.6 Mathematical reliability modelling


The objective of developing reliability models is to derive, on the basis of historical data, a
mathematical representation of the reliability behaviour of a system or a component in or-
der to estimate the probability of failure events occurring in the future.

4.6.1 Basic equations and terminology


The probability density function (pdf)

    

 


(3)

with  0 for all times t 0 describes the probability of a failure occurring within a time
interval dt, divided by the length of the interval dt. The area under the pdf sums up to one,
corresponding to the fact, that the probability of a failure between t=0 and infinity equals to
one:


   1 (4)


The cumulative distribution function (cdf, also called unreliability) denotes the probability of
failure before time t and is equal to the integral of the pdf until time t:


       
 (5)


Closely related to the probability of failure F(t), the reliability function R(t) describes the
probability of survival until time t:


     1     
 (6)


Note that for the time t approaching infinity, F(t) → 1 and R(t) → 0.

Finally, a key quantity in reliability analysis and modelling is the hazard function
75(122)

    
| 
 


(7)

which can be understood as the probability of failure in the interval dt under the condition
that no failure has occurred until time t, divided by the length of the interval dt. It is related
to the previous quantities according to:

 
  
 1  
(8)

The knowledge of one of the distributions f(t), h(t), F(t) or R(t) thus implies the knowledge of
all other distributions as well.

4.6.2 Univariate models


As the equations above suggest, the reliability behaviour of equipment is often expressed as
a function of time (i.e. the component age) only. Instead of t being a time, it could also be
understood as another variable characterizing the usage of the equipment, e.g. the number
of kilometers run by a car or the number of parts manufactured by a machine. The property
these options have in common is that a single usage-variable is used to characterize the reli-
ability of the equipment, while no concomitant information like e.g. condition-information
obtained from CMS is taken into consideration.
One of the most widely used life distributions in reliability analysis is the Weibull distribu-
tion. It is very flexible and can, through the choice of suitable parameters, model a variety of
reliability behaviours (see e.g. [Rausand and Hoyland (2004)] for details about the Weibull
distribution).
The time to failure T is called to be Weibull-distributed with the parameters η and  when
the distribution function, the density function and the hazard function are given by:

 !
  1     "

(9)

  !$%
 !
  ∙    "
  
(10)

  !$%
  ∙
 
(11)
76(122)

Note that the Weibull distribution is only defined for t>0.


For the purpose of illustration, a univariate Weibull distribution is fitted to the bearing fail-
ure data introduced in Section 4.2. The Weibull analysis is based on both failures and sus-
pensions together with the corresponding component age. Figure 37 (left) shows the
Weibull probability plot for the data. The approximate alignment of the data points along
the straight line

(12)

confirms the suitability of the Weibull distribution to characterise the reliability behaviour of
the component under consideration. The scale parameter η and the shape parameter β in
Equations (9)-(12) is estimated by means of the maximum-likelihood method. The obtained
estimate of the shape parameter β = 2.3 indicates that the hazard function of the analysed
bearings increases over time, as it is typically the case in presence of accumulating damages
and ageing (see also Chapter 3).

Figure 37: Weibull probability plot of the bearing failure data including suspensions (left); Weibull
cumulative distribution function of the bearing lifetime (right)

Figure 37 (right) displays the resulting distribution of the probability of failure over the bear-
ing age, based on the estimated Weibull parameters. From this, the expected lifetime is cal-
culated to 12.8 years. The 95% confidence interval shown in the graph reveals that 95% of
the bearings are expected to fail between the age of 3.0 and 25.3 years. The high width of
the confidence interval of more than 22 years obtained in the present case clearly shows the
strong limitations of the univariate model in the context of maintenance planning. Obvious-
ly, lifetime predictions with significantly higher levels of certainty and therefore with consid-
erably narrower confidence interval are required in order to base maintenance-planning
decisions upon them.
77(122)

4.6.3 Multivariate models


The limitations of univariate models discussed above can be overcome by means of multivar-
iate lifetime models which include the effect of covariates.
Covariates can be stress factors that influence the life of the monitored component (so-
called external covariates), e.g. quantities related to the mechanical stress on a wind-turbine
component or environmental conditions like ambient temperature. Alternatively, covariates
can be quantities related to the technical condition of a component, like e.g. the vibration
signals or the debris particle concentration in lubricant provided by CMS (internal covari-
ates).
Multivariate lifetime models which include the effect of covariates are e.g. the Accelerated
Failure Time Model, the Proportional Hazard Model, the Additive Hazard Model, the Propor-
tional Odds Model and the Prentice William Peterson Model. See Vlok (1999) for a concise
overview and Gorjian et al. (2009) for a more detailed review and comparison of these mod-
els.

4.6.4 Proportional Hazards Model


Among the multivariate models named above, the Proportional Hazard Model (PHM) origi-
nally developed for biomedical applications by Cox stands out by having been most compre-
hensively investigated and very widely used. Among the areas within which PHM has been
applied in the past are e.g. aircraft engines [Jardine and Anderson (1985)], nuclear reactor
cooler pumps [Montgomery et al. (2006)], brake discs in high-speed trains [Bendell et al.
(1986)], machine tools [Mazzuchi and Soyer (1989)] and power-supply cables [Kumar and
Klefsjö (1994)]. Vlok applied the PHM in a context having some similarity with the task in this
project, for prognosis of the remaining lifetime of water-pump components including bear-
ings using vibration-CMS data [Vlok (1999)].
Cox showed that the hazard rate of a component can be expressed by the product of an ar-

being a function of the covariate vector &̅. In reliability applications, a common and widely
bitrary (and possibly even unspecified) baseline hazard function h0(t) and a functional term

applied form is the exponential form of the functional term in combination with a Weibull
baseline hazard function:

; )    ∙ exp-.̅ ∙ &̅/



  !$%
 ∙  ∙ exp0.% ∙ &%  .1 ∙ &1  ⋯ 3
(13)

 

). While t denotes the time (calendar age, working age or another usage parameter of the
Herein, β, η and γi are the model parameters, which are summarised in vector of unknowns

component under consideration, see Section 4.6.2), &̅ is the vector of covariates.
78(122)

In the present work, the PHM is applied in its fully-parametric form given by Equation (13)
with t denoting the calendar age of the generator DS bearings and the CMS trend data as a
single time-dependent internal covariate. In the following, the covariate z(t) is thus consid-
ered a scalar quantity. Consequently, the PHM contains three unknown parameters (, η
and γ) which are to be estimated from the failure data and the CMS data histories.
Note that there is no distinction between different failure modes of the considered bearings,
as the provided data does not allow this level of detail in analysis. Due to the immensely
large amount of data that would be required for reliability modelling on the level of detail of
different bearing failure modes, it is neither considered to be of interest for the practical
application of the prognosis methodology.
The choice of suitable covariates requires expert knowledge. In the present case, this choice
has been made the monitoring experts at the project partner SKF who selected for this work
the CMS trend data as a well-proven indicator of bearing damage. In cases where covariates
have to be chosen among a variety of potential condition indicators, statistical hypothesis
testing can be used to verify that a covariant is significant and thus is in fact related to the
lifetime of a component.
In the present report, the description of the theory underlying the PHM is mainly limited to
the aspects that are required for applying the model in practice. For an in-depth introduction
to the PHM including the mathematical derivation of its key equations, please refer to [Ban-
jevic et al. (2001)].
In contrast to the univariate lifetime models presented in the previous section, the multivar-
iate PHM model offers the potential of providing bearing-specific predictions of the residual
life which are continuously updated during the monitoring process when new condition in-
formation is provided by the CMS.
In the following sections, the different steps required for the implementation of the progno-
sis method are presented, including the key equations.
The following nomenclature is used, which follows to a large extent the nomenclature in
[Banjevic, et al. (2001)]:
• ): vector of PHM model parameters (, , .

0  5  5%  ⋯  567  5 : the covariate observation points for the ith component
• z(t): time-dependent covariate; here: CMS trend data

history; here: the bearing age in the moment of measurement of a certain CMS trend da-
ta point, with Ti denoting the age at failure or suspension
&9  & 5 59  with j = 0,1,…,ki: observed covariate values in the component history i;
5

here: the jth data point in the (possibly reduced) CMS trend history of bearing i
• (Ti, Ci, z(i)(s); s ≤ Ti) with i = 1,2,…,n: sample of n independently observed (component)
histories, where Ci is the censoring indicator
79(122)

4.6.5 Maximum-likelihood based parameter estimation


The method of maximum likelihood is applied to estimate the unknown parameters of the
PHM model (, η and γ).
For right-censored data and independent censoring, the likelihood is (see e.g. [Banjevic et al.
(2001)], [Lawless (2003)]):

:∝ < 5 , & 5 5  ; ) ∙ < 9 , & 9 9  ; )
5 9
(14)
=>5?@ABC DE?F >?? G5CDA5BC

In this, the hazard function is given by Equation (13). The non-conditional reliability function
for history i can be calculated using (see e.g. [Banjevic et al. (2001)])

67 $%
59L% !
59 !
-5 , & ; )/   H I J. ∙ &9 K ∙   "N
5 5
 
(15)
9M

with the covariate value between two data points being either equal to the mean of the pre-
vious and the next data point according to

& 5   0.5 ∙ Q&9 &9L% R for 59    59L%


5 5
(16)

or assuming a constant value at the level of the previous data point according to

& 5   &9 for 59    59L% .


5
(17)

Equation (15) is a discrete approximation of the continuous equation


-, & /     , )
 "
5

 
 !$%
      J. ∙ & 5 K
"
  
(18)

 !
    exp -. ∙ & 5 /
 "
 

which relates the reliability function and the hazard function.


For the likelihood function follows, by substituting Eqs. (13 ) and (18) into Eq. (14):
80(122)

(19)

The parameter set maximising this likelihood function, , is the one representing the
historical reliability behaviour of the given sample and its relation to the covariate values in
the best way. It is therefore this parameter estimate which is used in the PHM model in or-
der to make predictions about the future reliability.

Figure 38: Example of the log-likelihood distribution over β and η for optimal γ. The red circle
marks the optimum parameter set which maximizes the likelihood function.

4.6.6 Goodness-of-fit test for the PHM


A review over different graphical and analytical goodness-of-fit tests suitable for the PHM is
provided by Vlok (1999). Out of these, one graphical and one analytical method have been
applied in the present work and are thus shortly presented in the following.

4.6.6.1 Residual plots

Residual plots are constructed using the Cox-generalized residuals for the PHM. These are
defined by

(20)
81(122)

with i=1,2,…, n and the ui being calculated according to:

67 $%
59L% !
59 !
S5  I J. ∙ &9 K ∙   "
5
 
(21)
9M

By ordering the Cox-generalized residuals ri and plotting the ordered residuals r1*≤ r2*≤ … ≤rn
against their expected values

1 1 1
T5  T5,E  ⋯
U U1 UV 1
(22)

the fit of the Weibull-PHM can be assessed. A distribution of the resulting data points (Ei; ri*)
along the line y=x indicates a good fit of the model. The visual assessment is facilitated by a
transformation of the data points (Ei; ri*) to (1-exp(Ei); 1-exp(ri*)), see [Vlok (1999)], which
results in an approximately equally spaced distribution of the data points along the horizon-
tal axis. Again, for a Weibull PHM fitting the underlying data well, the residual plots will be
distributed along the line y=x.

4.6.6.2 Z test

A statistical test for which the distribution of the test statistic under the null hypothesis can
be approximated by a normal distribution is called a Z-test. In the present case, the null hy-
pothesis to be tested is H0: The estimated model parameters (, η and γ) are equal to the
true values describing the reliability behaviour of the modelled bearings, i.e. the model fits
the data.
For the PHM model, the Z-test is implemented by calculating, from the ordered residuals
r1*≤ r2*≤ … ≤rn, the Z-score

∑[\] A∗
W with W5  A7∗ .
7^] Y7 $.ZE$%
_.ZE$%
(23)
[

The p-value corresponding to the obtained value of Z (based on a normal distribution) can be
used to make inference about rejecting or not rejecting the null hypothesis. The null hypoth-
esis is usually rejected if the p-value is smaller than or equal to the significance level α.

4.6.7 Prediction of covariate behaviour


Once the parameters of the proportional-hazards model have been estimated from the fail-
ure and suspensions events and the CMS data histories, and the goodness-of-fit has been
confirmed, a model is available which quantitatively relates the probability of failure with
82(122)

the corresponding covariate (here: vibration-CMS data) history. However, because covari-
ates obtained from CMS signals are per se time-dependent, it is necessary to predict their
future behaviour in order to allow a prognosis of the residual life.
Vlok (2011) discusses stochastic and non-stochastic covariate behavior as well as techniques
to predict future covariate behavior in both cases. In the stochastic case, covariates can only
be predicted to lie within certain confidence bounds while an exact prediction is not possi-
ble. A situation to which this applies and for which a stochastic covariate would obviously be
most appropriate is the example of using wind speed as an (external) covariate.
In the non-stochastic case, it is assumed that the future covariate behavior can be predicted
with reasonable accuracy beyond the present time t, based on the covariate history up to
this point in time. In the context of condition monitoring where CMS signals related to the
(often gradually deteriorating) component condition are used as covariates, this is usually
the case. In case of a developing damage, often clear trends are visible in the data history,
which allow a reasonably accurate extrapolation into the future (see Figure 39 for an illus-
trating example). To capture the variability of the CMS data is of less interest, but the covari-
ate prediction model should be capable of properly representing the trend in the data. Due
to this reasoning and the higher appeal to maintenance practitioners pointed out by [Vlok
(2001)], covariates are considered non-stochastic in the present work. They are modelled by
means of polynomial or exponential functions, the parameters of which are obtained
through non-linear regression using the Levenberg-Marquardt algorithm (“nlinfit” command
in MATLAB).

Covariate history Covariate prediction

Figure 39: Example of history-based covariate prediction by means of a regression function

For the prediction of future covariate behaviour, it is recommendable to choose a regression


function which is as closely as possible related to the monitored physical phenomenon, i.e.
in the present case the defect growth in metallic components. According to Figure 40, this is
usually characterized by an increasing defect-growth rate towards failure. An additional
property of interest in this context is the fact that a defect will either remain constant in size
or will grow, but will not become smaller over time. Note that service maintenance applied
83(122)

to a bearing by means of re-greasing might temporarily reduce the vibration level at that
bearing, that this will, however, not reduce an already present defect. To constrain the re-
gression functions to either constant or positive slopes is thus a reasonable approach even
for components subject to preventive maintenance (excluding any preventive replace-
ments).

Figure 40: Damage growth characteristics of metals and composites (left, [Jones (1999)]) and of
corrosion vs. crack-like defect growth (right, [Isaksson and Dahlgren (2011)])

For the prediction of covariates representing CMS signals closely related to component con-
dition, regression functions with the following properties are considered most suitable:
• exclusion of negative values
• constant or positive slope for all future points in time
• a growth-rate increasing with the covariate level (corresponding to the defect-growth
rate increasing with the size of a defect)
These properties are fulfilled by exponential regression approaches, while polynomial re-
gression functions usually have the significant drawback of providing physically implausible
predictions in many cases.
Both 2nd-order polynomials as well as different exponential regression functions have been
investigated in the scope of this project. According to this, the best covariate representation
and prediction has been achieved with the multi-step approach described in the following.
Besides the long-term horizon covering the whole available CMS data history of a compo-
nent, a short-term horizon of e.g. 1-6 months is used for regression in order to capture also
recent trends with faster damage-growth dynamics in the data.
(1) Having defined a suitable short-term horizon, the regression function
z(t)=exp(at)+b is fitted both to the whole CMS data history of a bearing and to the se-
lected number of most recent months (the short-term horizon named above) in the CMS
data.
84(122)

(2) Of the two parameter sets obtained, the one with the steeper slope (corresponding to
the larger value of a, i.e. showing the faster dynamics) is chosen in order to provide a
conservative prediction of the future covariate behaviour. Note that for DS trend values
z ≤ 1, always the long-term horizon is used for covariate prediction, with the purpose of
avoiding unrealistically short residual-life prognoses which can result from trend data
variations at irrelevantly low levels.
(3) If the procedure described above provides negative coefficients a for both the long-term
and the short-term regression, the data is approximated by means of a horizontal line,
fitted to the data in the short-term horizon.

Figure 41: History-based covariate prediction obtained using the implemented multi-step exponen-
tial approach; “6m” indicates the length of the short-term horizon of 6 months chosen in this case;
the red data points are those the plotted regression line is based upon.

Figure 41 illustrates the results of this multi-step prediction approach using the trend data
histories of four bearings: In the case of turbine WT24 shown in the top left figure, the high-
est positive slope was obtained from the long-term horizon covering the whole data history.
The DS bearing in turbine WT38 displays an apparently random variation of the vibration
level, with a minor decreasing trend. In accordance with the reasoning described above, it is
approximated by a horizontal line representing the average value from the short-term hori-
85(122)

zon. The cases of turbines WT12 and WT38 show faster dynamics in the short-term horizon
than in the long-term horizon. The covariate prediction is thus based on the data from the
most recent 6 months in both cases.
Summarising, the described multi-step covariate-prediction approach based on the regres-
sion functions
b c V b 0
&  a
c V b  0
outperforms polynomial regression approaches with respect to physical plausibility and is
capable of representing both short- and long-term trends in the CMS data in an accurate
way.

4.6.8 Residual-life prognosis


Based on the optimal parameter set for the Weibull PHM determined according to Section
4.6.5 (being valid for the entirety of considered components of comparable type) and on the
prediction of the future covariate development z(t) (being specific for every individual wind-
turbine generator bearing), the hazard function of this individual component can be calcu-
lated for future points in time:

  !$%
  ∙ ∙ exp. ∙ &
 
(24)

The non-conditional reliability functions for t>Ti (with Ti the present time) is given by:


 !
, &  5  ∙    exp . ∙ &
 "

(25)
d7

The conditional expectation of the residual life can be calculated using:

fd 



e5   T0  5 | 5 3  7
5 
(26)

Based on the present age of the component and its extrapolated CMS data history, this is the
expected value of the remaining lifetime. Obviously, the expected age at failure texp.fail.
equals the sum of the present age and the expected residual life μ:

Bgh.=>5?.  T0| 5 3  5 e5  (27)


86(122)

The conditional probability of failure F(t)=P(T≤t|T>Ti) of a failure occurring until a future time
t, knowing that the component has survived until the present time Ti, is obtained using:


  1 
5 
(28)

Finally, the 95%- confidence bounds for the age at failure, i.e. the lower bound t0.025 and the
upper bound t0.975) correspond to the time in the future at which the cumulative distribution
function takes on values of 0.025 and 0.975, respectively:

.1Z 
.1Z   1   0.025
5 
(29)

.jkZ 
.jkZ   1   0.975
5 
(30)

sumptions that (a) the estimated Weibull-PHM parameters , , . are the true parameters
It is important to note that the obtained confidence bounds are based on the implicit as-

and not afflicted with any uncertainty, and (b) that the future covariate behaviour is predict-
ed without any error. As also discussed by Vlok (2001), this is questionable in case of the
limited sample size and the strongly variable covariate behaviour often found in practical
cases like the present one. When assessing the results, it must therefore be kept in mind
that the confidence bounds calculated here can only represent the spread of the considered
components’ reliability behaviour resulting from the estimated optimal Weibull-PHM.
In analogy with Eqs. (29) and (30) above, the recommended time of replacement (sometimes
also called “Just-in-time point”) for a chosen level of risk r of unexpected failure, which a
CMS analyst would be willing to accept, can be calculated. This is given by:

A5C6 ?BoB? 
-ABh?>nB /  1  p
5 
(31)

4.6.9 Quality measures for the prognosis model


In order to assess the quality of the prognosis model (comprising the PHM as well as the co-
variate prediction model), two measures are introduced:
The model accuracy is quantified using the average of the squared errors on the residual-life
estimates compared to the real times of failure:
87(122)

v ƒ
u ‚
1 u 1 1‚
qrrs  ∙ I uw ∙ I -   / ‚
U5 Bgh.=>5?. =>5?@AB
u ‚
hABx.
(32)

u ‚
>?? =>5?Bx >?? hADyEDCBC
nD„hDEBEC z5G5E Bo>?@>5DE
E7 t GDA5{DE |}~€. 

The lower the measure MSSE of an assessed prognosis model, the higher is its prognosis accu-
racy in the evaluation horizon.
The uncertainty of the model is measured by the summed widths of the residual-life confi-
dence intervals according to:

v ƒ
u ‚
1 u 1 ‚
q@En  ∙ I u ∙ I .jkZ  .1Z ‚
U5 whABx.
u ‚
(33)

u ‚
>?? =>5?Bx >?? hADyEDCBC
nD„hDEBEC z5G5E Bo>?@>5DE
E7 t GDA5{DE |}~€. 

Therefore, the lower the measure Munc, the lower is its prognosis uncertainty in the evalua-
tion horizon.
Due to the fact that in the condition-monitoring context, the importance of a good predic-
tion accuracy and low uncertainty is highest during the time close to failure, an evaluation
horizon of 2 months before observed failure is used for the calculation of the above two
quality measures.
88(122)

4.7 Results and discussion


The prognosis model developed in the present project is capable of predicting the residual
life of a generator bearing based on its age and its CMS trend data history. While the preced-
ing chapter summarised the theory underlying the different parts of the prognosis model,
the present chapter aims to provide an understanding of the results obtained from this
model, to highlight different aspects investigated during its development together with the
best identified solution, and to discuss the potential and limitations of the final model.

4.7.1 Lifetime prognosis results in case of a turbine with bearing


failure
For the purpose of illustration and discussion of the results provided by the prognosis model,
the case of one specific wind turbine with a failure of the DS generator bearing is consid-
ered: For turbine WT144, Figure 42a shows the CMS trend data history (after reduction to
2-week maximum values) from the beginning of the monitoring period at a component age
of 3 months until the failure of the bearing at an age of 1074 days. For three points in time
which correspond to a component age of 600, 815 and 1020 days, respectively, the Figure
42a to c illustrate the residual-life prognoses the developed method would have provided
based on the CMS data measured until that time.
At the age of 600 days, the absolute vibration level is low. The regression analysis of the data
history until that time indicates an only slightly increasing trend level, which is extrapolated
into the future for the purpose of lifetime prediction (see the dark-green line in Figure 42a.
Figure 42b shows the resulting expected residual life of µ(t = 600 days)=734 days together
with the 95%-confidence bounds (71 days…1490 days), to which Figure 42c provides the cor-
responding conditional probability density distribution of the residual life over the bearing
age. Note the flat, wide shape of the conditional pdf and the high width of the confidence
interval of almost 4 years resulting from this.
At the age of 815 days, a significantly increased CMS trend level (hc) of 4.9 gE is observed,
together with an further increasing trend (see the green regression line in Figure 42a). Cor-
respondingly, the expected residual life is reduced to 340 days, with the 95%-confidence
interval spanning a still wide period of 27...646 days.
At the third marked point on the time axis, at a bearing age of 1020 days, a further rise in the
considered vibration level to 8.7 gE together with an increasingly steep trend slope is ob-
served (light-green line in Figure 42a). The expected residual is estimated to 147 days, with a
considerably narrower conditional pdf and a corresponding 95%-confidence interval ranging
from 11 to 284 days.
At a component age of 1074 days, 54 days later, the bearing is considered failed and the
turbine is stopped for replacement.
89(122)

(a)

(b)

(c)

Figure 42: CMS data history and prediction (a), predicted residual life incl. confidence intervals (b)
and predicted conditional lifetime pdf for the drive-side generator bearing in wind turbine WT144
90(122)

The case exemplarily discussed here illustrates well the significantly enhanced certainty in
lifetime prediction compared to that obtained from the only univariate Weibull-model dis-
cussed in Section 4.6.2. In addition, the lifetime prognosis provided by the developed meth-
od is dynamically updated when new CMS data becomes available while the prediction was
static in the univariate case. Finally, the case discussed here illustrated well how the certain-
ty of the prognosis improves with an increasing CMS trend level and therefore usually pro-
vides increasing certainty towards the time of failure.

4.7.2 Aspects investigated during model development

4.7.2.1 Data basis for PHM parameter estimation

In Sections 4.6.4 and 4.6.5 of the foregoing chapter, the equations underlying the PHM
model and the maximum-likelihood based estimation of its parameters have been intro-
duced in a generally valid form. In particular, the likelihood function given in Eq. (19) is for-
mulated in a way that allows to include both exactly observed lifetimes (i.e. component
histories ending with failure) and right-censored data (component histories in which no fail-
ure had occurred until the end of observation) in the estimation of the PHM parameters.
The inclusion of censored histories in the parameters estimation is, however, subject to con-
troversial statements in the literature. While Montgomery et al. (2006) state that both fail-
ure-terminated and censored component histories should be included, Vlok et al. (2004)
claim that the inclusion of not yet failed components in the construction of the lifetime
model leads to biased coefficients. The latter authors thus recommend basing the lifetime
model solely on histories with failures.
Due to the fact that both approaches apparently have been capable of providing useful
models in the past, in the present project the two alternatives have been implemented and
compared. Based on this, the approach yielding the better model quality has been chosen
for the final model.
Figure 43 shows the results of the graphical goodness-of-fit test described in Section 4.6.6.1.
The figures on the left display the residual plots for the case in which the Weibull PHM pa-
rameters are estimated based on all 211 valid component histories, thus including both fail-
ure-terminated and censored histories. The figures on the right show the corresponding
residual plots for the case in which the model parameter estimation is based on the 9 fail-
ure-terminated component histories only.
The residual plots clearly indicate a major difference in the goodness-of-fit of the two mod-
els: The model fit obtained based on both failure- and suspension-terminated histories (left)
is poor. In particular, the model fails to represent the histories including failure as the large
number of outliers below the 10%-confidence bound at a residual level of 0.1 shows, all of
which are residuals corresponding to failure-terminated histories. Due to the fact that the
91(122)

intended utilisation of the model in wind-turbine condition-monitoring in the present pro-


ject requires good model accuracy in particular in the case of impending failure, the PHM
obtained based on all component histories is considered insufficient.
On the contrary, the PHM determined solely based on failure-terminated histories repre-
sents the data well: With the exception of a single outlier (related to history 1, WT3), the
residuals spread around the level of the horizontal line r=1 in the top right figure and are
clearly aligned with the straight line y=x in the lower right figure.

Figure 43: Residual plots for Weibull PHM constructed based on all component histories (left; o -
suspension, x - failure) and based on solely failure-terminated histories (right)

The Z-test as an analytical goodness-of-fit test introduced in Section 4.6.6 confirms the
graphical results in Figure 43. Recall that the null hypothesis to test is “H0: The model fits the
data”, which implies that the Z-score of its residuals would be normally distributed. In the
first case described above, which is based on failure- and suspension-terminated component
histories, the Z-score calculated using Eq. (23) (see Section 4.6.6.2) is Z=1.23, corresponding
to a (two-sided) p-value of p=0.22. In the second case based only on failure-terminated his-
tories, Z=-0.59 and therefore a p-value of p=0.56 is obtained. A small p-value indicates that
the null hypothesis is rejected with high confidence. Accordingly, the significantly larger p-
value obtained in the second case supports the model fit to be considerably better when the
censored histories are not taken into account.
92(122)

All PHM models discussed in the following are thus constructed with parameters estimated
from the 9 failure-terminated bearing histories.

4.7.2.2 Covariate selection

The developed Weibull-PHM includes CMS-data as an internal covariate. Out of the four
CMS signals provided for analysis in this project (see Section 4.2), the generator DS-bearing
specific ones of interest in this context are “trend DS hc” and “trend DS lc”, i.e. CMS trend
data measure in the high-rotational-speed class (hc) and in the low-rotational-speed class
(lc), respectively.
The most straightforward modelling option is to use these absolute CMS trend values as
covariates in the PHM. However, the fact that different wind turbines (also of identical type)
can show very different base levels of vibration in good technical condition suggests that an
alternative to this is worth being investigated as well: the utilisation of relative trend values
as covariates, which are obtained by means of normalizing the absolute trend data with a
reference level measured at the beginning of the monitoring period.
On this background, the following four potential covariates have been tested and the quality
of the resulting prognosis models has been compared:
1) &%  pU
…† ‡
2) &1  pU
…† ˆ‡
3) &‰ 
ABEx Šr Gn
>oBA>yB ABEx Šr Gn ?BoB?C =AD„ GB =5AC „DEG D= „DE5DA5Ey

4) &‹  >oBA>yB ABEx Šr Gn ?BoB?C =AD„ GB =5AC „DEG D= „DE5DA5Ey


ABEx Šr ?n

Table 4 summarises the measures of model accuracy MSSE and model uncertainty Munc which
have been defined in Section 4.6.9. It is found that the use of CMS trend data from high rota-
tional speeds as covariate (z1, z3) yields the better prognosis models than the use of data
from low rotational speeds (z2, z4).

Table 4: Comparison of the prognosis-model quality (quantified in terms of the inaccuracy measure
MSSE and the uncertainty measure Munc) resulting from different covariate choices

DS trend data, hc DS trend data, lc


(data from high rotat. speeds) (data from low rotat. speeds)
Absolute trend level MSSE = 5.91e+4 days² MSSE = 7.93e+4 days²
Munc = 354 days Munc = 479 days
Normalised trend level MSSE = 6.04e+4 days² MSSE = 7.38e+4 days²
Munc = 419 days Munc = 554 days
93(122)

A comparison of the suitability of absolute vs. normalised hc trend values reveals that the
use of the absolute DS hc trend data as covariate provides the best prognosis model, both
with respect to the accuracy and the certainty of the residual-life prediction (corresponding
to the lowest values of MSSE and Munc in Table 4). Consequently, this option has been select-
ed for the final model.
The choice of CMS data from high rotational speeds is additionally supported by the obser-
vation by CMS analysts that some damages are detectable only at higher rotational speed
[Heintz (2010)]. This stresses the importance of using hc data in order to obtain the best
possible lifetime prognosis.
The normalisation of data with a reference value can only provide useful results if the moni-
toring of the bearing under consideration started while it was still in good condition. In the
present data, this seems not to be the case for at least one wind turbine (WT147), at which
increased CMS trend levels were observed already from the beginning of the monitoring
period. The bad performance of the normalised covariates concluded here might thus to
some extent be caused by the analysed data set. If the abovementioned prerequisite is ful-
filled in future analyses using different data sets, the performance of lifetime prediction
based on normalised trend values might thus be better than the present case suggests.
Based on the data analysed here, however, the procedure of data normalization is not rec-
ommended.
Only absolute levels of DS hc trend data are used as covariate in the following.

4.7.2.3 Data-reduction intervals

As discussed in Section 4.5, a reduction of the original CMS data to its maximum values in
intervals of a predefined length has the benefits of reducing the amount and variation of the
input data to the prognosis and of enhancing the comparability of the data.
Data reduction intervals of 1 week, 2 weeks and 1 month have been tested in the present
work in order to identify the most suitable procedure for application in the final prognosis
model. The choice of data-reduction scheme influences the lifetime-prognosis calculation in
different steps: On the one hand, it has an impact on the optimal parameter set obtained for
the Weibull-PHM; on the other hand it determines the data history (and data density) based
on which the future covariate behavior is predicted.
Figure 44 shows the prognosis results obtained for wind turbine WT144 based on data-
reduction to maximum values using the three different interval lengths (a) 1 week, (b) 2
weeks and (c) one month.
The upper plot of each subfigure shows the reduced data history together with the covariate
predictions based on the data up to that time. Note that the cases (a) and (b) allow the use
of a short-term horizon of only 1 month for the covariate prediction procedure (see Sec-
tion 4.6.7), while this needed to be set to 2 months in case (c) in order to have sufficient da-
ta points in the short-time horizon to allow a prediction.
94(122)

(a) 1-week data-reduction


intervals

(b) 2-week data-reduction (c) 1-month data-reduction


intervals intervals

Figure 44: Prognosis results for wind turbine WT144 obtained using data-reduction to maximum
values in intervals of (a) 1 week, (b) 2 weeks and (c) 1 month
95(122)

The lower plots display the corresponding predicted lifetimes and their 95%-confidence
bounds for each of the points in the CMS data history. Note that the earliest prognosis is
provided upon availability of three data points in the history. Due to this, the first prognosis
in case (c) is obtained approx. 2 months later than in case (a).
A reduction of the data to maximum values in longer intervals has the advantage of provid-
ing less rotational-speed related variation in the data, which can be expected to be a better
indicator of the technical component condition. A drawback of reducing the original CMS
data to only one value per month is that single outliers (as e.g. in Figure 44 the single data
point with a particularly high trend level measured at a component age of 900 days) have a
higher weight in the parameter estimation process in this case. As such outliers are very un-
likely to be related to component condition, this is an undesired effect, which is diminished
in case of a higher data density. Another disadvantage of using the 1-month-based data-
reduction procedure, which becomes obvious from Figure 44c, is the resulting less frequent
update of the lifetime prognosis.
In addition to this mainly qualitative reasoning, the model quality measures introduced in
Section 4.6.9 have been utilized for a quantitative assessment of the three options. The re-
sults are summarised in Table 5. Note that due to the short computing time of the devel-
oped model, which does not exceed several seconds to a few minutes in any of the three
cases, the computational effort does not need to be taken into account in the choice to be
made.

Table 5: Comparison of the prognosis-model quality (quantified in terms of the inaccuracy measure
MSSE and the uncertainty measure Munc) resulting from different options of data-reduction

1-week data-reduction 2-week data-reduction 1-month data-reduction


intervals (case a) intervals (case b) intervals (case c)
MSSE = 6.62e+4 days² MSSE = 5.91e+4 days² MSSE = 9.68e+4 days²
Munc = 398 days Munc = 354 days Munc = 418 days

Based on the discussion above and the quantitative performance of the different resulting
models indicated in Table 5, the best-performing 2-week data-reduction scheme is found to
be the most suitable one, which is therefore applied in the developed final model.

4.7.3 Potential and limitations of the final prognosis model


As presented in the previous sections, different influential factors have been investigated
and the options providing the best prognosis quality with respect to the occurred failures
have been included in the final prognosis model. Its properties are summarized once again in
the following list:
• The parameters of the Weibull-PHM are estimated based on the failure-terminated his-
tories.
96(122)

• The absolute DS hc trend values are used as covariate.


• The original vibration data is reduced to maximum values over intervals of 2 weeks.
• A short-term horizon of 1 month is used for covariate prediction, while the long-term
horizon includes the whole CMS history.

The results discussed in the following have been obtained using the final prognosis with the
above properties.
Figures 45(a-d) and 46 show five different cases, which are suitable to discuss both the po-
tential and the present limitations of the developed prognosis model.
Figure 45a illustrates the case of wind turbine WT145, the generator DS bearing of which
showed outstandingly high vibration levels before it was replaced at an age of 742 days. A
significantly increasing trend level at the age of 300 days leads to a low predicted residual
life. The subsequent stabilization of the trend level suggests that the underlying damage is
not further developing at the same rate, which is reflected by higher and similarly stable
values of the predicted residual life in the age range of 380…700 days. An extraordinary in-
crease in the trend level in the weeks preceding failure correctly relates to a very short pre-
dicted residual life.
In the case of wind turbine WT189 shown in Figure 45b, the generator DS bearing failed at a
still significantly increased vibration level at an age of 1888 days. The case illustrates well
how the predicted residual life decreases with increasing trend level and trend slope, while
the certainty of the prediction increases at the same time. A comparison of the true remain-
ing lifetime represented by the red line in Figure 45b with the predicted lifetime intervals
reveals the good performance of the prognosis model in this case.
The prediction capabilities of the present model are, however, limited by the wide range of
vibration levels at which bearings have been replaced and thus are considered to have failed.
A major source of uncertainty lies in the fact that any information about the true bearing
condition at the time of replacement is missing.
A consequence of this is observed in in the case of wind turbine WT3 shown in Figure 45c:
Herein, a step-increase in the trend level at the age of approx. 220 days is followed by stable
trend values around 1.5gE. After an initially low remaining-life prognosis related to the step-
increase in vibration, the observed stabilisation of the vibration level from t=230 days results
in an increasing expectation for the remaining life. At a trend level as low as 1.6gE and a
bearing age of only 350 days, the considered bearing is replaced, while this would not have
been expected based on the high residual-life prognosis at this time. It may be suspected
that in this case the bearing has been replaced already at an early stage of damage and that
it would have reached a critical condition not before a later point in time.
97(122)

Figure 45: Prognosis results in four different cases; (a) failure at high vibration level, (b) failure at
intermediate vibration level, (c) failure at low vibration level, (d) a bearing with a developing dam-
age that had not failed yet at the end of observation; red line in (a-c): true remaining life
98(122)

Figure 46: Prognosis results in the case of a bearing without indication of damage

The case of WT3 discussed here shows that the present data basis used for development of
the prognosis model is to some extent subject to subjective factors: While an operator with
a higher risk appetite may initiate bearing replacement not before very strong indicators of
damage (as e.g. in case of WT145), a precautious operator would order a replacement al-
ready at the first sign of developing damage (as assumably occurred in the case of WT3). Due
to the abovementioned lack of information about the true bearing condition at replacement,
the bearings are simply classified as “failed” in both cases. Not surprisingly, this has a nega-
tive impact on the accuracy and certainty of the mathematical model based upon this data.
To overcome this problem and improve the quality of quantitative residual-life predictions
for a potential future application, it is recommended to collect, together with every re-
placement event, information about the true component condition and to classify as a fail-
ure only the cases with a confirmed critical level of damage.
Figure 45d illustrates the case of wind turbine WT199 containing a bearing with a developing
damage that had not failed yet at the end of observation (i.e. at the end of the available CMS
history). However, according to the most recent residual-life prognosis, this bearing can be
expected to fail within the next two years with a confidence of 95%. Figure 45d provides a
suitable example for discussing the impact of the covariate-prediction procedure on the
prognosis result: The majority of expected-lifetime values and the corresponding confidence
bounds show a systematic, uniform decrease with increasing component age and trend lev-
el. However repetitively in between, significantly lower residual-life predictions are obtained
99(122)

(see the lower plot in Figure 45d). A comparison with the covariate prediction functions
(plotted green in the upper plot in Figure 45d) shows that all outstandingly low values of the
predicted residual life correspond to cases in which a recent strong increase in the trend
level forced a pessimistic (steep) extrapolation of the trend-level history (based on the short-
term evaluation horizon, see Section 4.6.7). Although the use of long-term and short-term
based covariate predictions causes some variation in the prognosis results over time, it is
considered the most suitable approach, as in every instance with an observed steep increase
in the trend levels, it should be conservatively assumed that this steep trend increase is go-
ing to continue into the future.
Finally, the case of a bearing in WT213 with low vibration levels over the whole monitored
time is shown in Figure 46. After a slight initial increase in the trend level until t=600 days,
the trend values stabilize at a level as low as 0.7 gE. An interesting effect to be discussed by
means of this case is the fact that the corresponding residual-life estimates decrease steadily
in spite of the constant trend levels. The effect of decreasing life predictions can therefore
only be assigned to the impact of the component age. However, is it plausible that bearings
of different age but with identical vibration levels have different risks of failure? Although
the degree of this impact is surprising, it is considered possible: It could be explained e.g. in
the way that an aged bearing has suffered from material fatigue, so that an existing defect of
a certain size could be expected to grow faster than in non-aged material. Besides the fact
that the data analysed in the present work clearly reveals this impact of component age on
the failure risk, also results obtained from analogous cases of Weibull-PHM application, e.g.
by Montgomery et al. (2006), confirm this effect to be existent.
However, the degree to which the predicted residual life decreases only due to increasing
component age but in spite of constant vibration values (see Figure 46) appears questiona-
ble. A possibly overestimated impact of age in the prognosis model could be explained by
the fact that suspended histories were not taken into account in the parameter-estimation
procedure for the final model. As discussed in Section 4.6.6, this choice was made based on
the very limited capability of the Weibull-PHM to appropriately represent the complete set
of data of failure-terminated and censored histories. In a potential future development of
the methodology presented here into a practical application it is therefore recommended to
investigate also the suitability of alternative multivariate lifetime models (see also Sec-
tion 4.6.3) to represent the full set of data.
100(122)

5 Summary and conclusions


The WindAM-RCM project has covered two main topic areas: the application of the concept
of Reliability-Centred Maintenance (RCM) to wind turbines, and the optimal utilisation of
data from condition-monitoring systems (CMS) in the wind-turbine maintenance process.
During the initial phase of the WindAM-RCM project, a state-of-the-art study has been car-
ried out. The subject of this study has been closely aligned with the abovementioned focus
areas in the project: The concepts of Reliability-Centred Maintenance, quantitative mainte-
nance optimization (QMO) and Reliability-Centred Asset Maintenance as a combination of
the former are subject of the first part of this study; the second part is focused on condition-
monitoring systems and their utilisation in the maintenance process. For each of these top-
ics, the present report has provided a general introduction, summarised previous research
and presented the state of the art concerning the application in wind power systems. In ad-
dition, relevant terminology for this project has been compiled together with the definitions
provided in standards. The literature study has revealed that the methods of interest - RCM,
QMO, and RCAM – have been applied to wind power plants only in a few single projects until
today. Regarding CMS for wind turbines, a large number of publications has been found that
deal with data acquisition and data processing for condition monitoring as well as with diag-
nosis, while the literature focusing on failure prognosis for wind turbines was very sparse.
Within the project part on Reliability-Centred Maintenance, a limited-scope RCM analysis of
the wind turbines Vestas V44-600kW and V90-2MW has been carried out. The executing
RCM workgroup included Göteborg Energi AB as an owner of the analysed wind turbines,
Triventus Service AB as a maintenance service provider, SKF as a provider of condition-
monitoring services and a wind-turbine component supplier, as well as the WindAM re-
search group at Chalmers. Combining the results of failure statistics and assessment of ex-
pert judgement, the analysis has been focused on the most critical subsystems with respect
to failure frequencies and consequences: the gearbox, the generator, the electrical system
and the hydraulic system, the pitch system and the rotor. For these subsystems, the RCM
analysis has identified the most relevant functional failures and their failure causes as well as
suitable measures to prevent either the failure itself or to avoid critical secondary damage.
It has been found that a considerable number of preventive measures proposed by the RCM
workgroup for the V44-600kW turbine are implemented in the V90-2MW series. This reflects
the learning curve in wind-power technology development and indicates an increasing focus
on reliability and maintainability. A highly interesting result of the RCM analysis is the identi-
fied central role that vibration plays as a failure cause for mechanical failure of a variety of
components. From this it can be concluded that measures for prevention or early detection
of bearing damages and other sources of excessive vibration, like e.g. vibration-monitoring,
are particularly effective to enhance the overall reliability of the system.
In addition to the analysis of wind-turbine subsystem specific failures and appropriate coun-
termeasures, comprehensive background information regarding the current maintenance
practices has been obtained during the RCM study. Challenges which are currently impeding
101(122)

the operation and maintenance of wind turbines from becoming more cost-effective have
been identified and solutions have been proposed. Component standardization, enhanced
training of maintenance personnel, and the utilisation of data-based quantitative methods
for decision support in wind-turbine maintenance are concluded to be important steps to
improve the reliability, availability and profitability of wind turbines.
The results of the RCM study form the basis for the development of quantitative models for
maintenance strategy selection and optimization within the framework of the Reliability-
Centred Asset Maintenance (RCAM) approach. A first quantitative model for maintenance
strategy assessment has been developed in a thesis project supervised within the WindAM-
RCM project. In addition, the results of the RCM study are utilised in the PhD student project
of P. Bangalore titled “Load- and risk-based maintenance of wind turbines”, which is carried
out within the Swedish Wind Power Technology Centre (SWPTC) at Chalmers since autumn
2011. In addition to providing the basis for quantitative maintenance management models,
the results of the RCM study can also be used as a feedback of field experience for further
improvement of wind-turbine design.
The second main part of the WindAM-RCM project was dedicated to approaches for opti-
mally utilising CMS data in the maintenance process. Based on the outcome of the literature
study and discussions with the project partners, the CMS-data based prognosis of the resid-
ual life of wind-turbine components has been identified as main focus in this part of the pro-
ject. This has been carried out in close cooperation with the SKF Wind Industry Service
Centre in Hamburg, from which over 600 CMS-equipped wind turbines are being monitored.
The frequently failing generator bearings were selected as the wind-turbine component for
which the prognosis methodology was implemented and tested. Vibration and failure data
from over 200 turbines of identical type has been used for this purpose.
So far, lifetime models of wind-turbine components had been limited to univariate, usually
purely age-based reliability models, which did not take into consideration information from
CMS and which are of very limited use for maintenance planning due to their usually very
high uncertainty in the prediction of failure times. Alternatively for wind-turbine mainte-
nance planning, the visual or mathematical extrapolation of trends in the vibration data is
being used for a subjective assessment of the risk of failure.
In contrary to this, the proportional-hazards model (PHM) used in the present work is capa-
ble of integrating failure and CMS data, which is a prerequisite for quantitatively assessing
the risk of failure and predicting the residual component life. Using the CMS and failure his-
tory of a fleet of turbines as input data, the prognosis model provides the distribution of the
residual life of the monitored component based on its age and its CMS vibration levels. In
this way, a turbine-specific prediction is provided, which is continuously updated when new
CMS data becomes available. The accuracy and certainty of the lifetime prediction increase
with increasing vibration levels and thus usually in vicinity of failure. The certainty of predic-
tion is significantly higher than that obtained from univariate reliability models (compare
Figure 37 and Figure 44).
102(122)

At present, the prediction accuracy of the developed prognosis model is limited (a) by the
variation in bearing condition and vibration levels at failure, which is closely related to a lack-
ing clear definition of bearing “failure”, (b) by lack of CMS-data from the initial operating
phase of wind turbines, and (c) by the lack of data from wind turbines of an age exceeding 8
years. However, these limitations can in principle be overcome by means of more compre-
hensive data collection in the future.
The presented lifetime-prognosis method provides a promising option for the optimal utili-
sation of wind-turbine CMS data in the maintenance process. While in the present work, the
method has been implemented and tested for the specific case of generator bearings of a
selected wind-turbine type, it is in principle applicable to any CMS-monitored wind-turbine
component. In light of this, the method offers the opportunity to introduce residual-life
based warning and alarm procedures in the monitoring process, being either an alternative
or a complement to the present use of manually set alarm thresholds and the mainly subjec-
tive assessment of the failure risk of a monitored component. In this way, the condition-
based maintenance of wind turbines could be optimised.
As a concluding remark, it shall be noted that the broad practical application of quantitative
methods in maintenance decision-support tools will require the structured and automated
collection of in-depth failure and maintenance data of wind turbines. Thus, further and in-
tensified efforts towards such systematic data collection and management, as e.g. using the
RDS-PP component designation structure combined with the EMS designation structure for
maintenance activities (see VGB PowerTech (2007), VGB PowerTech (2003)) as well as Com-
puterised Maintenance Management Systems (CMMS), are strongly needed. This is a pre-
requisite for tapping the full potential of quantitative maintenance optimisation for a further
cost-reduction of wind energy.
In summary, the following main contributions have been achieved in the WindAM-RCM pro-
ject:
• a presentation of the state of the art in Reliability-Centred Maintenance and related
quantitative approaches for maintenance management, of Condition-Monitoring Sys-
tems and procedures as well as of current practices in wind-turbine maintenance,
providing a basis for the present and future work on these topics;
• the application of the concept of Reliability-Centred Maintenance to wind turbines,
based on both statistical failure-data analysis and the experience of maintenance practi-
tioners, which has demonstrated the potential of the RCM method and provides a broad
knowledge base regarding the dominating failures, their causes and suitable remedial
measures in wind turbines; and
• the development of quantitative, data-based models for (a) maintenance strategy selec-
tion and optimisation and (b) enhanced condition-based maintenance, which offer the
potential to overcome the presently dominant reliance on subjective expert judgement
in maintenance management in the future.
103(122)

6 Related work
In parallel to the post-doc project described in the present report, two PhD student projects
have been on-going within the WindAM group at Chalmers during the project period, which
are closely related to the WindAM-RCM project and have offered a variety of opportunities
for discussion and collaboration:
• The project “Reliability modelling and optimal maintenance management for wind power
systems” is carried out by PhD student Francois Besnard in cooperation with Vattenfall
Wind Power and funded within Vindforsk III. The main subjects and selected results from
this work have been summarised in Section 2.4 of the present report.
• The project “Load- and risk-based maintenance of wind turbines” is carried out by
Pramod Bangalore in cooperation with Göteborg Energi AB, SKF and Triventus Service AB
and funded through the Swedish Wind Power Technology Centre (SWPTC) hosted at
Chalmers. The research application for this PhD student project was developed within
the WindAM-RCM project, based upon the outcomes of the RCM study. The main subject
of the project is to investigate the dependency of wind-turbine failures on environmental
factors e.g. like wind speed, wind turbulence, humidity or grid disturbances. With the
overall objective to achieve cost-efficient wind-turbine maintenance, it aims at develop-
ing a method for flexibly adapting the maintenance of certain wind-turbine components
based on the stress these components have been subjected to in the history.

In addition, three student thesis projects have been supervised within the WindAM-RCM
project, which are complementing the present work:
• Malin Sallhammar: “Reliability Analysis of Wind Power Systems - A statistical analysis of
failures in Swedish wind turbines”, master’s thesis at Chalmers in cooperation with LTH
Lund, 2011
M. Sallhammar carried out a statistical analysis of failure data from wind turbines located
in Sweden for the periods 1989-2005 and 2006-2009. She investigated the failure rates
and resulting downtime of wind turbines from different classes of rated capacity, the
seasonal variation of failure rate and downtime, and in particular the reliability charac-
teristics of the most relevant wind turbine subsystems (i.e. identification of a decreasing,
constant or increasing failure rate over time).
• Bertrand Kerres: “A comparison of wind turbine life-cycle costs and profits resulting from
different maintenance strategies”, project thesis at Chalmers in cooperation with RWTH
Aachen and Göteborg Energi AB, 2011
Bertrand Kerres analysed the revenues and costs from a wind turbine, and developed an
application capable of calculating and comparing the total cost of different maintenance
strategies over the lifetime. This was applied in a case study for the Vestas V44-600kW
system. The work has been carried out in close collaboration with Göteborg Energi which
104(122)

supported the work with access to maintenance reports and cost data for the V44-
600kW turbine, with wind data and with valuable discussions.
• Nelly Forsman: ”An analytical tool for the evaluation of wind power generation”, mas-
ter’s thesis at Göteborg Energi AB / Chalmers, 2011
Nelly Forsman developed a methodology for a SCADA-data based tool for detecting devi-
ations and degradations in wind-turbine performance by comparison with turbine-
specific historic data. Using the data histories of wind speed, wind direction, air density
and generated power, normalized power curves are derived for each wind turbine, which
represent the reference signature. The present performance data is compared to this
reference signature and deviations above a predefined threshold range generate an
alarm. The method has been implemented for Göteborg Energi’s V44-600kW turbines lo-
cated in the harbour of Göteborg. It provides a tool capable of detecting changes in
wind-turbine performance as a potential indicator of maintenance need as well as of
identifying measurement errors.
105(122)

7 Publications and communication


The results and knowledge created in this project have continuously been disseminated and
externally communicated - through presentations at conferences and seminars, scientific
publications, but also through the involvement of master’s thesis students as described
above. In the following, a compilation of the publications and reports created during the
WindAM-RCM project is given; in addition, the visited conferences and seminars (with
presentations of WindAM-RCM work as stated) carried out during the project period are
listed. Finally, the last section summarises the study visits undertaken within the project.

7.1 Publications and reports


• Fischer, K.; Besnard, F.; Bertling, L.: "Reliability-Centered Maintenance for Wind Turbines
Based on Statistical Analysis and Practical Experience", IEEE Transactions on Energy Con-
version, In press, available online since December 2011, DOI 10.1109/TEC.2011.2176129.
(related to Chapter 3 of the present report)
• Besnard, F.; Patriksson, M.; Strömberg, A.-B.; Wojciechowski, A.; Fischer, K.; Bertling, L.
(2011): “A Stochastic Model for Opportunistic Service Maintenance Planning of Offshore
Wind Farms”, Proceedings of the IEEE PES PowerTech 2011, Trondheim, Norway, 19 - 23
June 2011.
• Fischer, K.; Besnard, F.; Bertling, L.: ”A Limited-Scope Reliability-Centred Maintenance
Analysis of Wind Turbines“, Proceedings of the European Wind Energy Conference EWEA
2011, Brussels, Belgium, 14-17 March 2011. (related to Chapter 3 of the present report)
• Fischer, K.; Bertling, L.: “RCM analysis of the wind turbines Vestas V44-600kW and V90-
2MW”, Report at the Division of Electric Power Engineering, Department of Energy and
Environment, Chalmers University of Technology, Göteborg, Sweden, January 2011. (re-
lated to Chapter 3 of the present report)
• Besnard, F.; Fischer, K.; Bertling, L.: “Reliability-Centred Asset Maintenance – A step to-
wards enhanced reliability, availability, and profitability of wind power plants”. In Pro-
ceedings of the IEEE PES Conference on Innovative Smart Grid Technologies Europe,
Göteborg, Sweden, 10-13 October 2010. (related to Chapter 2 of the present report)
• Fischer, K.: “Maintenance Management of Wind Power Systems by means of Reliability-
Centred Maintenance (RCM) and Condition Monitoring Systems (CMS): Results of the
State of the Art Study”, Report at the Division of Electric Power Engineering, Department
of Energy and Environment, Chalmers University of Technology, Göteborg, Sweden, June
2010.
Two further publications are currently under preparation. These will present
• the model for maintenance strategy assessment developed in the thesis work of
B. Kerres (see Section 3.5 and Chapter 6 in the present report), and
106(122)

• the CMS-based residual life prognosis model described in Chapter 4 of this report.

7.2 Conferences and seminars


• Conference ”Vindkraftsforskning i fokus 2011” (Vindforsk/Vindval/SWPTC), Chalmers,
18-19 January 2012: Oral presentations (Pramod Bangalore, Francois Besnard, Katharina
Fischer)
• Forskningsdagen (Research day) at Göteborg Energi, Göteborg, 7 December 2011: Oral
presentation (Katharina Fischer)
• SKF Wind Farm Management Conference, Barcelona, Spain, 11-12 May 2011: Oral
presentation (Katharina Fischer)
• Seminar ”Statistical reliability and lifetime data analysis”, London,
20 April 2011 (Katharina Fischer)
• IEA Topical Expert Meeting “International Statistical Analysis on Wind Turbine Failures”,
Kassel, Germany, 30-31 March 2011: Oral presentations (Francois Besnard, Katharina
Fischer)
• European Wind Energy Conference EWEA 2011, Brussels, Belgium, 14-17 March 2011:
Paper and oral presentation (Katharina)
• Conference ”Vindkraftsforskning i fokus 2010”, Chalmers, 24-25 November 2010
(Katharina Fischer)
• IEEE Innovative Smart Grid Technologies (ISGT) Europe, Gothenburg, 11-13 October
2010: Oral presentation of joint article (Francois Besnard, Katharina Fischer, Lina
Bertling)
• Vind2010, Gothenburg, 15-17 September 2010: Oral presentation and poster (Katharina
Fischer)
• Vindforsk Seminar on Operations and Maintenance, Stockholm, 15 June 2010: Oral
presentation (Katharina Fischer)
• SKF Wind Farm Management Conference, Copenhagen, 9-10 June 2010: Poster presenta-
tion of the projects WindAM-RCM and WindAM-Opt (Katharina Fischer, Francois
Besnard)
• Nordic Wind Power Conference O&M, Stockholm, 25 May 2010: Oral presentation
(Katharina Fischer, Lina Bertling)
• ELKRAFTDAGEN, Göteborg, 25 March 2010: Poster presentation of the projects WindAM-
RCM and WindAM-Opt (Francois Besnard, Katharina Fischer)
• VGB Conference "Maintenance of Wind Power Plants", Bremen, Germany, 24-26 Febru-
ary 2010, (Katharina Fischer)
107(122)

7.3 Study visits


• Lillgrund Offshore Wind Farm, Öresund, Sweden, 1 February 2010
• SKF ERC, Nieuwegein, The Netherlands, 19 May 2010
• Kenersys K82-2MW turbine in Gårdsten, Göteborg, during Vind2010 conference,
17 September 2010
• SKF Wind Industry Service Centre (WISC), CMS-monitoring centre, Hamburg,
1 September 2010 and 17 June 2011
108(122)

Appendix

A.1 Relevant terminology

Reliability is defined as “the ability of an item to perform its required function under given
conditions for a given time interval”. [SS-EN 13306]

Availability is defined as “the ability of an item to be in a state to perform a required function


under given conditions at a given instant of time or during a given time interval, assuming
that the required external resources are provided”. [SS-EN 13306]

Note: The term Availability is not used consistently in the context of wind power and should
thus be specified in detail to avoid ambiguity. Within the present project, the most common
and straightforward definition will be used, describing availability as the ratio of operational
time divided by the nominal time, the latter usually being a period of one year:

Nominal time - Downtime


Availability =
Nominal time

In this definition, downtime due to external causes like failures of the grid or absence of
wind is not eliminated and is thus included in the same way as downtime after wind turbine
failures or due to preventive maintenance. [Ribrant (2006)]

Maintenance describes the combinations of all technical and corresponding administrative


actions, including supervision actions, intended to retain an entity in, or restore it to, a state
in which it can perform its required function. [IEC 50(191)]

Corrective maintenance is defined as “maintenance carried out after fault recognition and
intended to put an item into a state in which it can perform a required function”. [SS-EN
13306]

Preventive maintenance is defined as “maintenance is carried out at predetermined inter-


vals or according to prescribed criteria and intended to reduce the probability of failure or the
degradation of functioning of an item.” [SS-EN 13306]

Scheduled maintenance is defined as “preventive maintenance carried out in accordance


with an established time schedule or established number of units of use”. [SS-EN 13306]
109(122)

Condition based maintenance is defined as “preventive maintenance based on performance


and/or parameter monitoring and the subsequent actions. Performance and parameter mon-
itoring may be scheduled on request or continuous”. [SS-EN 13306]

Repair describes the part of corrective maintenance in which manual actions are performed
on an entity. [IEC 50(191)]

Failure is the event when a required function is terminated (or exceeding the acceptable
limits). [IEC 50(191)]

Fault is defined as “abnormal condition that may cause a reduction in, or loss of, the capabil-
ity of a functional unit to perform a required function”. [IEC 61508]

Note: According to [NASA (2002)], the term “fault” describes a defect, imperfection, mistake,
or flaw of varying severity that occurs within some hardware or software component or sys-
tem. It is a general term and can range from a minor defect to a failure.
110(122)

A.2 Example of an event report

Figure 47: Example of an event report provided to the wind turbine operator by the condition-
monitoring centre. Source: Kjellberg (2009)
111(122)

A.3 Failure report sheet

Figure 48: Failure report sheet including component categorization, by courtesy of Swedpower /
Vattenfall Power Consultant
112(122)

A.4 Test for equality of life distributions by means of the


proportional-hazards model
As described in Section 4.2, the analysed data comprises 211 component (DS generator
bearing) histories from a fleet of 202 wind turbines. All turbines in the fleet are of identical
model, but the sample comprises turbines with
• two different rotor diameters (denoted Type A and B),
• two different generator manufacturers (Type C and D)
• and two different bearing dimensions (Type E and F).
Table 6 summarises the representation of the different rotor, generator and bearing types in
the total of 211 component histories, as well as the number of failures in each class.

Table 6: Representation of turbine types (i.e. rotor diameter), generator manufacturers


and bearing types in the component histories
(in brackets: number of component histories ending with failure)

Wind turbine Type A Wind turbine Type B Total


Generator Generator Gen.unknown Generator Generator Gen.unknown
BearingType
Type C Type D (Type C or D) Type C Type D (Type C or D)
E 0 87(4) 7(1) 0 27(0) 0 121
F 4(0) 7(0) 4(0) 74(4) 0 1(0) 90
Total 4 94 11 74 27 1 211

As described by Lawless (2003), the semi-parametric form of the proportional-hazards model


can be used for testing the equality of distributions. In the present work, this model has
been used to test if the hypothesis that the monitored bearings have identical lifetime dis-
tributions even if they are operated in the turbines of type A and B, in generators of manu-
facturers C or D and have the dimensions E or F (thus, to test the assumption (1) stated in
Section 4.3: “All DS bearings in the sample have identical lifetime distribution and can thus
be described by a single reliability model”).
The three null hypotheses to test in the present context are:

Œ :    Ž 
Œ0 :    …  (34)
Œ0 : T    

For this purpose, the semi-parametric PHM

; .    ∙ exp.̅ ∙ &̅


(35)
113(122)

is utilized, in which the baseline-hazard function h0(t) covers the impact of time and internal
covariates and can remain unspecified.
The covariate vector & consists of constant, binary indicator-covariates taking on the value 0
or 1 according to whether an individual is associated with the first or the second distribution
(A or B, C or D, E or F). A test of γ=0 is a test of the corresponding hypothesis.
According to Lawless (2003), the hypothesis testing can be carried out by fitting the PHM
and using the Wald statistic for evaluation. For the present case of constant covariates, the
parameters γ and their standard errors can be determined using the “coxphfit” function
available in MATLAB.
High values for all obtained p-values indicate that none of the null hypotheses formulated
above can be rejected. Based on the limited amount of available failure data, none of the
indicator covariates related to the rotor diameter, the generator manufacturer or the bear-
ing size is found to be statistically significant. It is thus justified to assume that all analysed
bearings have identical lifetime distributions and to fit a single parametric PHM with solely
time-dependent internal covariates to the whole data sample as it is done in the present
work.
114(122)

References
[Agrolager (2012)]: Agrolager, Website, accessed January 2012,
http://www.agrolager.de/popup_image.php?pID=12060796&osCsid=2a7a2fd137534d16745
44cb44aa52c64
[Andrawus et al. (2006a)] Andrawus, J.A.; Watson, J.; Kishk, M.; Adam, A. (2006a): Determin-
ing an Appropriate Condition-based Maintenance Strategy for Wind Turbines. 2nd Joint In-
ternational Conference on Sustainable Energy and Environment (SEE 2006), B-027 (O),
Bangkok, Thailand, 21-23 November 2006.
[Andrawus et al. (2006b)] Andrawus, J.A.; Watson, J.; Kishk, M. (2006b): The selection of a
suitable maintenance strategy for wind turbines. International Journal of Wind Engineering,
Vol. 30(6), pp. 471-486.
[Andrawus et al. (2006c)] Andrawus, J.A.; Watson, J.; Kishk, M. (2006c): Asset Management
Processes in the Wind Energy Industry. Proceedings of the 2nd Joint International Confer-
ence on “Sustainable Energy and Environment (SEE 2006)” 21-23 November 2006, Bangkok,
Thailand, p. 269-274.
[Andrawus et al. (2007a)] Andrawus, J.A.; Watson, J.; Kishk, M. (2007a): Modelling System
Failures to Optimise Wind Turbine Maintenance. International Journal of Wind Engineering,
Vol. 31(6), 2007, pp. 503-522.
[Andrawus et al. (2007b)] Andrawus, J.A.; Watson, J.; Kishk, M. (2007b): Wind Turbine
Maintenance Optimisation: principlesof quantitative maintenance optimisation. Wind Engi-
neering, Vol. 31(2), 2007, pp. 101–110.
[Andrawus et al. (2008)] Andrawus, J.; Watson, J.; Kishk, M.; Gordon, H. (2008): Optimisation
of Wind Turbine Inspection Intervals. Wind Engineering , Vol. 32 (5), p. 477-490.
[Andrawus (2008)] Andrawus, J.A. (2008): Maintenance optimisation for wind turbines. PhD
thesis, Robert Gordon University, Aberdeen, UK, April 2008.
[Anjar et al. (2011)] Anjar, B.; Dalberg, M.; Uppsäll, M. (2011): Feasibility study of thermal
condition monitoring and condition based maintenance in wind turbines. Elforsk report
11:19, Stockholm, May 2011.
[Amirat et al. (2009)] Amirat, Y.; Benbouzid, M.E.H.; Al-Ahmar, E.; Bensaker, B.; Turri, S.
(2009): A brief status on condition monitoring and fault diagnosis in wind energy conversion
systems. Renewable and Sustainable Energy Reviews, Vol. 13 (2009), p. 2629–2636.
[Balschuweit (2009)] Balschuweit, R.: Beanspruchungs- und Schadensanalyse von Windener-
gieanlagen am Beispiel der Vestas V90-2MW. Diplomarbeit, TFH Berlin in cooperation with
Vattenfall, Germany, 2009.
115(122)

[Banjevic et al. (2001)] Banjevic, D., Jardine, A.K.S., Makis, V., Ennis, M. (2001), A control-
limit policy and software for condition-based maintenance optimization. INFOR, Vol. 39, pp.
32–50, 2001.
[Becker (2008)] Becker, E. (2008): Telemonitoring of wind turbines – for your condition-
based maintenance. Prüftechnik Condition Monitoring, magazine “antriebstechnik“, May
2008.
[Bellens and Chemweno (2010)] Bellens, R.; Chemweno, P. (2010): Business case for condi-
tion based maintenance of wind turbine gearboxes. Master’s thesis, Faculty of Engineering,
K.U. Leuven, Belgium, 2010.
[Bendell et al. (1986)] Bendell, A., Walley, M., Wrightman, D.W., Wood, L.M. (1986): Propor-
tional hazards modelling in reliability analysis – an application to brake discs on high speed
trains. Quality and Reliability Engineering International, 2, pp.45-52, 1986
[Bertling (2002)] Bertling, L. (2002): Reliability-Centred Maintenance for Electric Power Dis-
tribution Systems. PhD thesis, KTH, August 2002.
[Bertling et al. (2005)] Bertling, L.; Allan, R.; Eriksson, R. (2005): A reliability-centered
maintenance method for assessing the impact of maintenance in power distribution sys-
tems. IEEE Transactions on Power Systems, Vol. 1.
[Bertling et al. (2006)] Bertling, L.; Ackermann, T.; Nilsson, J.; Ribrant, J. (2006): Förstudie om
tillförlitlighetsbaserat underhåll för vindkraftssystem – fokus på metoder för
tillståndskontroll. Elforsk report 06:39, May 2006.
[Besnard (2009)] Besnard, F. (2009): On Optimal Maintenance Management for Wind Power
Systems. Licenciate thesis, KTH, Stockholm, December 2009.
[Besnard et al. (2009)] Besnard, F.; Patriksson, M.; Strömberg, A.; Wojciechowski, A.;
Bertling, L. (2009): An Optimization Framework for Opportunistic Maintenance of Offshore
Wind Power System. Proceedings of the IEEE PowerTech Conference 2009, Bucharest, Ro-
mania, 28 June-2 July 2009.
[Besnard et al. (2010a)] Besnard, F.; Fischer, K.; Bertling, L.: Reliability-Centred Asset
Maintenance – A step towards enhanced reliability, availability, and profitability of wind
power plants. Proceedings of the IEEE PES Conference on Innovative Smart Grid Technolo-
gies Europe, 11-13 October 2010, Gothenburg, Sweden.
[Besnard et al. (2010b)] Besnard, F.; Nilsson, J.; Bertling, L. (2010): On the Economic Benefits
of using Condition Monitoring Systems for Maintenance Management of Wind Power Sys-
tems. 11th International Conference on Probabilistic Methods Applied to Power System
(PMAPS), Singapore, 14 - 17 June 2010.
[Besnard and Bertling (2010)] Besnard, F.; Bertling, L. (2010): An approach for condition
based maintenance optimization applied to wind turbine blade. IEEE Transactions on Sus-
tainable Energy, 1(2), 2010.
116(122)

[Besnard et al. (2011a)] Besnard, F.; Patriksson, M.; Strömberg, A.-B.; Wojciechowski, A.;
Fischer, K.; Bertling, L. (2011): A stochastic model for opportunistic maintenance planning of
offshore wind farms. Proceedings of the IEEE PowerTech Conference 2011, Trondheim, Nor-
way, 19-23 June 2011.
[Besnard et al. (2011b)] Besnard, F.; Fischer, K.; Bertling, L.; Attermo, P. (2011): Analysis of
Maintenance Support Organization for Offshore Wind Farms. Presentation at the Offshore
EWEA Conference, Amsterdam, The Netherlands, 29 November-1 December 2011.
[Carlson et al. (2001)] Carlson, O.; Helmer, M.; Liljegren, C.: Analys av havererade
asynkrongeneratorer. Elforsk report VINDKRAFT – 4/01(in Swedish), 2001. Available at
http://www.vindenergi.org/teknikbev/Info_4-01.pdf.
[Cox (1972)] Cox, D.R. (1972): Regression models and life tables (with discussion), Journal of
the Royal Statistical Society, B34, pp.187-220
[Crabtree (2010)] Crabtree, C.J. (2010): Survey of Commercially Available Condition Monitor-
ing Systems for Wind Turbines. Download: http://www.supergen-
wind.org.uk/docs/Survey%20of%20commercially%20available%20CMS%20for%20WT.pdf
[Crabtree (2011)] Crabtree, C.J. (2011): Condition Monitoring Techniques for Wind Turbines.
PhD thesis, School of Engineering and Computing Science, Durham University, UK.
[Dekker (1996)] Dekker, R. (1996): “Application of maintenance optimization models: a re-
view and analysis”, Journal of Reliability Engineering and System Safety, 1996, 51(3), p. 229-
240.
[Engels et al. (2009)] Engels, W.; Obdam, T.; Savenije, F. (2009): “Current developments in
wind – 2009, Going to great lengths to improve wind energy”, Technical Report ECN-E-09-96,
ECN.
[EWEA (2010)] European Wind Energy Association EWEA (2010): Annual Report 2009 - A
Breath of Fresh Air, available online.
[EWEA (2011)] European Wind Energy Association EWEA (2011): Wind in power - 2010 Euro-
pean statistics, available online.
[Faulstich et al. (2008)] Faulstich, S.; Hahn, B.; Jung, H.; Rafik, K.; Ringhandt, A. (2008): Ap-
propriate failure statistics and reliability characteristics. Proceedings of the German Wind
Energy Conference DEWEK 2008, Bremen, Germany.
[Faulstich et al. (2009)] Faulstich, S.; Hahn, B.; Jung, H.; Rafik, K. (2009): Suitable failure sta-
tistics as a key for improving availability. European Wind Energy Conference EWEC 2009,
Marseilles, France, available at
http://ewec2009proceedings.info/allfiles2/520_EWEC2009presentation.pdf.
[Fischer et al. (2011a)] Fischer, K.; Besnard, F.; Bertling, L.: Reliability-Centered Maintenance
for Wind Turbines Based on Statistical Analysis and Practical Experience, IEEE Transactions
on Energy Conversion, In press, available online since December 2011, DOI
10.1109/TEC.2011.2176129.
117(122)

[Fischer et al. (2011b)] Fischer, K.; Besnard, F.; Bertling, L.: ”A Limited-Scope Reliability-
Centred Maintenance Analysis of Wind Turbines“, Proceedings of the European Wind Energy
Conference EWEA 2011, Brussels, Belgium, 14-17 March 2011.
[Germanischer Lloyd (2007)] Germanischer Lloyd (2007): Guideline for the Certification of
Condition Monitoring Systems for Wind Turbines, Edition 2007. Germanischer Lloyd Indus-
trial Services GmbH, Hamburg, Germany.
[Gorjian et al. (2009)] Gorjian, N., Ma, L., Mittinty, M., Yarlagadda, P., Sun, Y. (2009): A re-
view on reliability models with covariates. Proceedings of the 4th World Congress on Engi-
neering Asset Management, 28-30 September 2009, Athens
[Gray and Watson (2009)] Gray, C.S.; Watson, S.J. (2009): Physics of Failure approach to wind
turbine condition based maintenance. Wind Energy, published online August 2009.
[Hameed et al. (2009)] Hameed, Z.; Hong, Y.S.; Cho, Y.M.; Ahn, S.H.; Song, C.K. (2009): Condi-
tion monitoring and fault detection of wind turbines and related algorithms: A review. Re-
newable and Sustainable Energy Reviews, Vol. 13 (2009), p. 1–39.
[Hargreaves and Planitz (2009)] Hargreaves, D.J.; Planitz, A.: Assessing the energy efficiency
of gear oils via the FZG test machine. Tribology International, Vol. 42 (6), p. 918-925, 2009.
[Heintz (2010)] Heintz, B. (2010): Optimierung und Standardisierung der Analyse und Diag-
nose von Schäden am Antriebsstrang von Windkraftanlagen. Diplomarbeit, Technische Uni-
versität Kaiserslautern, Germany
[Hyers et al. (2006)] Hyers, R.W.; McGowan, J.G.; Sullivan, K.L.; Manwell, J.F.; Syrett, B.C.
(2006): Condition monitoring and prognosis of utility scale wind turbines. Energy Materials,
Vol. 1(3), 2006, p. 187-203.
[IEC 50(191)] IEC 50(191): International Electrotechnical Vocabulary (IEV) - Chapter 191 -
Dependability and Quality of Service. International Electrotechnical Commission, Geneva,
1990.
[IEC 61508] IEC 61508: Functional Safety of Electrical/Electronic/Programmable Electronic
Safety-Related Systems, Parts 1–7, International Electrotechnical Commission, Geneva, 1997.
[Isaksson and Dahlberg (2011)] Isaksson, E., Dahlberg, M. (2011): Damage prevention for
wind turbines, Phase 2 - Recommended measures. Elforsk report 11:18, Stockholm
[Jardine and Anderson (1985)] Jardine, A.K.S., Anderson, M. (1985): Use of concomitant vari-
ables for reliability estimation. Maintenance Management Int., 5, pp. 135-140
[Jardine and Tsang (2006)] Jardine, A.K.S.; Tsang, A.H.C. (2006): Maintenance, Replacement,
And Reliability: Theory and Applications. CRC Press, Boca Raton, FL, USA.
[Jardine et al. (2006)] Jardine, A.K.S.; Lin, D.; Banjevic, D. (2006): A review on machinery di-
agnostics and prognostics implementing condition-based maintenance. Mechanical Systems
and Signal Processing, Vol. 20, p. 1483–1510.
[Jonasson (2001)] Jonasson, K. (2001): Tillståndsövervakning av vindkraftverk – Utvärdering
av system utfört av SKF Nova. Elforsk report 01:30, 2001, available at: www.elforsk.se.
118(122)

[Jones, R.M. (1999)] Jones, R.M. (1999): Mechanics of Composite Materials. 2nd ed., ISBN 1-
56032-712-X, Taylor and Francis Inc., 1999
[Kerres (2011)] Kerres, B. (2011): A comparison of wind turbine life-cycle costs and profits
resulting from different maintenance strategies. Project thesis, Chalmers / RWTH Aachen,
2011.
[Khan et al. (2005)] Khan, M.M.; Iqbal, M.T.; Khan, F. (2005): Reliability and condition moni-
toring of a wind turbine. Canadian Conference on Electrical and Computer Engineering 2005.
Available at www.ieeexplore.ieee.org.
[Kjellberg (2009)] Kjellberg, H. (2009): Condition monitoring and preventive maintenance.
SKF Sverige AB, Presentation at the UK – Sweden Wind Power Forum, Stockholm, Sweden,
3rd December 2009.
[Kolerus (1995)] Kolerus, J. (1995): Zustandsüberwachung von Maschinen. 2nd ed., Expert
Verlag, Renningen-Malmsheim
[Koutoulakos (2008)] E. Koutoulakos: Wind turbine reliability characteristics and offshore
availability assessment. Master’s thesis, TU Delft, 2008.
[Kraemer and Fritzen (2008)] Kraemer, P.; Fritzen, C.-P. (2008): Damage Identification of
Structural Components of Offshore Wind Energy Plants. Deutsche Windenergie Konferenz
(DEWEK2008), Bremen, Germany, 2008.
[Kugellagerzentrale (2012)] Kugellagerzentrale, Website, accessed January 2012,
http://kugellagerzentrale.de/Rillenkugellager/SKF-6332M-C3.html
[Kumar and Klefsjö (1994)] Kumar, D., Klefsjö, B. (1994): Proportional hazards model: an ap-
plication to power supply cables of electric mine loaders. International Journal of Reliability,
Quality & Safety Engineering, 1, 3, p. 337-352
[Kung (2009)] Kung, P. (2009): Condition monitoring solutions for wind turbines. QPS Pho-
tronics, Presentation at CMVA 2009, Vancouver, Canada.
[Lading et al. (2002)] Lading, L.; McGugan, M.; Sendrup, P.; Rheinländer, J.; Rusborg, J.
(2002): Fundamentals for remote structural health monitoring of wind turbine blades—a
preproject. Risø National Laboratory, Roskilde, Denmark, 2002.
[Lawless (2003)] Lawless, J.F. (2003): Statistical Models and Methods for Lifetime Data. 2nd
ed., John Wiley & Sons Inc., Hoboken, New Jersey
[Lin et al. (2009] L. Lin, F. Sun, Y. Yang, Q. Li: Comparison of Reactive Power Compensation
Strategy of Wind Farm Based on Optislip Wind Turbines, SUPERGEN Conference 2009, 6-7
April 2009, Nanjing, China.
[Lindqvist and Lundin (2010)] Lindqvist, M.; Lundin, J.: Spare Part Logistics and Optimization
for Wind Turbines – Methods for Cost-Effective Supply and Storage. Master’s thesis, Uppsala
University, Sweden, April 2010.
119(122)

[Lucente (2008)] Lucente, M. (2008): Condition Monitoring System in Wind Turbine Gearbox.
Master’s thesis (in cooperation with NTNU, Trondheim, Norway), XR-EE-ETK 2008:004, KTH,
Stockholm, March 2008.
[Lynette (1985)] Lynette, R. (1985): Wind Power Stations 1984 Performance and Reliability.
EPRI AP-4199, R. Lynette & Associates, Bellevue, Washington.
[Lynette (1986)] Lynette, R. (1986): Wind Power Stations 1985 Performance and Reliability.
EPRI AP-4639, R. Lynette & Associates, Bellevue, Washington.
[Manwell et al. (2009)] Manwell, J.F.; McGowan, J.G.; Rogers, A.L.: Wind Energy Explained –
Theory, Design and Application. 2nd Edition, John Wiley & Sons, 2009.
[Mazzuchi and Soyer (1989)] Mazzuchi, T.A., Soyer, R. (1989): Assessment of machine tool
reliability using a proportional hazards model. Naval Res. Logistics, 36, pp.765-777, 1989
[Montgomery et al. (2006)] Montgomery, N., Lindquist, T., Garnero, M-A., Chevalier, R.,
Jardine, A.K.S. (2006): Reliability functions and optimal decisions using data for EDF primary
pumps. 9th Int. Conf. on Probabilistic Methods Applied Power Systems (PMAPS), Stockholm,
11-15 June 2006
[Moubray (1997)] Moubray, J.: Reliability-Centered Maintenance, Industrial Press Inc., New
York, USA, 1997, ISBN 0-8311-3078-4.
[Moubray (2000)] Moubray, J. (2000): The case against streamlined RCM. Download from:
www.relogica.com
[NASA (2002)] NASA (2002): Fault Tree handbook with Aerospace Applications. NASA Office
of Safety and Mission Assurance, Washington DC, USA.
[Nilsson (2006)] Nilsson, J. (2006): Maintenance management of wind power systems – Cost
effect analysis of condition monitoring systems. Master’s thesis (in co-operation with Dr.
Thomas Ackermann and Vindforsk), XR-EE-EEK 2006:008, KTH, Stockholm, April 2006.
[Nilsson and Bertling (2007)] Nilsson, J.; Bertling, L. (2007): Maintenance management of
wind power systems using Condition Monitoring Systems – Life Cycle Cost analysis for two
case studies in the Nordic system. IEEE Transactions on Energy Conversion, March 2007.
[Nilsson (2009)] Nilsson, J. (2009): On Maintenance Management of Wind and Nuclear Pow-
er Plants. Licenciate thesis, KTH, Stockholm, October 2009.
[Nilsson (2010)] Nilsson, K.: Torque estimation of double fed induction generator using a
dynamic model and measured data. Master’s thesis, Lund University, Lund, Sweden, No-
vember 2010.
[Nowlan and Heap (1978)] Nowlan, F.S; Heap, H. F. (1978): Reliability-Centered Mainte-
nance. AD-A066-579, United Airlines, San Francisco, CA, USA. Download:
http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA066579&Location=U2&doc=GetTRDoc.pdf
[Pawitan (2001)] Pawitan, Y. (2001): In all likelihood – Statistical modelling and inference
using likelihood. Clarendon Press, Oxford.
120(122)

[Petersson (2005)] Petersson, A.: Analysis, Modeling and Control of Doubly-Fed Induction
Generators for Wind Turbines. PhD thesis, Chalmers, Göteborg, Sweden, 2005.
[Quinlan (1987)] Quinlan, P.J. (1987): Reliability centered maintenance applied to wind park
operations. Proceedings of the Sixth ASME Wind Energy Symposium, Dallas, USA.
[Raadnui (2009)] Raadnui, S. (2009): Magnetic Chip Detector (MCD) wear particle analysis.
Department of Production Engineering, Faculty of Engineering King Monkut’s Institute of
Technology, North Bankok.
[Rausand and Høyland (2004)] M. Rausand, A. Høyland: System Reliability Theory - Models,
Statistical Methods, and Applications, Wiley-Interscience, 2004.
[Ribrant (2006)] Ribrant, J. (2006): Reliability performance and maintenance – A survey of
failures in wind power systems. Master’s thesis (in co-operation with Dr. Thomas Ackermann
and Vindforsk), XR-EE-EEK 2006:009, KTH, Stockholm, April 2006.
[Ribrant and Bertling (2007)] Ribrant, J.; Bertling, L. (2007): Survey of failures in wind power
systems with a focus on Swedish wind power plants 1997-2005. TEC-00272-2006, IEEE
Transactions on Energy Conversion, March 2007.
[RP] Communication with Rihard Pasaribu, SKF Engineering and Research Centre, Nether-
lands, 2011.
[Ruo and Olver (2008)] Ruo, P.; Olver, A.V.: Hydrogen in lubricated contact. Proceedings of
the 2nd PhD Conference on Artificial Intelligence for Industrial Applications, Hannover, Ger-
many, 14-15 May 2008.
[Scheffler et al. (1986)] Scheffler, R.L.; Wehrey, M.C.; Yinger, R.J.; Akins, R. (1986): Guidelines
for Testing Wind Turbines. EPRI AP-4682, Southern California Edison Co., Rosemead, CA,
USA.
[Schmidt (2010)] Schmidt, R. (2010): A New Phase in Turbine Condition Monitoring. M4 Wind
Services, available at windsystemsmag.com, April 2010.
[Seeliger et al. (2006)] Seeliger, A.; Schaaf, C.; Lachmann, J.K.; Messner, A.; Steinhausen, C.;
Schlecht, B.; Schulze, T.; Hermsmeier, J.; Nahrath, T. (2006): Combination of Multi Body Sys-
tem Simulation, Electrical Simulation and Condition Monitoring – A powerful research devel-
opment. Proceedings of the European Wind Energy Conference EWEC 2006, Athens, Greece,
May 2006, available at
http://www.ewec2006proceedings.info/allfiles2/876_Ewec2006fullpaper.pdf.
[SKF (2010)]: personal communication with Inge Aasheim, SKF
[SKF Reliability Systems (2000)] SKF Reliability Systems (2000): Vibration Diagnostic Guide.
Available at www.skf.com.
[SKF (1994)] SKF (1994): Bearing failures and their causes - Product information 401, 1994.
[SS-EN 13306] Svensk Standard SS-EN 13306: Underhåll terminologi - Maintenance
terminology. SIS Förlag AB, Stockholm, 2001.
121(122)

[Swedpower database] Felanalys. Database of failures for Swedish wind turbines 1989-2005.
Data compiled by SwedPower AB, Stockholm, on behalf of Statens Energimyndighet STEM
and ELFORSK.
[Timmerman (2009)] Timmerman, H. (2009): SKF WindCon - Condition Monitoring System
for Wind Turbines. Presentation at the New Zealand Wind Energy Conference 2009, availa-
ble at http://windenergy.org.nz/documents/conference09/HarryTimmerman.pdf.
[US DOE (2005)] US DOE (2005): Wind power today: federal wind program highlights. US
Department of Energy, Energy Efficiency and Renewable Energy, DOE/GO-102005-2115,
April 2005.
[Verbruggen (2003)] Verbruggen, T.W. (2003): Wind Turbine Operation & Maintenance
based on Condition Monitoring. Final report of the project WT-Ω, report no. ECN-C--03-047,
April 2003.
[Vestas General Specifications V42/V44-600kW (1995)] Vestas Wind Systems A/S (1995):
“Variabel Slip Vindmølle V42-600kW, V44-600kW” and “El-tekniske data Vestas V42/V44-
600kW”, Vestasvind Svenska AB, Falkenberg, Sweden.
[Vestas product brochure V90-2MW (2007)] Vestas Wind Systems A/S: V90-1.8MW & 2MW
– Built on experience. Product brochure, Randers, Denmark, 2007.
[Vestas product brochure V90-2MW (2009)] Vestas Wind Systems A/S: V90-1.8MW/2MW.
Product brochure, Randers, Denmark, 2009.
[Vestas product brochure V90-2MW (2010)] Vestas Wind Systems A/S: V90-1.8MW/2MW.
Product brochure, Randers, Denmark, 2010.
[Vestas General Specification V90-1.8/2.0MW (2005)] Vestas Wind Systems A/S: General
Specification V90-1.8MW/2MW OptiSpeed Wind Turbine, Randers, Denmark, 2005.
[VGB PowerTech (2003)] VGB PowerTech (2003): Richtlinie EMS – Ereignis-Merk-mal-
Schlüsselsystem. VGB-B 109, Germany, 2003.
[VGB PowerTech (2007)] VGB PowerTech (2007): Guideline Reference Designation System
for Power Plants, RDS-PP - Application Explanations for Wind Power Plants. VGB 116 D2, 1st
Ed., Germany, 2007.
[Vlok (1999)] Vlok, P.-J. (1999): Vibration Covariate Regression Analysis of Failure Time Data
with the Proportional Hazards Model. Master’s thesis, University of Pretoria, South Africa
[Vlok (2001)] Vlok, P.-J. (2001): Dynamic Residual Life Estimation of Industrial Equipment
based on Failure Intensity Proportions. PhD thesis, University of Pretoria, South Africa
[Watson and Xiang (2006)] Watson, S.; Xiang, J. (2006): Real-time condition monitoring of
offshore wind turbines. Proceedings of the European Wind Energy Conference, Athens,
Greece, May 2006.
[Wind Turbine Gearbox Inspection Report (2009)] ZF Service GmbH: Wind Turbine Gearbox
Inspection Report, provided by Göteborg Energi / Triventus, 2009.
122(122)

[Yang et al. (2010)] Yang, W.; Tavner, P.J.; Crabtree, C.; Wilkinson, M. (2010): Cost-Effective
Condition Monitoring for Wind Turbines. Transactions on Industrial Electronics, Vol. 57(1),
January 2010.

View publication stats