You are on page 1of 22

Articles on the state of the art of developments in building

services engineering
Welcome to the ’Blue Pages’. This is where guest editors from the Editorial Advisory Panel for Building
Services Engineering Research and Technology address current developments in building services
engineering practice and research. The articles are very short, on one theme of current interest, and do
not go through the longer refereeing process for conventional research papers. This is to encourage
consultants and contractors to discuss their latest developments in a non-commercial manner.
Academics will also be asked to outline current research areas in universities and colleges. Without the
constraint of full refereeing there is scope in the articles to provoke interest and to raise issues for
discussion. Consequently the Editor of BSER&T is happy to receive and publish letters furthering the
discussion of ’Blue Pages’ articles. If you have a suggestion for an article or topic you would like to see
included in the ’Blue Pages’ please inform the Editor, Barry Copping. Comments on the ’Blue Pages’ are
also welcome. ’Blue Pages’ will only succeed if readers of the journal wish the feature to succeed and
contribute to its success.

Identifying faults in HVAC systems


Introduction ing and ongoing performance monitoring, in real applica-
tions.
The short articles published here offer an overview of the var-
A L Dexter (Department of Engineering Science, University ious aspects of the research activities. The first article sum-
of Oxford, UK) marises the results of surveys of practitioners undertaken in a
number of countries to determine what are thought to be the
This issue of the ’Blue Pages’ is devoted to work undertaken most common and the most important faults in differcnt
for a major international research project entitled Annex 25: types of HvAC systems. The second article discusses the eco-
Real-time simulation of HVAC systems for building aptirrrisation, nomic and other benefits of installing a fault detection and
fault detectton and diag1wsis, as part of the International Energy diagnosis system. The following five articles provide intro-
Agency (IEA) Energy Conservation in Buildings and ductions to the main types of fault detection and diagnosis
Community Systems programme. The project was co-ordi- methods that were considered. The last three articles present
nated by the Laboratory of Heating and Ventilation, some of the results of applying FDD schemes to heating sys-
Technical Research Centre of Finland (VTT) and involved tems, air-conditioning systems and chillers, and discuss the
over forty research engineers from building controls manu- advantages and disadvantages of the different approaches.
facturers, plant manufacturers, government research laborato- Fuller descriptions of the results of the work will soon be
ries, and educational institutions in fourteen countries. The published by the VTT in a two-volume final report. The first
United Kingdom was represented by the Building Research volume is a ’source book’ which introduces the basic tech-
Establishment and the Universities of Loughborough and niques ; the other volume is a collection of technical reports
Oxford. which describe in detail the FDD schemes, the practical appli-
cations, and the results of the joint evaluation exercise.
The original aims of the research were:

(a) to create a database of the faults that occur in different Reference


types of HvAC systems (hot water heating systems, heat
1 HyvärinenJ and Kohonen R (eds) Building Optimisation and Fault
pumps and chillers, air-conditioning systems and ther-
mal storage systems); Diagnosis System Concept (Espoo, Finland: Technical Rescarch Centre of
Finland (VTT) Laboratory of Heating and Ventilation)
(b) to investigate methods of detecting and identifying dif-
ferent types of faults, and of optimising performance;

(c) to evaluate the most suitable modelling approaches for Survey of typical and important faults
the real-time simulation of HVAC plant in fault detection
and diagnosis (FDD) schemes. Juhani Hyvdrinen (VTT Building Technology, Finland)
A framework for the Annex was defined during a one-year
preparation phase, before the three-year working phase was 1 Introduction
started in 1991. The project was subsequently extended for a
further year to allow comparative testing of the FDD schemes The faults in a building may occur on many levels. The
on a number of reference HVAC systems. The IEA has recently
building’s technical systems consist of subsystems which in
approved the launch of a follow-up project entitled turn consist of components. A fault at any of these levels can
Computer-aided evaluation of HvAC system performance: The cause degradation in the building’s technical performance
practical application of fault detection and diagnosis tech- and technical availability (Figure 1).
niques in real buildings. The aim of the new project is to
work with industrial partners, such as building controls man- The number of different components, subsystems and sys-
ufacturers, and building owners and operators, to demon- tems in a building is large. It is not feasible to develop a
strate the benefits of on-line evaluation, for both commission- method for each of these. Attention must be focused on the

B13
-

energy and water savings: inefficiently operating processes


can be detected earlier than otherwise;
-
increased quality of living: faults can be detected before
they affect the indoor quality or the quality of other
’products’ produced by the HVAC system, and the mainte-
nance work can be scheduled so that it does not cause

unnecessary inconvenience to the occupant;


-
reduced maintenance costs when the maintenance work can be
planned beforehand: faster location of faults, scheduling of
work, usage of maintenance staff with appropriate skills,
ordering of spare parts, and shorter process down time;
Figure 1 Tasks and objectives of building optimisation -

health and safety: detecting the faults early decreases the


likelihood of personal injury and damage to equipment
most important in terms of some (inevitably subjective) eval- and property, and reduces the heath risk which may be
uation criteria. Also, the knowledge of fault modes and the associated with processes operating incorrectly.
ways in which faults appear to the user is important. Examples of the criteria used in assessing the important faults
Therefore, essential matters that must be resolved in develop- in thermal storage systems are given in Table 1. Relative
ing fault detection and diagnosis methods and systems are: importance is obtained from the responses to the question-
---
to identify those central entities (subprocesses and com- naire.
ponents) on which the development of the fault detection Table 1 ’A fault is assessed to be important because ..’
and related methods should focus;
-

to collect data on existing ways in which faults are detect-


ed ;
-
to collect data on how possible faults are remedied.

This information can then be used in the various stages of


designing and building a fault detection and diagnosis sys-
tem.

2 Results

In Annex 25 this task was one of the first. The objective was
to define typical HvAC system descriptions and to produce an
off-line database on the problems and faults in these systems.
The task consisted of preparing functional descriptions of
HVAC systems, and of analysing these reference systems with 3 Typical faults
different techniques. The systems considered were heating
For the reasons explained above it is obvious that the list of
systems, air handling systems, chillers and heat pumps, and
thermal storage systems. typical or important faults is a subjective matter. However, in
the following, abbreviated lists are presented for each refer-
There were two major outcomes from the analysis. Firstly, ence HVAC system to give some idea of the nature of the
component and fault lists were compiled for the reference sys- results obtained.
tems. These lists give an idea of what kinds of faults exist in
the systems that were studied, and also catalogue the faults
3.1 Fleating systems
generally considered important.
Two surveys were carried out for entire heating systems
Secondly, it became clear that the ranking of faults is strongly
dependent on the viewpoint of the individual or the group of (France/CSTB, Germany/University of Stuttgart) and two

individuals making the assessment. For example, in the types of heat generation systems were analysed by a group of
analysis ofVAV air handling system the results were grouped experts (Finland/VTT). The heating systems comprised heat
according to three different professional groups: designers, generation subsystems, a heat distribution network and the
radiators in the rooms. The important faults in heating sys-
constructors, and maintenance personnel. Different lists were
tems can be categorised as either design or operating faults.
obtained as the outcome.
The main operating faults are:
As well as varying by profession, the results of ranking varied
according to building type. In the heating system analysis, for
-

heating curve badly tuned


example, the evaluation criteria were weighted differently --
incorrect calculation of optimum start or stop by the
according to building type. Heat and warm water must be operational mode controller
available in hospitals, for instance. This led to a different
ranking as compared with that for a building in a rural area
-

poor combustion
where the services personnel are not easily available and early -

scaling and dirt accumulation in the boiler’s heat


warning of developing faults may be necessary. exchanger _

The main benefits to the ’end user’ of applying an FDD system -

leakage in the three-way/two-way valves controlling the .

are as follows: secondary circuits


B14
-

deposit of scale in the domestic hot water heat exchanger. Cost-benefit analysis for FDD systems
3.2 Airhar~gs~rnscmd thermalstarag~esy~rm
J E. Braun (Ray W Herrick Laboratories, Purdue University,
An exhaustive analysis was carried out for both air handling systems USA)
Gapan/Nagoya University, USA/NIST, Switzerland/Landis&Gyr)
and for thermal storage systems JapwNagoya University). The Costs are a critical factor in the development and application
analysis was based on questionnaires returned by professionals. Two of fault detection and diagnostic (FDD) systems. In the devel-
examples are shown in Tables 2 and 3. The format follows the prac- opment phase, the costs associated with sensors and electron-
tice used in reliability engineering. ics must be balanced with the benefits that the FDD system
will provide. The end product will not be applied unless the
33 Challers mrd he~t pum~s consumer believes that the cost savings associated with the
FDD system will be greater than the installed costs.
Vapour compression refrigeration equipment forms the largest part
of installed commercial refrigeration capacity. An extensive list of The installed costs include the costs of sensors, electronics,
vapour compression cycle machine faults was compiled by a group wiring, and software setup. The benefits associated with an
of experts (Canada/CANMET, The Netherlands/TNO, FDD system include:
Switzerland/ETH and SEC Sulzer Energieconsulting, -

reduced service costs


JqnmNapp University). Three faults were considered to be of the
highest priority due to their friequency and to their adverse effects on -

improved comfort
equipment efficiency and mechanical integrity. These faults are: -
reduced energy costs
-
lack of refrigerant -

longer equipment life


-

the presence of air in the refrigerant circuit -

improved safety
-
the presence of refrigant in the lubricating oil. -

reduced environmental impact.


Automation can reduce service costs by reducing the man-
4 Conclusions power required to monitor and diagnose problems and by
eliminating unnecessary service tasks. Comfort can be
Analysing HvAC systems from the reliability point of view gives improved through earlier detection of comfort problems and
some insight into where to concentrate the development of FDD sys- reductions in equipment downtime. Improved maintenance
tems. The fault lists given here are only examples but give some idea scheduling can improve the equipment operating efficiencies
of the kind of information that can be obtained. The information is and reduce energy costs. An FDD system could also detect and
mainly qualitative but there are also methodologies for quantitative possibly avoid operating conditions that would compromise
analysis. equipment life (e.g. liquid entering a compressor), safety (e.g.
too high a condensing pressure), and the environment (e.g.
Apart from generating a catalogue of faults, a valuable result has leaking refrigerant).
been the list of criteria applied in their assessment. The criteria may
have wider applicability than the exercises carried out in Annex 25. The installed costs could be easily quantified after FDD sys-
It would also seem that the relative importance of the criteria is the tems are fully developed and have been installed in several
same in the different evaluations since, in all cases, the criteria con- applications. However, the cost savings associated with FDD
sidered to be of greatest significance were those relating to energy systems are much more difficult to quantify, particularly
consumption, environment conditions and frequency of occurrence. those costs not associated with service and energy.

Table 2 Typical faults in air handling units

Table 3 Typical faults in thennal storage system

B15
Matsunawa et a1.~2~ describe three different approaches for per- costsdue to fouling approach zero asymptotically as the foul-
forming a cost-benefit analysis for FDD systems. In the first ing time increases. In the limit, the fouling goes to zero and
approach, the decision to develop or apply a FDD system is the optimal scheduler reduces the service frequency to zero.
based upon the evaluation of a positive life-cycle saving asso- In contrast, the extra cost of the regular service schedules
ciated with those benefits for which cost savings can be deter- approaches the annual service cost as the fouling time increas-
mined. A secondary qualitative evaluation of the additional es. The additional cost for the constrained service schedule
benefits is used to provide additional justification for the FDD also approaches zero as the fouling time increases, but at a
system. The second approach is similar to the first except that slower rate than the optimal scheduler. For a fouling time of 5
cost savings must be assigned to all benefits associated with a years in Figure 2, the minimum operating costs are about a
FDD system. The third method is also similar to the first third of those for twice per season service and about two-
except that weights are assigned to both the quantifiable (i.e. thirds of those for once per season and constrained service.
cost savings) and non-quantifiable benefits to determine a
decision regarding the applicability of an FDD system. In a The regular maintenance schedulers always provide excessive
case study, Takesa~3> demonstrated the first cost-benefit service whenever the service interval is short enough to
ensure that comfort conditions are maintained. For the
approach and studied the relationship between the life-cycle results of Figure 2, the twice/season regular service scheduler
savings and different assumptions regarding energy and maintains comfort for shorter fouling times than the
maintenance savings. However, it is difficult to draw any con-
clusions regarding the economic benefits of FDD systems from once/season schedule at the expense of higher service costs for
this study since the energy and maintenance savings were longer fouling times. The costs associated with the simplified,
only hypothesised. In a recent study Rossi and Braun(4) esti- near-optimal scheduler are close to those for the optimal solu-
mated the combined energy and service savings associated tion under all conditions. The average difference between the
with optimal scheduling of the cleaning of condensers and optimal and near-optimal schedules is less than 0.2% for all
results presented in Figure 2.
evaporators for a rooftop air conditioner. In addition, a simple
algorithm for estimating optimal service times was developed. The results of Rossi and Braun only demonstrate the cost
Figure 2 shows sample results from this study associated with benefits associated with FDD for heat exchanger fouling.
evaporator fouling. The operating costs, plotted on the verti- Undoubtedly, more significant cost savings could be achieved
cal axis, are normalised using the base operating cost with no for FDD applied to more severe faults. In order to perform
fouling. The base cost is the total annual energy cost with no cost-benefit analysis properly, additional research is necessary
fouling for cooling. The normalised cost is the actual cost to evaluate the cost savings associated with all benefits pro-
minus the base cost divided by the base cost and represents vided by an FDD system.
the fractional extra cost due to reduced efficiency and
increased service associated with the fouled heat exchangers.
The horizontal axis is a characteristic fouling time defined as
the calendar time required for the evaporator to foul com-
pletely. 2 Matsunaw K, Iizuka H, Itoh H, Miyasak F and Takase T Cost Benefit

All the results shown in Figure 2 were generated for service Analysis and Economic and Environmental Evaluation in IEA Annex 25:
Building Optimization and Fault Diagnosis Source Book (1995)
costs of $60 per cleaning and energy costs of $0.10 per kWh.
3 Takase T Actual Application of Cost Benefit Analysis by LCC Method in
Minimum operating costs are compared with those for two IEA Annex 25: Building Optimization and Fault Diagnosis Source Book
regular service intervals (representative of current practice), (1995)
those for a strategy where service is only performed when a 4 Rossi T M. and Braun J E Minimizing Operating Costs of Vapor
constraint is violated (e.g. a comfort violation), and those for Compression Equipment with Optimal Service Scheduling Int. J.
the simplified, near-optimal strategy. As expected, Figure 2 Heating, Ventilating, Air-Conditioning and Refrigerating Res. 2(1) (1996)
demonstrates that the optimal solution always results in lower
costs than in the alternative service strategies. The savings
associated with optimal maintenance scheduling increase
with fouling time. For optimal scheduling, the additional

Innovative approaches to FDD


H Yoshida (Kyoto University, Japan)

1 Introduction

A real HVAC system (or its components) can be modelled by a


mathematical formula with some simplifications, and the sys-
tem output under normal operation can be estimated on the
basis of the model. Since the response of the system changes
when there is a malfunction or when a fault occurs, the mea-
sured output from the real system will deviate from the out-
put estimated by the model. The fault, therefore, can there-
fore be detected by carrying out statistical tests on the values
of the difference between the two outputs (the residuals or
innovations). The better the model represents the system
characteristics, the higher the reliability and the faster the ’

Figure 2 Cost comparisons for alternative maintenance strategies fault detection.

B16
A straightforward model structure can be obtained if the 2.2 State-space model approach
characteristics of each component in the system are described
by equations derived from the basic laws of physics (e.g. heat By utilising the knowledge of the structure of a system and its
transfer theory). However, reliable values of the model para- components, and assuming linearity, the dynamics of the sys-
meters are often not available because either the design data
tem can be expressed by the following state-space modelt5~:
or the manufacturer’s data are too idealised to describe the
actual operation. In practice, an identification process is usu-
ally required to estimate the parameters of the model from where x is the state-space vector, v and w are linear transfor-
operational data. Fault detection based on testing whether or mations of the unknown input disturbances d to the system,
not the identified model parameters change is also possible.
and A, B and C are known matrices. For the discrete time
Since precise simulation of system behaviour is not the ulti- case, the state space vector x and output y can be estimated
mate goal in fault detection, a simple model can well be used optimally using a Kalman filter. When nonlinearity cannot
in most cases. Even very simple ’black-box’ models, which are be ignored, the model must be formulated with nonlinear
built without any knowledge about the physical characteris- terms. In this case an extended Kalman filter can be used to
tics of components, can be used. Furthermore, models may be estimate x andy. As in the ARMAX case, a statistical test of the

built on the basis of the dynamic characteristics or of the stat- innovations


ic characteristics of the system. Thus there may be several
alternatives when modeling a system and an appropriate
model should be selected considering merits and demerits of
each approach. can be used to detect the presence of a fault.
There are two basic methods of fault diagnosis using innova- 2.3
tions. One is based on rules acquired by analysing the charac- Physical model approach
teristics of the innovations when different types of faults are Each component of the HvAC systems is modelled by one or
present. The other is a multiple fault approach in which it is more equations derived from a theoretical understanding of
assumed that only one of n defined possible faults can be pre- the physical processes occurring in the system (e.g. heat and
sent at a given time and that a model can be identified which mass balances)<6>. Generally the equations of the model are
describes the behaviour of the system in the presence of each quite sophisticated. For example, effectiveness of a counter
fault. The actual fault is diagnosed by observing which model flow cooling coil might be formulated as follows:
generates the smallest innovations. c =
[1-exp(-rrt’tr{1-w))]I[1-wexp(-rrru(1-w))]
where t~’rv is the number of transfer units and w is the heat
2 Different approaches
capacity ratio. As explained earlier, the design or manufactur-
ers’ values of these parameters usually need to be refined
2.1 ‘Black box’niodel approach using a constrained optimisation technique and operational
data.
The autoregressive exogenous (ARX) and the autoregressive moving
average exogetww (ARMAX) models are typical examples of the
’black box’ type of model. The ARMAX is more sophisticated 3 Advantages and disadvantages of the approaches
than the ARX model and the number of parameters is general-
ly smaller. These models are commonly used because of their If the occurrence of a fault changes the dynamic behaviour
simplicity. The mathematical expression of the ARMAX mod- significantly, the ARMAX or the state-space model approaches
els is as follows: can be used successfully for fault detection. However, since
operational data must be processed in real time, the costs
associated the data acquisition and transmission system will
be high. Nonlinearity can be handled in both approaches but
the feasibility of applying these approaches to real plant needs
to be investigated because little experience has been accumu-
where n is the measured output at time 11, a- are the autore- lated so far in the field of ttvAC systems. In terms of model
gressive parameters, M is the order of the autoregressive part, building, the ARMAX approach is very simple and robust, but
un is the measured input at time n, bi are the exogenous para- its diagnostic capabilities are limited because no knowledge of
meters, K is the order of the exogenous part, c; are the moving the physical structure is used and only innovations are tested.
average parameters, N is the order of the moving average part, In the state-space or physical model approaches, initial values
en is the residual at time n (considered to be zero mean white of the parameters have to be specified, even though they may
noise) When N 1, the equation denotes an ARX model.
=
be refined during an operational stage. This may prove to be
Mathematically this model is a linear system; however, sys- the major disadvantage of these approaches, since it may be
tem input nonlinearities can be handled by incorporating
very diffcult to obtain this information from designers and
nonlinear terms in the input u. The predicted output by the
model is obtained as follows.
plant manufacturers.

where~I(z’) = 1 + alz-1 + + aAlz-JI and z-1 denotes a back-


...

5 Yoshida H, Iwami T, Yuzawa H and Suzuki M Typical faults of air-


ward shift operator, for example ~’xn xn-1. The terms B(z-1)
=

conditioning systems and fault detection by ARX model and extended


and C(z’) are defined similarly. A common criterion for Kalman filter
ASHRAE Trans. Pt 1102 (1996)
determining the best model or the best parameter set is the 6 Haves P, Salisbury T I and Wright J A Condition monitoring in HVAC
least-square criterion which requires minimisation of the subsystems using first-principles models ASHRAE Trans. Pt 1 102
residuals<5>. (1996)

B17
Parameter estimation approaches to FDD order to reduce the modelling and measurement errors at
high frequency.
Advantages
L G Oisson and P E Blomberg (KTH, Division of Building
Services Engineering, Stockholm, Sweden) -

The results can be interpreted simply.


-
The occurence of simultaneous faults can be handled.
1 Introduction Disadvantages
-

Long computing times may cause significant delays in


Parameter estimation methods cover a very broad range of
response to a fault in on-line applications.
procedures, from those which search through dynamic mod-
els with given structures, to direct calculation of the parame- -
For satisfactory diagnosis the system needs to be contin-
ters of analytical steady-state models or expressions. In all uously disturbed.
cases the methods make extensive use of measured data.
Three quite different approaches are described briefly in this 3 Characteristic curves
article; their common feature is the use of thresholds for deci-
sion making. Characteristic curves are visual representations of the behav-
iour of physical processes which are based on measured data.
2 Physical models Manufacturers often use such representations to describe the
properties of their products, rather than to provide design
Parameterised physical models are used in which the parame- information. However, such information is usually restricted
ters serve as indicators of a predefined set of faults in each to steady-state behaviour. -
component or subsystem. The values of the parameters are Generally the characteristic curves are polynomial fits, often
estimated by iterative calculation, as shown in Figure 3. in quadratic form. Physical insights into the function of the
component are exploited to help structure the measured data,
and mathematical models are sometimes used to extrapolate
the operating range. Thus characteristic curves work as an
efficient means of defining the correct behaviour of compo-
nents and, if properly structured, can serve as a reference
when analysing in situ measurements recorded during the
normal operation. An FDD procedure of this type is shown in
Figure 4.

Figure 3 Block diagram for the FDD procedure. V~,(p) is the objective func-
tion, y,~f(t) are the predicted outputs and x(t) and yet) denote the measured
inputs and outputs respectively. M(~ p, DD) is a continuous-time model.

Changes in the parameter values, from values which are rep- Figure 4 Block diagram of FDD procedure based on characteristic curves. x,
resentative of correct behaviour, indicate various faults and and xz are measured signals; ~ is the physical quantity (or quantities) of
how serious these are. Parameters may be associated with interest. Subscript c denotes stored values and M the corresponding value

physical constants as well as variables; faults may be specified predicted by the model.
by single parameters or combinations of them. The changes
in the parameters can be used for both detection and diagno-
sis. Taking a coil with control valve as an example, the follow-
ing quantities may be defined:
-
the estimated temperature effectiveness of the coil All measured signals used must be close to steady state values.
-
The normalised measured flow rate through the pump Depending on the models x, and x~ may include common
measurements, and (D and 4),, the stored and current values
It is assumed that the parameters are almost constant during of the features, may represent several physical quantities, par-
each sampling period. The model M(x, p, DD) is based on heat ticularly in a complete mapping of the behaviour of a compo-
balances and approximations to the dynamic effects and the nent.
behaviour of the valve. The output signalsyM(t) are the air
It should be noted that the parameters are normally constant
and water outlet temperatures. For correct behaviour, the
and faulty behaviour is detected by a discrepancy between (k
parameters should take a value of approximately unity. When and 4$~~. However, further analysis is often needed to identify
both the air and water side responses are monitored, the para-
the particular fault. This might be carried out by systemati-
meters used for detection can also be used for diagnosis.
cally combining information from the various characteristic
The parameters are estimated using existing software that functions, or by comparing measured and expected values of
minimises an objective function ~N defined by the sum of the important signals, such as the fluid temperatures and their
squares of the differences between measured and predicted setpoints. Taking a coil subsystem as an example, the follow-
outputs. It is usually advantageous to filter the residuals in ing quantities may be defined:
.

B18
-

<Pc is the correct temperature effectiveness (both water- real system. In the example considered
here, faults like fluid
side and air-side effectiveness may be used) stored as a leakage or blockage and fouling may be resolved by using the
function of flow rates x,. FDS.

- <PM is the current air-side effectiveness calculated from The directions associated with particular faults will depend
the input and outlet temperatures X2 and the definition of on the ordering and the definition of the cps. This
procedure
effectiveness. can be extended to any arbitrary number of parameters. It

might also be appropriate to associate fault cases with particu-


Comparison of (Dc and ~41 for the air side might then be used lar sectors rather than directions in the ~s.
to detect a possible fault while the water side values must be
considered for diagnosis, unless the differences are very large. The diagnosis is performed in two steps. First the type of
fault is identified by comparing the directions (or sectors)
Advantage with those stored for each fault case and then the size of the
-

provides a simple method of on-line fault detection fault is determined by calculating the distance from the ori-
gin. At each stage the statistical uncertainties, which arise
-
uses easily accessible measured data to define correct from poor estimation of the steady-state features from mea-
behaviour. sured data, the sensitivity of parameters to modelling errors
Disadvantages etc., must be taken into account by normalising the FDS
according to the likely errors and introducing appropriate
-
relies on near steady-state conditions thresholds.
-
can result in complex diagnostic routines for some sub- Advantages
systems. -

allows a predefined set of faults to be considered

4 Characteristic parameters (Fault direction method)


-

diagnosis is performed on-line.


Disadvantages
This approach is also based on steady-state physical or empir-
ical models of the behaviour of the system to be analysed. A
-

steady-state data must be extracted from the measure-


ments
set of characteristic parameters (crs) is defined such that the
parameters are almost constant during normal operation but -

requires a large number of sensors to give a complete


change when a fault occurs. By structuring such changes mapping of a system.
properly, a fault direction space (FDS) can be formed, where
the direction of the vector of &eth;CPs indicates the presence of a
particular fault, and the distance from the origin (the fault-
free condition) indicates the severity of the fault. The behav-
iour of a heat exchanger or coil is considered to illustrate the
method. Two steady-state heat balance equations are used to Classification approaches to FDD
generate two parameters: cPt f 1{xt) and CP2=!;(X2)’ where x,
and x2 are sets of measured signals (e.g. temperatures and M Stylianou (EDRL, CANMET)
pressure drops); xt and x~ may include common measure-
ments. The ftinctionsf, and f2 are model-based dimensionless
quantities or parameters, which are constant during fault-free 1 Introduction
operation, and vary in a unique way in the presence of a par-
ticular fault. The corresponding two-dimensional FDS is Generally fault detection and diagnosis systems employ one
shown in Figure 5. of the following two approaches:
A FDS which has been predefined by means of laboratory (a) Estimation
experiments or computer simulations of a typical plant, with (b) Pattern recognition.
and without faults, can serve as a reference for the tests on the
Estimation-based procedures are used in systems which
acP2 employ physical or statistical models to estimate relevant
process variables. In general, process system models are used
to estimate the values of the process variable. These values are
subsequently compared with measured process variables, and
the resulting differences are supplied to a separate module
where the diagnosis is made.
Fault detection and diagnosis using pattern recognition, on
the other hand, can be characterised as a set of procedures
X~~~. ACPI
which extract recognisable patterns present in the output data
V4 of systems. Each structure or pattern is associated with a par-
ticular fault which the pattern recognition algorithm or pat-
tern classifier can ’leam’ to recognise.

The classifiers, in general, ’learn’ the different faults using


data from known examples (’training data’). This procedure,
Figure 5 FDS for the heat exchanger example; ACPI and ACP2 are the also called ’supervised learning’, aims at minimising the clas-
observed changes and J fi the corresponding fault vectors (The circles indicate sification error by modifying the classifier’s parameters. Once
the statistical uncertainties.) the classification error has been reduced to a predetermined

B19
threshold, learning is complete and the classifier is ready to systems custom-made. This is a problem faced by all clas-
are
test a new data set. sifiers using supervised learning, and since each
equipment/system is different, solutions need to be found
The following three classifiers, each of which has been used accordingly.
by participants in the Annex 25 project, will be described: the
Bayesian classifier, the multi-layered perceptron-based classifier These classifiers also require knowledge of the probabilities
and a case-based reasortiaig classifier using topological case-based with which particular faults occur. This requirement may
nwdelling (TCBM). pose some problems since the probability with which particu-
lar equipment or system faults occur is generally unknown.
Although it may be estimated, the effect of this on the perfor-
2 Bayesian classifiers mance of the classifier is not easily quantified.

This family of classifiers is based on the Bayes decision rule.


This rule, originally formulated by Thomas Bayes 3 Artificial neural network classifier
(1702-1761), states that a measurement vector x belongs to a There are a number of different types of neural networks<7>,
class (or fault) c if and only if the conditional probability of
the fault (class) c given x, is higher than the conditional prob- several of which have been used as classifiers for fault detec-
abilities of the other faults. tion and diagnosis. Of these the most widely used is the mul-
tilayer perceptron (MLP) trained using the back-propagation
The conditional probabilities are computed using the Bayes training algorithm.
theorem which requires knowledge of the probability of
The training in this type of classifier is performed by intro-
occurrence of individual faults (the a priori probabilities) and
--

the joint probability density functions of c and x. As these ducing the input vector to the nodes forming the input layer i
as shown in Figure 7, while the class/fault vector is intro-
joint probability density functions are unknown, the Normal duced to the nodes in the output layer k. In the case of fault
distribution is assumed. This assumption facilitates the devel-
detection and diagnosis, the input vectors introduced in layer
opment of the classifier, but the parameters of the normal dis- i are real-valued numbers, while the class/fault vector which is
tribution, i.e the means and variances, need to be estimated. the output of layer k takes values signifying the presence or
These means and variances are estimated using training data, absence of a fault. The nodes in layer k are therefore trained
i.e. data representing known faulty conditions. The training to take, say, 0 if the fault is absent and 1 if it is present.
data are thus reduced to their means and variances and are
stored for eventual use when a new data set is introduced for
classification. The classification decision, may be viewed as
establishing the ‘closest’ fault to a new data set x. Figure 6
shows two statistical distributions, one representing Fault 1
and a second representing Fault 2. Classification in this case
is achieved by computing a (statistical) ’distance’ between x
and the mean of each of the faults present in the training data
(ZI and z2). The fault which is ’closest’ to x is then selected as
the most likely fault x belongs to.

2.1 Advantages and disadvantages


The practical implementation of Bayesian classifiers generally
assumes normally distributed data, and requires training data
for each of the faults the classifier is to recognise. This train-
ing data requirement can pose problems, since in many cases
the equipment and systems present in building mechanical
Figure 7

As the input vector is propagated through the network, it is


modified by weights present in the node connections as well
as by the (linear or nonlinear) functions associated with the
nodes. Once this vector reaches the output layer k, it is com-
pared with the desired output, namely the class/fault vector.
The difference between the desired output vector and the
actual one is computed and the resulting error vector is prop-
agated backwards and is used to modify the network parame-
ters. Repeated exposure to the training data reduces the error
to a predetermined threshold. Once this threshold is reached,
the network is said to have converged and is ready to classify

new input vectors.


B20
3.1 Advamages and disadvantages Expert system approaches to FDD
Several neural network topologies and training algorithms
have been used for classification, each having its own set of K Baer (University of Stuttgart, Germany)
advantages and disadvantages. In general however, these do
not assume any statistical distribution and have been shown
to perform well in both linear and non-linear classification 1 Introduction
problems. The three-layer perceptron, using the back-propa-
gation training algorithm, in particular, has been shown to be Fault diagnosis in complex systems is a traditional and very
able to classify input vectors into classes whose complexity is successful application of expert SySteMS(12). One of the best
only limited by the number of nodes in the hidden layer<’). known experts systems of this domain, Mycin, supports med-
This aspect in the design Of MLPS, i.e. an optimum number of ical diagnosis of meningitis and other bacterial infections. It
nodes in the hidden layer, is still a matter of trial and error. If was the first large expert system to reach the level of human
an insufficient number of nodes is selected, the MLP is under- experts and had a strong impact on further developments in
fitted, limiting its capability with complex classification prob- expert systems<’ 3>.
lems. If, however, too many nodes are selected, the MLP is
overfitted. This leads to unpredictable results when the input An essential characteristic of expert systems is the strict sepa-
vector does not correspond closely to the training data. ration of expert knowledge of a given domain from general
knowledge about problem solving. The most important issue
is the definition of the knowledge base, which represents the
4 Topological case-based modelling (TCRM) domain-specific knowledge in a well structured and for-
malised form. The evaluation of the knowledge the process
-

TCBM(9) is anapproach which has strong similarities to near- of problem solving -

is performed by an inference engine,


est-neighbour-based algorithms such as instance based karning which links the supplied knowledge in a suitable way, draws
(IBL)CIO). These types of algorithms are characterized as memo- conclusions upon it and attempts to arrive at the solution in a
ry-based algorithms in that, unlike the Bayesian classifiers, step-by-step manner. Most systems are designed to deal with
they do not describe the particular problem as a statistical uncertain knowledge in the form of incomplete or inconsis-
model, but rather as a collection of cases (memories). More tent facts.
specifically, whereas Bayesian methods summarise the data
technical considerations in the efficient development
statistically, memory-based reasanirtg (MBR) algorithms use spe- Major
cific cases to perform classification. of expert systems are how to acquire knowledge from the
experts, how to express and formalise this knowledge and
Generally MBR algorithms assume that similar cases have sim- how to select the most suitable method for reasoning about
ilar classifications belonging to particular ’neighbourhoods’. the given knowledge(&dquo;). Many different techniques have been
These algorithms use some measure of similarity between the proposed to deal with each of these issues (see Figure 8).
existing cases and a new case. This measure is computed for Details are available in the literature (e.g. Reference 12).
each of the cases present in the case-base ’neighbourhood’ and
passed on to the classification function. The classification
function, in turn, using the ’topology’ of the case-base, classi- 2 Approaches to FDD
fies the new data. Finally the case-base is updated on the basis
of the success of the individual cases belonging to the particu- Symptoms are the noticeable or measurable effects of faults. A
lar ’neighbourhood’. fault may manifest itself in one or more symptoms. If it is
I
..... -.

4.1 Advantages and disadvantages


The TCBM algorithm provides an alternative to the previous
two classifiers in that, unlike the Bayesian approach, it does
not require knowledge of the statistical distribution of the
data to arrive at classification. In addition, unlike the MLP
classifier, it is transparent and provides the user with infor-
mation that would be lost if the MLP classifier were used,
namely the related faulty conditions. However, this type of
classifier requires significantly more memory to process each
new case which, in high-dimensional problems, may pose dif-
ficulty to the processor carrying out the classification.

7 Dubuisson B and Vaezi-Nejad H Introduction to Formal Neural


Networks in Building Optimisation and Fault Diagnosis Source Book
ed.Hyvarinen J(1996)
8 Lippman R P An Introduction to computing with Neural Nets IEEE
ASSP Mag. 4 (1) 4-22 (1987)
9 Tsutsui H Topological Case Based Modelling in Building Optimisation and
Fault Diagnosis Source Book ed Hyvarinen J (1996)
10 Aha D W, Kibbler D and Albert M K Instance-Based Learning
Algorithms Machine Learning 6 37-66 (1991)
11 Fukunaga K Introduction to Statistical Pattern Recognition (New York: Figure 8 Overview of commonly used techniques for implementing expert
Academic) (1990) systems

B21
assumed that other software is used to preprocess the mea- 2.3 Knowledge-based approaches
sured data so as to detect the occurrence of a fault (e.g. by
The term ’knowledge based approaches’ covers all expert sys-
comparing simulated or reference data to measured data), the tems which use other types of knowledge representation
primary concerns of the fault diagnosis scheme are how to schemes rather than facts and rules. Such systems are often
represent the symptom/fault relationship and how to infer the called ’frame-based systems’. The information in the knowl-
associated fault or faults from a set of observed symptoms.
edge base is usually held in object-oriented models and a
sequential hypothesise-and-test inference process is used.
2.1 Associative networks This approach results in very powerful tools, because all
When associative network approach is used, the symp-
an
advantages of object-oriented programming can be used (clas-
are stored in the form of direct links
sification, hierarchies, (multiple) inheritance, message pass-
tom/fault relationships ing, polymorphism, etc.) and, in addition, declarative pro- .
between the symptoms and the faults (see Figure 9). The
links point to the faults which could cause the particular
gramming (i.e. rules) can be closely intermixed with proce-
dural programming (i.e. methods). As a result, the most
symptom. Given a set of symptoms, the reasoning process appropriate paradigm can be chosen for each of the diagnostic
finds all of the sets of faults which could explain the occur- subtasks (e.g. calculations by methods, diagnosis by rules). In
rence of the observed symptoms<’5>.
general, expert systems of this type are often explicit attempts
to model the underlying reasoning of the human diagnosti-
The diagnosis is the minimal set of faults which explains all
of the observed symptoms. The more symptoms point to a cian, also including abductive inference of the form:
given fault the more likely it is that this fault will be included given fact ’B’ and the association ‘IF , A > THEN < B >,
in the diagnosis~l6~. For details on how the term minimal may infer ’A’ is plausible
be defined more exactly, see Reference 15.
Although such inferences are quite obvious to human diag-
For example, suppose the fault detection program has detect- nosticians, rule-based expert systems are normally not able to
ed the two symptoms ’District Heat Outlet Flow Too Low’ draw such conclusions.
and ’Domestic Water Consumption Too Low’ to be present. An FDD system based on this approach could follow the
As can be verified by examination of Figure 9, there are three scheme shown in Figure 10 and use the simplified models of
faults capable of causing the detected symptoms, but only one a district heater substation shown in Figure 11.
fault (’Heat Exchanger Leak’) which is capable of causing
both symptoms. The resulting diagnosis is therefore that this In this scheme, the performance indices (PI) are the interface
fault has occurred<’ 7>. between the incoming data and the system model. Each PI
surveys a specific process variable at a specific location (e.g.

the heat exchanger outlet temperature at the secondary side).
2.2 Rule-based approaches In its simplest form, the PI compares the simulated and mea-
A typical diagnostic expert system which uses rule-based sured data by evaluating the difference between these two
deduction has a rule base consisting of conditional rules quantities. The simulation data may be generated by charac-
teristic curve modeIlingl’8>. If the resulting value is not within
(IF < premise(s) > THEN < conclusion(s) > ) } the range described by preset thresholds, fault diagnosis is
started. Through the use of a system model and a sensor
and an inference mechanism, based on modus poneru, of the description, the diagnosis scheme at this time is aware of the
form: type of the process variable (e.g. temperature) which is out of
bounds, its location (e.g. outlet of heat exchanger, Branch 2)
given fact ’A’ and the rule ’IF < A > then < B >’, infer ’B’ and which threshold has been exceeded (too high or too low).
In addition, it knows about the other sensors connected to the
Facts are expressed in the most simple form as attribute-value heat exchanger, and can therefore also make prognoses based
pairs. Applying this approach to the example shown in Figure upon the process variables controlled by these sensors. The
9, one such rule could look like: prognosis could be a pattern of symptoms like ’Outlet
IF < Domestic Water Consumption Too High is true > Temperature 22 is Too High’, ’Flow 1 is Too High’ and
’Pressure Drop at Control Valve is Too Low’. Using funda-
THEN < Leak In Domestic Water Pipe is true > mental relationships (e.g. equations based on the laws of
physics), these symptoms can be investigated further and
OR < Leak In Heat Exchanger For Domestic Water is true > considered in relation to each other. A rule expressing such a

Figure 9 Associative network for


diagnosis of faults in a district
.

heater substation

B22 . - 01 ................
fundamental relation as AP = R(91? might be expressed as
follows:
IF < Device Pressure Drop is Too Low >
THEN < Device Outlet Flow is Too Low >
OR < Device Resistance is Too Low >
In this example, the ’Device’ can be any of the objects
described in the system description. Finally, through apply-
ing such device-oriented rules together with pattern-oriented
rules, the diagnostic system should either identify one or sev-
eral faults as causing the symptoms or, in case of false alarms,
wait for the next time step and repeat the diagnosis.

3 Advantages and disadvantages


A general drawback of all the approaches which have been
described is that it is always difficult to ensure that the
knowledge is complete and correct, and to check the consis-
tency of the underlying knowledge bases. Other advantages
and disadvantages are summarised in Table 4.

12 Harmon P and King D Expertensysteme in der Praxis (Munchen/Wien:


Oldenbourg)(1986)
13 Buchanan B and Shortcliffe E Rule-based Expert Systems: the Mycin
Experiment of the Stanford Heuristic Programming Project (Reading,
MA: Addison-Wesley) (1984)
14 Kamimura K Rule-based Methods in IEA in Annex 25: Building
Optimisation and Fault Detection and Diagnosis Source Book
&mdash;

(Espoo, Finland: Technical Research Centre of Finland) (1995)


Figure 10 Combined fault detection and diagnosis system 15 Reggia J, Dana D and Wang P A Formal Model of Diagnostic
Inference Information Sci. 3 227-285 (1958)
16 Hinkelmann M and Schockle M Associative Networks in IEA Annex
25: Building Optimisation and Fault Detection and Diagnosis &mdash;

Source Book (Espoo, Finland: Technical Research Centre of Finland)


(1995)

Figure 111 Structure and models


used in the example diagnostic
expert system

B23
17 Peng Y and ReggiaJ Abductive Inference Models for Diagnosis Problem parameters exactly. He might observe, for instance, that an
Solving (New York: Springer) (1990) air-handling unit is actively cooling the air, even though the
18 Madjidi M Characteristic Curves in IEA Annex 25: Building Optimisation outside-air temperature is ’far low’
and Fault Detection and Diagnosis Source Book (Espoo, Finland:
&mdash;
too to expect such behav-
iour.
Technical Research Centre of Finland) (1995)

Broadly speaking, a qualitative description involves a coarse


Table 4 Comparison of expert systems based on associative networks, and reduction of the relevant information. Thus temperatures might
rule-based and knowledge-based approaches be replaced by temperature ranges, such as ’above 100°C’ or
’below 0°C’. Although such descriptions commonly arise in
rule-based approaches based on empirical expert knowledge of
the system, model-based descriptions are of special interest.
Qualitative models may be derived from first principles, the
approach used in qualitative physics, and the fault detection
and diagnosis (FDD) strategy relies on monitoring discrepan-
cies between observed qualitative values and corresponding
model-based predictions. Such strategies are generally robust,
being largely independent of the sizing parameters for a par-
ticular system or building. The trade-off, however, is that
they can be ambiguous in some situations the qualitative
-

description might be too coarse to reveal discrepancies even


when faults are present.

Another approach, which might be used in situations in


which formal qualitative models by themselves yield unac-
ceptably ambiguous results, involves methods based on fuzzy
sets~l9 ; see also Dexter and Glass(20). So-called membership
functions are associated with qualitative attributes such as
’high outside air temperature’. In general, unambiguous FDD
inferences will be valid only in certain qualitative operating
states<2’>. However, with the aid of defiizzifwation, such infer-
ences can be reduced to numerical values for arbitrary operat-
ing states, with the object of generating a measure of the dis-
crepancy between observed and predicted behaviour.

We present an example of the first approach. A fault detection


scheme using formal (’crisp’) qualitative models was applied
to the central air-handling unit (CAHU) of air-conditioning
Qualitative approaches to fault detection and
plants used in commercial buildings. The same method could
diagnosis be used for both a CAHU typical of North American HvAc
practice incorporating a heating coil, a cooling coil and a
-

A S Glass, P. Gruber and J Todtli (Landis & Gyr


bypass mixer and a plant representative of modem conti-
-

nental European practice incorporating a heat-recovery


-

Corporation, Zug, Switzerland) wheel (HRW). The qualitative approach used was first tested
using simulation software and then tested on a laboratory
CAHU incorporating bypass mixing. Finally the method was
1 Introduction
successfully demonstrated on real data from a commercial
building incorporating a heat-recovery wheel.
Many heating, ventilation and air-conditioning systems har-
bour undetected faults which can result in increased energy con-
sumption, increased emissions to the environment, or the occa- 2 Central air handling unit
sional failure of the system to meet its full perfornwr~ce specifi-
cations. Such faults can originate during the planrcirig, installa-
tion or commissioning as well as regular operatiort of the HVAC The type of plant investigated -
one incorporating a heat-

plant. In order to optimise both the ecorcorriy and the general recovery wheel -
is illustrated in Figure 12. Outside air is
serviceability of HVAC systems, it is highly desirable to detect heated or cooled to a preset supply-air temperature 7~ and dis-
such otherwise undetected faults automatically, wherever pos- tributed to ventilate the various zones in the building. Each
zone is separately controlled; the ventilation air can be cooled
sible, by intelligent rrurraitoring of the available building energy
or the zone can be heated by radiators as may be required.
management system data.
The task of the controller Cl is to ensure that the CAHU deliv-
There is a particular interest in qualitative methods. Most intel- ers supply air at the required temperature by operating the
ligent monitoring techniques rely on comparisons between heating coil, heat-recovery wheel and cooling coil in
observed system behaviour and behaviour predicted from sequence.
models of the system. Unfortunately reliable quantitative
models are often unavailable in practice, because many rele- Whenever the outside air temperature exceeds the return air
vant parameters of an HVAC system and building are very dif- temperature, it is advantageous to switch to an economy
ficult to measure and, in consequence, usually remain unde- mode in which the heat-recovery wheel is switched on and
termined. Nevertheless, an experienced service engineer can used to precool the outside air before it reaches the cooling
often correctly identify system faults without knowing these coil.

B24
Figure 12 Air-handling system
under investigation
- -------- -- -

3 Qualitative fault detection strategy where the linguistic mnemonics ’CLO’, ’NCLO’, ’OFF’, ’MAX’
and ’BMV’ stand for ’closed’, ’not closed’, ’off, ’maximum’
A wide variety of faults in the cAHU can be detected using stat- and ’between’. respectively.
ic qualitative models. The qualitative approach identifies oper- The measured values of the temperatures TOA’ TR, Ts (outside
ating regions of the controller separated by certain so-called air, return air and supply air respectively) are used to predict
‘landmark’ states, which are then related to corresponding
’landmarks’ of the relevant air temperatures (terminology corresponding qualitative values 1~~Q, UQ wand Uca
of the
controller outputs. The models used assume that the system
coined by Kuipers~zz~; see also Dexter and Glass(20». The over-
is operating under conditions approaching steady-state.
all scheme, illustrated in Figure 13, incorporates three
processes: Consequently, the analogue inputs are preprocessed to deter-
mine whether the system is indeed quasi-stationary. The
-
the quantisatian of controller outputs, reducing them to
qualitative values, design of the steady-state detector used in this work is
described by Glass et ~7.<~. Particularly important is the ques-
-
the prediction of qualitative controller outputs using mea- tion of tuning of the relevant steady-state detector parameters
sured temperature values, and - in this case time ’window’ and threshold which has
-

been described by Grube¡{24}.


-

testing whether the system is in a steady state.


Whenever the system is quasi-stationary and a mismatch
The analogue controller outputs Un, U~., Uc are reduced to between the observed and predicted qualitative controller
controller states ’heating’, ’heat recovery’ and ’cooling’ (plus outputs is noted, a fault will be detected.
transitions) by formally generating corresponding qualitative The prediction model used is illustrated graphically in Figure
values, and chosen from the following: 14. It is helpful to consider the temperature differences
- - - , ,

Figure 14 Graphical representation of prediction of qualitative controller

Figure 13 Qualitative model-based fault detection strategy outputs in terms ofsigns of To... - Ts and TR- T,

B25
and plot the temperature difference ARs against the difference
~05- The predicted qualitative controller output states occupy
a number of easily identifiable regions. The boundary
between cooling mode and HRW operation is the diagonal line
labelled U~. U~~m~~ in the lower right quadrant; the bound-
=

ary between heating mode and HRW operation is the same


diagonal in the left half of the figure. In practice, however, the
position of the diagonal line is difficult to determine without
accurate knowledge of detailed performance characteristics of
the heat-recovery wheel, so in a reduced prediction scheme
Figure 14 is separated by the signs of the three temperature
differences of equation 2.
It has been demonstrated that this qualitative predictor
model can be generated from a deep model in the form of a
PROLOG program (shown in Figure 15). It is structured as
follows: in the first part of the code the structure of the air-
handling unit is programmed - i.e. the way in which the
components are linked. The second part consists of the quali-
tative description of the components: heater, HRW, cooler and
controller. Finally, there are some auxiliary statements need-
ed for the logical clauses. Such deep models can be used in
future to generate qualitative prediction models for other
types of HvAC installations.

4 Tests on building data

For testing purposes, a medium-sized commercial building


occupied by the Landis & Gyr Corporation was selected in
Central Switzerland with the aim of:
-

determining approximately how often quasi-stationary


statesof the system could be expected to occur, and
-

applying the qualitative fault detection method to the


building data, but without knowing in advance whether
faults might be expected to be present.
A typical run is illustrated in Figure 16. The upper plot illus-
trates the outdoor, return and supply-air temperatures, for a
duration of 24 hours on 10 May 1995. The top three curves in
the second plot illustrate the controller outputs for the same
period, while the bottom curve is an alarm state. An alarm
was flagged if simultaneously:

-
the system was in normal operating mode (between the
hours of 05:30 and 19:30),
-
all three temperatures were quasi-stationary (tuning as
described by Gruber(2’;)), and Figure 15 PROLOG program for reduced fault detection scheme
-
a discrepancy between the qualitative temperature state sensor which yields measurements above the true value. In
and the qualitative controller state was observed. this case, independent evidence supports the latter.
In the event that such a discrepancy could possibly be
accounted for by quantitative differences between prediction 5 Conclusions
and measurement which are within the normal tolerances of
the system, the alarm is assigned the value 0.5; otherwise it The qualitative fault detection scheme described has proven
takes the value 1. its worth both in laboratory tests and on building data. In
particular, rw add£iixnal sensors are required, which adds to its
During the main operating period 05:30 to 19:30 the system economic viability.
was found to be quasi-stationary approximately 80% of the
time. As can be seen in the upper plot, the outside air temper-
The advantage of such methods is that they are comparatively
ature remains above the supply-air temperature from about independent of quantitative HVAC plant or building parame-
ters. The same method can be applied to plants of widely dif-
09:00 to 19:30. In spite of that the heat-recovery wheel con-
tinues to operate as late as about 15:00. As may be expected, fering sizes. Furthermore, the fault-detection procedures
the fault detection system results in a persistent alarm. applied to a central air-handling unit incorporating a heat-
recovery wheel on the one hand and one with bypass dampers
The observed symptom is compatible with both a cooling coil on the other were essentially the same. Thus the method may ’

valve that cannot switch off and an outside air temperature be considered generic.

B26
Figure 16 Building data test:
temperatures, controller outputs,
and qualitative fault detection
analysis on day when fault evident
The limitations of the procedure are typical for qualitative tently occupied buildings such as schools and offices. The
methods. Although faults may be present, qualitative discrep- research teams involved in the development of methods
most
ancies will not be observed in all operating states. Moreover, adapted to heating systems are those from VTT (Finland),
not all types of faults can be detected. Finally, once a fault has IKE (Germany), from Landis & Gyr (Switzerland), from the
been detected, the qualitative ’symptoms’ may often be insuf- University of Tsinghua (People’s Republic of China) and
ficient to diagnose the cause unambiguously. from CSTB (France). Their approach to this work made it
possible to develop methods for detection and diagnosis for
different purposes and using various techniques. Methods
applicable to buildings or to district heating schemes, and to
19 Zadeh L Fuzzy Sets Inform. Contr. 8 338-353 (1965) specific components in these systems, are presented here.
20 Dexter A L and Glass A S The use of qualitative models in fault detection
and diagnosis &sect;4.2 in Building Optimisation and Fault Diagnosis System 2 Methods for the system as a whole
Concept ed. Hyvarinen J and Kohonen R (Espoo, Finland: Technical
Research Centre of Finland (VTT) Laboratory of Heating and Landis & Gyr and CSTB have worked on methods designed
Ventilation) (1993) for the monitoring of energy consumption, and in particular
21 Dexter A L Fuzzy model-based fault diagnosis using fuzzy matching on energy signature methods. These methods should allow
Proc. IEE Part 142(6) 545-550 (1995) operating faults to be detected at the installation level and for
22 Kuipers B Commonsense reasoning about causality: Deriving behav- the building as a whole. They are based on analysis of the
iour from structure Artiftcial Intelligence 24(1-3) 169-203 (1984); 29(3) relation between energy consumption and outside tempera-
289-388(1986) ture. They make it possible to detect significant abnormal
23 Glass A S, Gruber P, Roos M and T&ouml;dtli J Qualitative Model-Based
Fault Detection in Air-Handling Units IEEE Control Systems Mag. 15(4) consumption. They were initially developed to analyse manu-
11-22 (August 1995)
al records or invoices regarding energy consumption. Energy
24 Gruber P Determination of the Tuning Parameters of the Steady State signature methods using data measured automatically and at
Detector for a Central Air Handling Unit (CAHU) I.E.A. Annex 25 Final reduced time intervals (hourly, daily) have been developed.
Technical Report AN25/CH/230994/3.0 These methods were also modified to take into account inter-
mittent occupation and the effect of diff’erent types of distur-
bances. Methods for consumption monitoring were also
developed, although they are still to be validated. The first in
situ tests, which were carried out in different schools in
Applying methods for detecting and diagnosing Montpellier (France), show that these methods have the
faults in heating systems advantage of being extremely simple, although they allow us
to detect only large-scale faults.
-

J-C Visier (CSTB, France) CSTB also developed a method for detecting and diagnosing
six types of faults associated with the boiler (poor combus-
1 Introduction tion, scaling or fouling-up), the control valves (leaks) and the
control/programming device (heating curve which is too low,
There are many different types of heating systems for which very early or very late restart procedure). This method using
one might consider implementing methods for detecting and neural networks was tested by generating computer simula-
diagnosing faults. Within the framework of IEA Annex 25, tion results representative of schools or offices. The addition
the work by the heating systems working group focused on of a flue gas temperature sensor to those already used for con-
district heating installations and central heating installations. trol makes it possible to diagnose these six faults with a prob-
As regards the latter, special attention was paid to intermit- ability of a successful diagnosis of 90% and a risk of false
B27
alarms estimated at 2%. Validation on site is now to be under-’
is indicative of a leak. This method was validated from labo-
taken. ratory measurements. VTT has also developed a method,
similar to that proposed by the University of Tsinghua, for
The University of Tsinghua developed a method designed to detecting the main faults in district heating substations
detect blockages in the networks of district heating schemes.
This method uses the pressure and flowrate measurements
(lower exchanger effciency, leakage from the primary to the
secondary loops, blockages in the primary or secondary loop,
already taken at each substation. The heat losses are comput- leakage from the primary or from the secondary loop to the
ed for each of the branches of the network with no control
devices from measurements taken on these branches. A
outside, shutdown of the secondary circulation pump). This
method requires measuring supply and return temperatures
hydraulic model of the network, in which its topology is and pressures in the primary and secondary loops. From
described, and an optimisation procedure for calculating the these eight measurements, a simple static method of prepro-
value of the head loss coefficients is used to do this. A varia-
tion in the head losses indicates a pipe blockage. The method,
cessing the data makes it possible to compute two synthetic
variables whose values should be constant under normal
which is costly in terms of calculation time, was applied to the
network of a district heating scheme in the town of Shenyang.
operating conditions. The diagnosis is carried out by
The method is used off line at regular time intervals to check
analysing the variations in both variables. Using only two
variables makes it possible to carry out the diagnosis using a
whether there is any abnormal change in the head loss coeffi- 2D plot of the variations in one variable plotted against the
cients. variations in other (see Figure 18). The method was validated
on the heating network at Shenyang in China.

3 Methods for specific components


Finally, IKE developed a method which can be applied more
generally. The fault detection is based on comparing mea-
VTT developed a series of methods designed to detect and sured process variables with reference data, where the refer-
diagnose faults in two-way control valves and their associated ence data are generated by system simulation. The simulation
hydraulic components. These methods requires the fluid flow model is generated using characteristic curves. The diagnosis
rate, the position of the valve and the differential pressure uses a knowledge base containing knowledge in object-orient-
across the valve to be measured, and are based on the use of a ed form about the symptoms, the faults, the system and its
simple static model of the system identified during a training topology. Whenever a difference between the measured and
period. The method uses indicators which are either the dif- reference data is detected, the diagnostic process attempts to
ference between the output of the model and the measured generate and tests hypotheses using IF... THEN... rules
variables, or the variations in the model’s parameters. These together with the models in the knowledge base. Currently
indicators can be inserted in a classifier which carries out the the system is being applied to district heating substations. It
diagnosis. These methods were applied to both the fuel oil is also planned to apply it to air handling units.
supply line of a burner and to a district heating substation. In
the first case, the validation was carried out on simulation
data and, in the second case, on data measured on a real 4 Results and future work
installation. The results obtained are encouraging since the
methods enable faults to be detected. However, false alarms A number of methods have been developed, though valida-
can frequently occur with such methods and the classifier
tion is still to be fully completed on site. The research has also
must be carefully designed if they are to be avoided.
allowed the different research teams to obtain an in-depth
understanding of a wide range of methods, and enabled them
VTT also developed a method for detecting leaks from dis- to select the method most suitable for a specific installation or
trict heating substation exchangers. This method requires plant component.
measuring primary and secondary flow rates as well as the
supply and return temperatures to the primary and secondary
loops. A dynamic model identified from initial measurements
is used to compute the energy consumed by the primary loop
on the basis of the measurement of the energy supplied to the
secondary loop. The consumption measured at the primary
loop is then compared with the consumption thus calculated.
The presence of a sustained deviation between the two values

Figure 17 Neural network struc-


ture

B28
Figure IS Diagnosis procedure
for district heating substation

Fault detection and diagnosis in air-condi- Certain abrupt taults can be detected by checking whether
tioning systems set-points are being maintained, and whether certain critical
variables remain within predetermined limits. In many cases
though, the operation of feedback control will conceal the
TI
effects of all but the most severe faults. To address these prob-
Salsbury, P Haves and J A. Wright (Loughborough lems, research has focused on model-based methods for
University, UK) analysing the sensor and control signals. The types of models
which have been applied to the task of detecting and diagnos-
1 Introduction
ing faults in AHUS include:
-

qualitative models
The purpose of air-conditioning systems is to control the -
neural networks
purity, temperature and humidity of the air in a space. In
most systems this is achieved using water-side (primary)
plant (i.e. boilers, chillers, water pumps, valves, etc) and air-
side (secondary) plant (i.e. fans, filters, dampers, heat
exchangers, humidifiers, etc). The secondary plant is general-
ly well instrumented and the sensor and control signals con-
tain potentially valuable information about the condition of
the different plant items. Modem building energy manage-
. ment systems (HEMS) facilitate the monitoring and storage of
these signals. The potential therefore exists to use the infor-
mation available to the BEMS for detecting and diagnosing
faults in the secondary plant.
An air-handling unit (AHU) consists of a number of different
components (e.g. heat exchangers, fans, dampers, filters, etc.)
which are susceptible to faults that can have a significant
effect on thermal comfort and energy consumption. Figure 19
depicts a typical AHU with heating and cooling coils, a mixing
box and supply and extract fans. Figure 19 Example of an air-handling unit

B29
-
state space models development process in order to assess the effects of unmod-
elled phenomena such as noise and unmeasured disturbances.
-

fuzzy models
Certain faults can be implemented in real systems without
-

’physical’ models causing lasting plant damage (e.g. a stuck valve or damper can
-
characteristic curves be implemented by disconnecting the actuator linkage). The
other faults can be implemented by simulating their effect.
For example the effects of a leaking valve on a heat exchanger
2 Model-based fault detection and diagnosis
may be simulated by altering the controller software so that a
zero control signal cannot be attained.
The use of mathematical models in fault detection and diag-
nosis systems has become popular in recent years<25-27>. Fault detection and diagnosis procedures developed at
Mathematical models are used to detect changes in the corre- Loughborough University have been tested in three phases
lation between the measured inputs and outputs . Two types using:
of mathematical model can be defined: analytical and empiri- -
HVACSIM+ simulations of office buildings
cal. Analytical models (sometimes called ’physical’ models)
are derived by considering the mechanisms which generate -
a full-size test rig at the Building Research Establishment
the signals and are based on physical laws that govern the (BRE)
behaviour of the system. Empirical models are inferred from
measured data obtained from the system of interest. A num- -
a real building on the Massachusetts Institute of
ber of candidate models are selected based on ’shallow’ Technology (MIT) campus.
knowledge of the functional behaviour (e.g. high order/low The tests at the BRE culminated in a live demonstration of
order, linear/nonlinear, etc.) and the model that is able to fit the fault detection and diagnosis technology in conjunction
the data most accurately is chosen. In practice, many analyti-
with Oxford University. Two different faults were introduced
cal models also include empirical relationships and would be
described more correctly as semi-empirical. during the demonstration: a leaking mixing box damper and
a supply air temperature sensor error. Both of the faults were
There are two main approaches to fault detection and diagno-

detected and correctly diagnosed by the fault detection and
sis using mathematical models: diagnosis systems. Tests carried out in a building on the MIT
campus involved gathering data from the BEMS under normal
-
Innovation generation
operating conditions with the system in fault-free and differ-
-
Parameter estimation. ent faulty states. The method tested at MIT used models of
heat exchangers as part of a parameter estimation procedure.
The innovation approach uses a model of the fault-free sys- Figures 20 and 21 show the results of testing the method with
tem to predict the outputs from the measured inputs. The dif- data collected from the system with a fouled cooling coil.
ference between the predicted and the measured outputs are Figure 20 shows the measured data and Figure 21 shows
known as ’innovations’. A fault is deemed to have occurred if three estimated parameters: the coil conductance UA, the
the innovations exceed a predetermined threshold. A diagno- valve characteristic parameter c, and the percentage leakage
sis can be produced by analysing the way in which the inno- through the coil valve (%LEAK). It can be observed that the
vations vary with operating point{28). estimated conductance falls (from its ’correct’ value) as the
model adapts to represent the observed behaviour of the
The parameter estimation approach involves estimating the
fouled coil. Static models were used in this example and the
values of the parameters of the model from measurements of
breaks in the lines in Figure 21 correspond to the periods
the system inputs and outputs. The parameters estimated are
when the measurements were rejected because of their vari-
assumed to be relatively constant when the system is operat-
ing correctly, but to change significantly in the presence of ability.
faults. A diagnostic schedule is produced by analysing the
way in which the estimated parameters vary. If analytical 4 Conclusions
models are used, the diagnosis task is simplified since such
models have physically meaningful parameters. The type and
the magnitude of the fault can therefore be ascertained direct-
A variety of model-based approaches to fault detection and
diagnosis in air-handling equipment have been developed.
ly from an analysis of the parameter variations. In both Methods based on parameter estimation and on innovations
approaches, either dynamic or static models can be used. have been tested and have demonstrated significant potential.
However, in the case of static models, additional on-line pre- Innovation-based approaches are particularly suited to fault
processing is required to ensure that transient data are not detection but do not allow diagnostics to be generated easily.
used.
Conversely, parameter estimation approaches can provide
useful diagnostic information directly, but can suffer robust-
ness problems due to the compounding of estimation and
3 Development and testing of FDD procedures
modelling uncertainties. Noise effects, unmeasured distur-
bances and modelling errors can cause problems for both
Ideally, fault detection and diagnosis procedures should be
developed and tested using real systems. However, in prac- approaches, and further development and testing in real
tice, it is difficult to implement real faults in a system, since buildings are required to establish which approach will be of
many of the faults of interest cause lasting damage. Computer
most benefit to the industry.
simulation is therefore a valuable tool to assist in the develop-
ment process. Simulation programs such as HVACSIM+~z9~
can be used to construct detailed simulations of air-condition-

ing installations in fault-free and various faulty states. It is 25 Gertler J Survey of Model-based Failure Detection and Isolation in
important, however, to use real systems at some stage in the Complex Plant IEEE Control Systems Mag. 8(6) 3 (1988)

B30
Figure 20 Data measured at MIT
using a cooling coil which had
been modified to simulate fouling

Figum 21 Parameter values esti-


mated from the measurements
shown in Figure 20

26 Isermann R Model-based Fault Detection and Diagnosis Methods 29 Park C, Clarke D R and Kelly G E An overview of HVACSIM+, a
Proc. American Control Conf., Seattle, Washington, USA p1605 (1995) dynamic building/ HVAC /control systems simulation program Proc. 1st.
27 Benouaret M., Dexter A L, Fargus R L, Haves P, Salsbury T I and Ann. Building Energy Simulation Conf, Seattle, Washington, USA (1985)
WrightJ A Model-based Approaches to Fault Detection and
Diagnosis in Air Conditioning Systems Proc. System Simulation in 30 Haves P, Salsbury T I and Wright,J A Condition Monitoring in HVAC
Building ’94 p52 (1994) ASHRAE Trans. Pt 1102 (1996)
Subsystems
28 Salsbury T I, Haves P and WrightJ A A Fault Detection and
Diagnosis Method Based on First Principles Models and Expert Rules 31 HyvarinenJ Building Optimisation and Fault Diagnosis Source Book
Proc. Tsinghua HVAC-95, Beijing, PRC (1995) (Technical Research Centre of Finland) (1996)

B31
Fault detection in chillers (a) System models: The entire chiller is considered as a ‘black
box’ with three independent inputs determining the out-
puts.
V E Bakker and H C Peitsman (TNO Building and (b) Component mosdels: Each major component of the chiller
Construction Research, Delft, The Netherlands) is considered as a ’black box’.

1 System identification and ARX models 2.1 System models


Faults can be detected by comparing the measured value of Independent variables are necessary to generate a good sys-
process variables with the value predicted by a mathematical tem model. The condenser and evaporator supply tempera-
model of the chiller<32}. A ’black-box’ model can be used to tures are both functions of external factors: the outside air
describe the relationship between the inputs and outputs. temperature and the internal load of a building. The instanta-
Such models are so described because they require little phys- neous power of the compressor can be used as a measure of
ical knowledge of the process. The parameters of a model are the internal load of the chiller.
estimated using a system identification scheme~33> and
input/output data collected from the chiller when it is operat-
ing correctly. 2.2 Component models
In this study on a laboratory chiller, one of the two types of
model used was a multi-input single-output auto regressive with
The configuration of the component models is based on a
exogenous inputs (ARX) model(3-t). Thermal capacitances and
study of the structure of refrigerant systems. The use of this
delays in the chiller result in process time dependencies. The knowledge makes the model a little bit ’grey’ (i.e. uses some
structure of the ARX model takes account of the probable exis- physical knowledge) instead of black (i.e. uses no physical
tence of such time dependencies by offering the possibility of knowledge). Every model has its own inputs and outputs
based on the physical inputs and outputs of the chosen com-
using the previous values of the inputs and output in the cal-
culation. Also, since real processes seldom have a direct linear ponents in the chiller system. The configuration of the refrig-
erant components of the chiller is shown in Figure 23.
relationship between the inputs and outputs of the process,
additional inputs, obtained by squaring or taking the product Figures 24 and 25 show typical variations in the discharge
of the original inputs, are sometimes included to improve the pressure P2 during correct and faulty (air is introduced into
structure of the model. the system) operation of the chiller. Figure 26 shows typical
variations in the residuals (the difference between the predict-
ed and measured discharge pressure) when the chiller is oper-
ating correctly. The threshold for detection is determined by
2 Chiller under test using statistical methods(&dquo;).
The refrigerant system used to assess the FDD scheme is a rec-
iprocating chiller. The laboratory chiller consists of a com-
pressor, a condenser, an evaporator, an expansion valve and a
subcooler. The filter-drier is not considered in the model.
3 Conclusions
The process instrun~entation diagram (PID) of the chiller plant is
presented in Figure 22. The example presented here illustrates the possibility of
Two classes of models are created from the measured training using ’black-box’ models for fault detection. The use of sys-
data: tem models makes it possible to detect faulty behaviour of a

Figure 22 Process instrument-


tion diagram (PID) of laboratory
reciprocating chiller
B32
Figure 23 Configuration of refrig-
erant components of chiller

Figure 26 Typical variations of residuals during correct chiller operation


Figure 24 Typical variation of discharge pressure P2 during correct chiller
operation system, whereafter component models can be used to locate
the defective component.
The goal of using system identification based on black-box
models is to allow faults to be detected in HVAC systems.
Further development of the black-box modelling techniques
is necessary to create reliable models which can be used in
practical applications. The results presented here demon-
strate the potential of the approach, although ongoing refine-
ment of the models should lead to further improvements.

The use of ’black-box’ models in model-based reasoning sys-


tems, which are capable of precise localisation of defective
components, has also been considered<35>. The combination of
’black-box’ fault detection with such reasoning systems
should increase the reliability of the FDD scheme.

32 Galen van P J M, Peitsman H C and van Duyvenvoorde A M Fault


Figure 25 Typical variation of discharge pressure 1’2 during faulty chiller detection to a real chiller plant in an office building using black model identifica-
operation tion IEA Annex 25 report (TNO Building and Construction Research,

B33
Department of Indoor Environment and Systems, The Netherlands) IEA Annex 25 report (TNO Building & Construction Research,
(1993) Department of Indoor Environment and Systems, The Netherlands)
33 Ljung L System Identification Theory for the User (New York:
&mdash; (1994)
Prentice Hall) (1987) 35 Bakker V E A Real-Time Fault Detection and Diagnosis Application for
34 Peitsman H C, van DuyvenvoordeA M and. B&ouml;ttger W OJ Application Thesis, University of Twente, Enschede, The Netherlands
HVAC systems
of
m
- odelsARX and Neural networks to a reciprocating chiller for fault detection (1995)

B34

You might also like