You are on page 1of 12

Automation in Construction 75 (2017) 79–90

Contents lists available at ScienceDirect

Automation in Construction

journal homepage: www.elsevier.com/locate/autcon

A System Identification Methodology to monitor construction activities


using structural responses
Ranjith K. Soman, Benny Raphael ⁎, Koshy Varghese
BTCM Division, Department of Civil Engineering, IIT Madras, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a methodology to use structural responses to monitor the progress of construction processes.
Received 23 May 2016 The methodology was implemented on a launching girder used for viaduct construction of a metro rail project. A
Received in revised form 6 December 2016 strain-based wireless sensor network was used for data acquisition. Structural responses from the launching
Accepted 12 December 2016
girder were used to identify the state of construction. A conventional System Identification Methodology was
Available online xxxx
tested for this application but was not accurate in inferring the stage of construction. Therefore a modified system
Keywords:
identification strategy using derived features and heuristics was used to infer the state of construction. The mod-
System identification ified methodology was found to be significantly more accurate than the conventional methodology and is well
Model falsification suited for applications in unstructured construction environments. Results from the case study confirm that
Construction progress monitoring the use of structural responses is feasible for measuring the progress of construction activities.
Heuristics © 2016 Elsevier B.V. All rights reserved.
Wireless sensor network

1. Introduction timely fashion which leads to delays in the implementation of control


measures. Hence, accurate monitoring is required for proper execution
The global construction industry contributes to around 11% of the of a construction project within the planned cost and time.
global GDP in 2011 and it is predicted to cross 13.2% by the year 2020 One solution to the above challenge is collecting the data automati-
[1]. The share of the construction industry in the global economy in 2015 cally. Automated data collection has already been employed in con-
is US$8.5 trillion which is expected to climb to US$10.3 trillion by 2020. struction for productivity and progress monitoring. This is achieved
A significant portion of this development comes from the construction using technologies such as imaging [7–10], RFID [11,12], distributed
projects of emerging economies which show a higher rate of growth [2]. sensor networks [13], LIDAR [14] for progress monitoring, material
Increasing population, growing economy and urbanization in these na- tracking, safety monitoring, etc. These technologies have inherent limi-
tions necessitate investments in various infrastructure projects. The overall tations such as lack of mobility. For instance, an imaging technique for
success of these projects depends heavily on completion of construction progress monitoring is constrained by its location and cannot be used
within planned time and cost. To ensure this, the productivity and progress for monitoring projects such as viaduct construction. Joshua and
of the project have to be monitored and controlled on a continuous basis. Varghese have addressed this issue to an extent by integrating a wear-
Although, theoretical frameworks for project monitoring and control are able sensor on workers to track productivity [15]. Although this ap-
available with which precise project performance indices can be computed proach is suitable for productivity assessment, this might not be
[3], massive amounts of accurate data need to be continuously generated suitable to track progress, especially in large projects since it is challeng-
for such computations. While manual monitoring is conventionally ing to consolidate a large amount of data from all the workers and final-
utilized, it is slow and inaccurate [4]. Further, there are overheads for ly evaluate progress from these data. These limitations could be tackled
manual monitoring, it is reported that 2% of the work in construction is if the monitoring system is employed on equipment or supporting
committed to tracking and recording progress data manually [5]. structure. Therefore, as the construction progresses, the monitoring sys-
Construction supervisors spend about 30–50% of their time tem travels with the equipment/supporting structure without the re-
interpreting and analyzing field data. In many situations, data for correct quirement of any other mobility arrangement.
analysis is not available, and the supervisors resort to optimistic A monitoring system that makes use of structural responses is one
reporting [6]. As a result of these inefficiencies in data collection and in- such technology which can be employed in an equipment or temporary
terpretation, appropriate corrective actions are not implemented in a structure. However, construction progress monitoring using the re-
sponses from structural responses of temporary structures and con-
struction equipment is an area that has not been explored yet.
⁎ Corresponding author. Employing sensors which could detect the structural state of equip-
E-mail address: benny@iitm.ac.in (B. Raphael). ment or a temporary structure and in turn mapping it to the extent of

http://dx.doi.org/10.1016/j.autcon.2016.12.006
0926-5805/© 2016 Elsevier B.V. All rights reserved.
80 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

construction would enable collection of data which could later be used calibrated [21]. Hence, these parameters can only be used for interpola-
for analysis of productivity or progress. This study explores the feasibil- tion and not extrapolation. In addition, it may provide wrong values of
ity of using structural responses from construction equipment to assess system parameters which compensate modelling and measurement er-
the productivity and progress of construction as shown in Fig. 1. rors at sensor locations. Also, it is not suitable to use this method where
Although the objective of the study is to monitor the progress and pro- model simplification or systematic bias is present [20,22–25]. These limi-
ductivity of the viaduct construction, this paper focuses on monitoring the tations of the model calibration method were demonstrated analytically
progress alone by employing a strain-based sensor network on a launching by Ben-Haim and Hemez [26]. They showed that increasing the depen-
girder, an equipment used in viaduct construction. Structural responses dency of the model on the test data would reduce the robustness of un-
during the different operations of viaduct construction are collected derstanding the process.
using a wireless sensor network based data acquisition setup. The data Another widely adopted method is the Bayesian approach. Bayesian
(structural responses) are analyzed to identify the state of viaduct con- statistical framework is a rational and powerful tool for model updating
struction using a model based system identification technique. System that is capable of managing measurement errors, incomplete experimen-
identification is the process of interpreting the physical characteristics of tal data, non-unique solutions and modelling errors [27,28]. Bayesian in-
the system from indirect observations like stresses and strains. Although ference uses Bayesian conditional probability to update the prior
the process may seem simple, there are several uncertainties associated knowledge of model parameters from measurements and other evi-
with the structural configuration of the equipment in a construction site, dences using a likelihood function [29]. Likelihood function is a probabil-
which has to be addressed in the framework. ity distribution which represents the likelihood of observing data when
The paper is organized into nine sections. Section 2 evaluates the the model is parameterized by a factor. It has the advantage that it can as-
feasibility of applying common system identification technologies to sign relative weights to the model parameters contributing to the likeli-
construction. Section 3 discusses the operations of launching girder in hood function. Initially, prior distribution is used for characterizing
detail. Section 4 gives a brief outline of the problem statement, the initial knowledge/behavior for calculating the relative weights. Subse-
scope of the work and the methodology followed. Section 5 discusses quently, the prior distributions can be updated by using experimental
in detail the system identification algorithm used in the study and data to improve classification [28]. This method has been used extensively
Section 7 deals with the implementation of the current system on an ac- in different applications since 1960 [30]. Its first application in structural
tive launching girder and Section 8 presents the results from this imple- identification dates back to the 1990's [31]. Since then, Bayesian approach
mentation. Section 9 presents the conclusion of the study. has been used to solve several complex model updating problems with
great efficiency. Bayesian methods have been used to solve higher dimen-
2. System identification concept and practices sion model updating problems using Monte Carlo simulations and its ef-
fectiveness was tested on structural dynamic problems [32]. Bayesian
System identification is the process of interpreting the physical char- uncertainty quantification has been applied to calibrate complex mathe-
acteristics of the system from the observations. Model calibration is one matical models related to turbulence modelling to predict quantities of in-
of the system identification approaches used to find the right model by terest [33]. Bayesian statistical framework was used to update models in
minimizing the difference between the observations and predictions. order to improve the match between experimental and corresponding
This method is still being widely used. One of the recent applications of model output in structural identification [34]. If the relation between
this method involves the prediction of the lifetime reliability of a bridge model parameter and the available data is complex, it is difficult to formu-
by calibrating the Young's modulus [16]. Similarly, field measurements late the likelihood function. This issue has been addressed using two stage
were used to update the baseline model of a multi-girder bridge in its con- approaches where the second stage is usually performed using heuristics
struction phase [17]. The same approach has also been used to update and [35,36]. Recent studies have avoided the use of heuristics by developing a
calibrate FEM models of masonry bridges from ambient vibration data fundamental theory for Bayesian two stage problem. An equation is de-
[18]. A weakness of this method is that a single answer to an inverse prob- rived to express the posterior distribution of model parameters identified
lem is rarely found; it is possible to get same responses at the sensor loca- [37]. This framework has been tested on ambient vibration data and
tions from many different combinations of parameter values [19,20]. Also, found that the resulting posterior distribution to be in strict concordance
the values of the parameter calibrated by model calibration may not have with Bayesian logic [38]. The System Identification Methodology based on
a physical interpretation outside the model structure within which it was Bayesian approach has been extensively used in laboratory experiments,

Fig. 1. Research question.


R.K. Soman et al. / Automation in Construction 75 (2017) 79–90 81

stochastic damage accumulation studies [39], model updating with un- consists primarily of five parts. Steel plate girder, front support, middle
certainty [40], structural damage detection [41], updating models using support, rear support and rear trolley as shown in Fig. 2. The different
incomplete model data [42], etc. Even though Bayesian updating is exten- stages (or state of launching girder) for the construction of a span and
sively used, a drawback of this methodology is that it requires the proba- its associated bending moment diagram (BMD is for illustration purpose
bility distributions, which represent modelling and measurement and may not be accurate on schematic) are as shown in Fig. 3.
uncertainties at the sensor locations. It is challenging to generate such
probability distributions for applications with complex uncertainties 3.1. Auto launching
such as in construction. Also, simplification of probability distribution
will lead to over conditioning of parameter values [19]. Therefore, Bayes- Auto-launching is the process when the launching girder is moved
ian inference requires complete knowledge of all errors correlations be- from the ‘constructed span’ to the ‘next to be constructed’ span of the via-
tween measurement locations in order to avoid bias [22,43]. duct as shown in Fig. 3:1. In this process, the front support of the
In 1930, Popper asserted that in science, models can only be falsified launching girder is released and launching girder is supported by remain-
and not validated by data [44]. This concept has then been used in dif- ing supports. Launching girder is launched to the subsequent span by
ferent studies. One of the pioneering researchers in this area, Albert pushing the plate girder forward to the next pier. Counterweight placed
Tarantola has stated “use all available a priori information to create se- at the rear side of the LG balances cantilever moment created by the
quentially models of the system, potentially an infinite number of them. front portion until the front support reaches the next pier. The support
For each model, solve the forward modelling problem, compare the predic- is then rested on the pier which marks the end of Autolaunching opera-
tions to the actual observations and use some criterion to decide if the fit is tion. The BMD of the operation is shown in Fig. 3:1 and characterized by
acceptable or unacceptable, given the uncertainties in the observations and, the high negative moment at the cantilever support.
perhaps, in the physical theory being used. The unacceptable models have
been falsified, and must be dropped. The collection of all the models that 3.2. Segment lifting
have not been falsified represents the solution” [45]. Discrete popula-
tion-based multiple model approach helps in solving inverse problems This operation takes place after the auto-launching. It involves lifting
that have non-unique solutions. each prefabricated segment from the ground sequentially and assembling
In this approach, multiple models are generated by assigning differ- it along the span to form the viaduct as shown in Fig. 3:2. Once all the seg-
ent combinations of values of variables that are not known precisely. ments are lifted, they are then aligned and moved into position to be
Models whose predictions match observations within estimated thresh- joined and placed on the piers. All the segments lifted remain suspended
olds are selected as candidates. Raphael and Smith [46] first proposed on the front portion of LG. Typical BMD of the operation is shown in Fig.
the use of discrete population-based multiple model approach to bridge 3:2. There is a positive moment on the front side of the LG while having
diagnosis. Later, the concept of error domain model falsification was de- a negative moment on the rear side. Please note that the BMD doesn't cor-
veloped by Goulet et al. [43,47,48] in which the idea that models can respond accurately to the schematic of the operation shown.
only be falsified was formally applied. It was shown that candidate
models can be identified from the measurement data better than 3.3. Post-tensioning
using model calibration and Bayesian inference [43]. Similar ideas can
be found in other domains such as environmental monitoring [49]. This process, which takes place after the segment lifting involves
Discrete population-based system identification has been used for joining all the segments through a post-tensioning process as shown
bridge diagnosis, leakage detection, environmental monitoring, etc. in Fig. 3:3. Cables are run through the segments and then tensioned to
[19,22,43]. The use of discrete population-based system-identification the designed stress with a hydraulic jack while the segments are
has not been explored for productivity measurement or progress mon- suspended on the front portion of LG. Once the stressing is done, the as-
itoring in construction. Although the principle of system identification sembled segments together act as a single structure. Load of the seg-
can be similar to previous works, the process for the same will have to ments is still borne by the front portion of the LG. State of the system
be made specific for construction progress monitoring. For example, during this process is depicted in Fig. 3:3. The bending moment is sim-
the process of modelling population for the system identification for ilar to segment lifting. However, the moment on the front side is partly
construction progress monitoring is different from that used for bridge negative in this case since the launching girder is fully loaded.
diagnosis. It is also essential to know what indicators should be identi-
fied to evaluate the progress. This paper evaluates the feasibility of ap- 3.4. Span loading
plying this approach to monitor the launching girder operations. The
conventional approach was modified in this work to improve the accu- The post-tensioned span is lowered onto the piers in this process.
racy of system identification by incorporating heuristics and derived Once the span rests on the pier, Segments are released from the
features, instead of using only direct measured data. launching girder. The span transfers its self-weight to the piers. LG is
free from the load of the segments. Once this process is done, LG is
3. Launching girder operations ready for auto launch to the next to be constructed span. Schematic of
this stage is shown in Fig. 3:4. The bending moment during this stage
This section describes the operations of a basic box type launching gird- is similar to that during the beginning of segment lifting but with a
er. It is observed that the overall progress of metro rail viaduct construction higher magnitude of the negative moment on the rear side.
can be ascertained by monitoring the operations of the launching girder. The bending moment changes continuously with respect to the op-
Launching girder (LG) is an equipment used in metro rail construction to eration. It starts with a BMD similar to span lowering as shown in Fig.
lift and assemble prefabricated concrete segments, which are then joined 3:4. When the support is released during launching, BMD on the front
to create the span of the viaduct. The LG is fabricated as a plate girder span will have a negative moment which increases with the increase
(box) and spans continuously over four supports as shown in Fig. 2. in the distance launched. Finally, when front support rests on the girder,
Launching girder is involved in several operations including the lifting BMD corresponds to the one at the start of launching. Then, during the
of segments, pre-stressing, post-tensioning and loading of the span. start of launching front side of the launching girder will have a positive
Therefore, monitoring the state of this equipment should give direct infor- moment which gradually shifts to negative moment as segment lifting
mation related to the progress of metro rail construction. Completion of progress. Also, post-tensioning and the end of segment lifting would
each cycle of the launching girder operations can be considered as a mile- have the maximum negative moment on the front span of the LG. How-
stone activity in the schedule of metro rail viaduct construction. The LG ever, these can be differentiated since the end of segment lifting will
82 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

Fig. 2. Different parts of launching girder.

have a sudden change in slopes of BMD while that of Post-tensioning 3, it can be understood that launching girder operations has inherent
would have gradual changes. Once the span is lowered onto piers, the uncertainties, and the System Identification Methodology should
BMD changes to the initial state before the launch. address these while inferring the state of operations. Therefore, it is essen-
These changes in stress patterns can be identified in real time by mea- tial that the model falsification methodology used for the study should
suring the structural parameters using sensors such as strain gauges, load enable provisions to include these uncertainties yet predict the state of
cells, etc. Once the strain pattern is obtained, the state of the LG can be rec- construction accurately for monitoring the progress and productivity.
ognized based on the strain pattern across the length of the LG. However, This study aims at evaluating the feasibility of using discrete popula-
the operations of LG has its uncertainties such as support settlement, var- tion-based system identification methodologies, or in other words falsi-
iations in counter weights, variations in the weight of segments, etc. Due fication methodologies, to effectively monitor the state of operation of
to these field uncertainties, the results from the base analytical model will the launching girder. Ideally, the System Identification Methodology
not be the same as the field measurements. For example, a segment should be powerful enough to identify the sub-activities (refer Table
weight is generally around 20 tons, but for some spans, the segments 1) from the sensor data. However, the scope of this paper is limited to
might weigh close to 16 tons only. In such cases, the response diagram identifying the four major operations of the launching girder as men-
might correspond to a different state of construction and reduces the ac- tioned in Section 3. The proposed system identification algorithm
curacy of identification. In addition, uncertainties in measurements and matches the observed readings to one of the simulated cases which cor-
errors in sensing may arise due to the nature of field conditions such as respond to a sub-activity (as shown in Table 1). However, the algorithm
a change in temperature, variations in structural parameter values, etc. maps the observation to the major operation under which the matched
The system identification algorithm should cater to these uncertainties sub-operation is identified. For example, the major operation
while identifying the state of the launching girder. ‘Autolaunching’ contains 16 sub-activities under it (Major operations
and their sub-activities are shown in Table 1). Although the proposed
4. Problem statement and evaluation methodology System Identification Methodology maps the observed reading to one
of the 16 sub activities, the output is shown as ‘Autolaunching’. In
Ideally, System Identification Methodology as mentioned in Section 2 other words, the sub-activities are not reported as output in the pro-
should cater to the monitoring of launching girder. However, from Section posed System Identification Methodology.

Fig. 3. Launching girder operations (schematic) and associated bending moments: 1. Auto launching; 2. Segment lifting; 3. Post-tensioning; 4. Span loading.
R.K. Soman et al. / Automation in Construction 75 (2017) 79–90 83

To monitor the state of the launching girder, a custom designed sen- thresholds of possible values of actual responses are defined by the fol-
sor network is deployed on the equipment. Responses from the lowing equations:
launching girder are analyzed using the above-said system identifica-
tion methodologies to arrive at the state of operation. In this study, im- Upper threshold bound ¼ Predicted Value þ errorðδÞ
portant launching girder stages include auto launching, segment lifting,
post-tensioning, and span lowering as pointed out in Section 5. Lower threshold bound ¼ Predicted Value−errorðδÞ
The evaluation methodology followed in this work is as follows.
A model instance is falsified if the difference between its predicted re-
• A strain-based data acquisition setup is deployed on an actual working sponse and the measured value is outside the threshold bounds at any
launching girder. sensor location. This system identification strategy based on model falsifi-
• The construction operations at the site are physically observed during cation has been proved to be efficient in various experiments [22,27].
a period of 2 months and the data from the acquisition setup is record-
ed during this period. Site activities are manually recorded in a log 5.2. System Identification Methodology B
book for comparison.
• System identification is performed using the data recorded in the ac- System Identification Methodology A with its strict limits on errors
quisition system using three methodologies. may not be suitable for an unstructured environment like construction.
• The accuracy of system identification is calculated by comparing the Uncertainties are inherent in construction processes which result in er-
inference made with the recorded observations in the log book for rors that cannot be modelled/simulated prior to the operations Strict
each time step. limits on error bounds will influence the rejection of candidate models
and increase chances of Type 2 error. Type 2 errors are the errors
5. System identification methodologies which fail to represent an effect that is actually present. For example,
if auto-launching is taking place and the algorithm fails to report it,
Three approaches to system identification are used and compared in the error can be termed Type 2. Even when measurement errors are
this study, which is as follows. present, the general trends in the structural response do not change.
Hence, system identification based on identifying the trends in structur-
• The System Identification Methodology adopted from Goulet et al. al responses is likely to be more reliable.
[50], termed as ‘System Identification Methodology A’ uses raw mea- The current work uses slopes between measurements at adjacent loca-
surement data. Modifications are made to the System Identification tions in addition to the measured readings and predictions. Even if the
Methodology A in the presumption to increase the accuracy of strain values have errors, trends such as increase or decrease in their values
inference. are likely to match the predictions of models. Therefore, in this methodol-
• A modified methodology termed ‘System Identification Methodology ogy, a model is falsified if the predicted slope between the readings at ad-
B’, uses derived features for identifying candidates. jacent locations is outside the threshold of the measured slope.
• A further modified methodology termed as ‘System Identification However, in this case, the threshold is computed differently. Consid-
Methodology C’, is an improvement of the System Identification er Fig. 5. Let y be the true value at a location A and δ be the combined
Methodology B by using domain-specific heuristics. error. Then,

Inferences made using system identification methodologies A, B and Upper threshold A0 ¼ y þ δ
C are evaluated by comparing it with the records in the log book and  
their accuracies are compared. Lower threshold A″ ¼ y−δ

5.1. System Identification Methodology A Similarly at B,


 
The System Identification Methodology used in the present study is Upper threshold B″ ¼ y þ h þ δ
built upon the framework developed by Goulet et al. [25]. In the context
of construction progress monitoring, the objective of system identifica- 
Lower threshold B0 ¼ y þ h−δ
tion is to identify the state of operation of an equipment which could be
linked to the progress of the entire construction. A typical falsification
process is shown in Fig. 4. The system identification process starts Let the distance between A and B be of unit value. Hence, the slope of
with defining a set of model classes representing different possible line segment AB is (h). Similarly, the slope of the line A′B′ is (h − 2 δ)
states of the system. Each model class contains a number of parameters and the slope of line A″B″ is (h + 2δ). Therefore, the threshold bounds
which correspond to the system properties such as geometry, material of slope in this framework is as follows,
characteristics, loading, boundary conditions, support conditions, etc.
Upper threshold bound ¼ Predicted Slope Value þ 2  errorðδÞ
Each combination of model parameters leads to a model instance
which is analyzed using finite element method in order to get a set of
Lower threshold bound ¼ Predicted Slope Value−2  errorðδÞ
predictions (which can be measured), for example, strain values at
each location.
The steps involved in System Identification Methodology B are as
The model prediction differs from the correct structural response at
follows:
a location by an amount equal to the modelling error. The measured
value also differs from the actual response by the value of the measure- i. Import the data from the server
ment error. However, neither correct values nor the error values are ii. Determine the slopes at the critical locations and compare it with
known precisely [26]. Therefore, a probability density function describ- slopes from simulated cases
ing possible errors needs to be estimated. In the absence of more de- iii. If a set of matching slopes are found, proceed to step (iv), else report
tailed information, a uniform probability density function might be an anomaly
used in which the minimum and maximum possible values for errors iv. Try the first set of matching slopes
are defined. The probability density function will have the errors from v. Predict the intermediate readings by following the pattern from
the measurement as well as the errors from the modelling. The matching simulated case
84 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

Measure responses
from actual case

Falsify
Generate Simulate each simulated
population case to get cases which Report cases
of cases by measurable lie outside which are not
varying responses for error falsified
parameters each case threshold of
the measured
response

Fig. 4. Falsification methodology.

vi. Compare the properties/patterns of the intermediate points with the vii. Check whether the matching property is either the previous infer-
simulated case, if it matches, report the state corresponding to that ence or the state following the previous state. If yes, report the in-
case. Else try the next set of matching slopes from step (iii) and re- ference. Else proceed to step (viii). This is shown in Blocks 9, 10,
peat steps (v) and (vi). If no more matching slopes are available, re- 11 of Fig. 6. For example, Segment lifting happens after the
port anomaly. Autolaunching. During Autolaunching, ‘previous inference’ will be
set to ‘Autolaunching’ because the inference was reported as
5.3. System Identification Methodology C ‘Autolaunching’ in the previous observation/reading. In step (vii),
the inference from step (vi) is checked whether it is ‘Autolaunching’
System Identification Methodology B is further modified by using do- or ‘segment lifting’. If it is either of those, that state is reported.
main-specific heuristics. The inference from System Identification Meth- viii. Set the inference to the state following the previous inference and
odology B is evaluated using domain-specific knowledge, such as repeat steps (v) to (vii). For example, if the previous state was
precedence relationships. For example, if the construction operation con- ‘auto launching’ and the current state is reported as ‘span lowering’,
sists of 5 activities which should be performed sequentially, Activity 5 this step sets the state as ‘segment lifting’. The inference is again
can only take place after activity 4, and activity 4 after activity 3. Therefore, inspected by going through steps (v) to (vii).
System Identification Methodology evaluates the current inference with
the previous inference and eliminates inferences that are impossible ac-
cording to the construction sequence adopted. These heuristics, when inte- 6. Uncertainties in system identification
grated with System Identification Methodology B, has the potential to
eliminate Type 1 error or false positives which might arise in System Iden- Reliability of the system identification methodologies is dependent
tification Methodology B. Evaluation based on heuristics are performed as on different types of errors. In many cases, various errors compensate
a final check after the falsification based on slopes is done. for each other such that model predictions match the measured values
The first step in the system identification framework is to generate a [20]. Two main categories of errors are measurement error and model-
population of model instances. The model instances represent different ling error.
stages of operation of the equipment as well as possible faults that could Measurement error (emeas) is the difference between the measured
happen due to human error, environmental effects, etc. The next step is and actual values in a single measurement. This is dependent on the
to quantify the errors associated with modelling and measurement. precision of the sensor, temperature dependencies for the sensor,
Then, for each model instance, the slopes between readings at sensor lo-
cations is calculated and stored in the database. Upper and lower
threshold bound for slopes at each location are computed and stored
in the database (DB-B) as shown in Fig. 6. Similarly, another database
(DB-A) stores the slope data of the critical locations.
The process of falsification in System Identification Methodology C is
shown as a sequence of blocks in Fig. 6. The sequence is as follows:

i. Import the data from the server (Block 1)


ii. Determine the slopes at the critical locations and compare it with
slopes from simulated cases. This step is represented by Blocks 2
& 3 in Fig. 6.
iii. If a set of matching slopes are found, proceed to step (iv), else re-
port an anomaly. Block 4 represents this step in Fig. 6
iv. Try the first set of matching slopes. This step is represented by
Block 5 in Fig. 6
v. Predict the intermediate readings by following the pattern from
matching simulated case. Block 6 represents this step in Fig. 6.
vi. Compare the properties/patterns of the intermediate points with
the simulated case, if it matches, proceed to step (vii). Else try the
next set of matching slopes from Step (iii) and repeat steps (v)
and (vi). If no more matching slopes are available, report anom-
aly as shown in Blocks 7,8,14 & 15 of Fig. 6. Fig. 5. Evaluation of slope error.
R.K. Soman et al. / Automation in Construction 75 (2017) 79–90 85

Fig. 6. Methodology for system identification C.

precision in the installation etc. Modelling error (emod) is the difference and the real model whereas e2 is the error which is introduced during
in the responses between an ideal model that represents the activities the computation of solutions of the mathematical model. The error e3
and the predicted responses of the simulation model. Modelling error is the component of error introduced due to incorrect assumptions
can be divided into three components e1, e2, and e3. The component e1
is the error due to the deviation in the behavior between mathematical

Table 2
Variation of parameters for creating the population of models.

Parameter Range

Counterweight Min 0 kg
Max 66,000 kg
Segment weight Min 15,690 kg
Max 20,594 kg
Front support settlement-right Min 0 mm
Max 28 mm
Front support settlement-left Min 0 mm
Max 28 mm
Rear trolley settlement right Min 0 mm
Max 28 mm
Rear trolley settlement left Min 0 mm
Fig. 7. (a): Strain gauge (magnified image); (b): Amplifier board with XBee (ZigBee
Max 28 mm
module).
86 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

made during the simulation of the numerical model such as boundary launching girder such as auto-launching, segment lifting, post-tension-
conditions, material properties, support conditions, etc. [20]. ing and span loading.
To monitor the state of the launching girder, a custom designed
wireless sensor network (WSN) was deployed on the equipment. Sen-
ThecombinederrorðeÞ ¼ Measurementerrorðemeas Þ
þ modellingerrorðe1 þ e2 þ e3 Þ sors were placed according to the measurement system design. A 120-
Ω resistance strain gauge with gauge factor of 2.03 was selected to min-
imize the drift, since the sensor needed to be deployed for several days
In this study, the primary objective is to illustrate and test a System on the LG. It has a precision of ± 0.042 Ω. Fig. 7(a) shows the strain
Identification Methodology based on derived features and heuristics in gauge used and Fig. 7(b) shows the ZigBee module which is connected
construction. The error component e3 is reduced to an extent by incor- to the strain gauge to transmit data to the data acquisition system. The
porating a range of values for material properties, support conditions, amplifier module along with the ZigBee transmitter has a 10-bit resolu-
etc. as shown in Table 2, which are the variables used in system identi- tion i.e. it could detect a change of 3 mv/V.
fication. Candidate models are selected such that the error between pre- As per the methodology discussed in Section 3, a population of
dictions and measurements are reduced. The error component e2 is models was created from the model classes. Initially, 33 base cases
expected to be small because of the use of discrete elements. It is a gen- (Table 1) were made which simulates the actual launching girder
eral thumb rule to take the combined modelling error component to be operations.
4% [20] for linear elastic beam members. Also, it can be noted that These constitute the four operations, namely, auto-launching, seg-
proposed methodology will have better accuracy than the existing ment lifting, post-tensioning and span lowering. Around 152,656
methodologies if the modelling error component is increased to a model instances were generated by varying the different parameters
higher value. This is because, by increasing the estimate of error, more in Table 2. The variation of parameters was decided from site studies
candidate models will have predictions matching the measurements and design data. The simplified linear static base model consisted of
within the error threshold. Using heuristics, many of these models 357 nodes and 356 elements. These models were then analyzed using
might be eliminated, thereby showing improvements in the accuracy FELT, a lightweight open source finite element analysis software.
of system identification using the proposed methodology. For each case, responses at 356 locations (Average strain/response
over each of the 356 elements) from the output of finite element analy-
7. Implementation and evaluation sis were tabulated. Entropy at each measurement point is computed
from the predicted responses of models. For the sensor placement, a
This section discusses the implementation of the methodologies measurement system was designed which is based on Shannon's entro-
discussed in Section 5 on a launching girder (LG) in operation. The py [51] as described in Soman et al. [52]. Positions and priority of the
steel plate girder is loaded during all the operations of the launching sensors placement are as shown in Fig. 8.
girder. Hence, the responses from the plate girder can be used to distin- Data from the WSN is automatically transmitted to a web server.
guish between different model instances. Deploying sensors on the This data is then downloaded to a client computer in which all the sys-
plate girder would enable us to identify various operations of the tem Identification methodologies are implemented. Modelling error

Table 1
Base cases to create model class.

Case Major operation Sub-activity Remarks

1 Autolaunching Start
2 Release middle support
3 Middle support to S7
4 Rear support to old middle support position
5 Middle support to S1
6 launch till rear trolley reaches rear support Launching – 0 m
7 launch till rear trolley reaches rear support Launching – 5 m
8 launch till rear trolley reaches rear support Launching – 10 m
9 launch till rear trolley reaches rear support Launching – 15 m
10 launch till rear trolley reaches rear support Launching – 20 m
11 launch till rear trolley reaches rear support Launching – 20.75 m
12 Move the rear support Launching – 20.75 m
13 Launch till next support Launching – 25 m
14 Launch till next support Launching – 30 m
15 Launch till next support Launching – 31 m
16 Front supported on the pillar
17 Segment lifting Segment trolleys on to the launching part
18 Attaching the lifting beam
19 Lifting segment
20 Lifting segment + transfer to segment trolley Segment 1
21 Lifting segment + transfer to segment trolley Segment 2
22 Lifting segment + transfer to segment trolley Segment 3
23 Lifting segment + transfer to segment trolley Segment 4
24 Lifting segment + transfer to segment trolley Segment 5
25 Lifting segment + transfer to segment trolley Segment 6
26 Lifting segment + transfer to segment trolley Segment 7
27 Lifting segment + transfer to segment trolley Segment 8
28 Lifting segment + transfer to segment trolley Segment 9
29 Lifting segment + transfer to segment trolley Segment 10
30 Lifting segment + transfer to segment trolley Segment 11
31 Transfer to segment trolley Segment 12
32 Post-tensioning
33 Span lowering End
R.K. Soman et al. / Automation in Construction 75 (2017) 79–90 87

was limited to an upper bound of 4%. Each observed set of readings was Methodology B has a lesser precision of 0.89. With modelling error ex-
matched to one of the four model classes. The four model classes consid- cluded from the analysis, System Identification Methodology B gave
ered in this study are auto launching, segment lifting, post-tensioning, 10.72% incorrect interpretations while System Identification Methodol-
and span lowering. The interpretations obtained from the system iden- ogy A interpreted 2.01% observations wrong, i.e. System Identification B
tification algorithm, for a time stamp is compared with the log book has a precision of 0.71 while System Identification A has a higher preci-
entry, which records the activity on the site, for the same time stamp. sion of 0.86. This is because slope trends happen to match even for some
The percentage of correctly matching readings is reported for all the wrong situations. For example, slope trends of the bending moments at
cases. Also, the percentages of wrong and zero matches are also the end of segment lifting and the beginning of span lowering are sim-
documented. ilar. Hence, end of the segment lifting may be interpreted as beginning
of span lowering.
8. Results and discussion In short, Methodology B has higher Type 1 error and methodology A
has higher Type 2 error. Both the methodologies have a significant re-
This section discusses the results of the implementation of the three duction in the accuracy of falsification when the modelling error is con-
system identification methodologies on the launching girder. The sys- sidered. This is because more candidate models will have predictions
tem identification algorithm outputs either the state of the system or matching the measurements within the error threshold. Hence, more
an anomaly. An anomaly is a zero match to the model predictions. than one classification will be identified as the state of launching girder
These results are compared with the log book details of that particular resulting in higher magnitudes of Type 1 and Type 2 error.
time and checked for consistency. The data from the Wireless Sensor System identification based on derived features and heuristics (Sys-
Network was transmitted to the server once every minute. However, tem Identification Methodology C) shows better performance than the
the log book entries were updated manually only once every hour. others and has an accuracy of 92.2% when modelling error is ignored
Therefore, some of the data recorded from the WSN was ignored since and an accuracy of 88.13% when the modelling error is considered. In
there was no log book entry for that time stamp. Furthermore, the con- other words System Identification Methodology has a recall of 0.92
struction site did not operate every day during the period the data was when modelling error is ignored and a recall of 0.88 when modelling
collected. Effectively, 19,724 readings collected over a period of error is considered. It could filter out the Type 1 error from System Iden-
2 months were used for the comparison. The percentage of match be- tification Methodology B by using heuristics. Also, this methodology
tween the inferences and the actual recorded values is used as a metric could eliminate the influence of modelling error to an extent. The Sys-
to quantify the accuracy of the System Identification Methodology. Fig. 9 tem Identification Methodology C filters out the candidate models
shows the percentage of match between the inferred and actual values which have predictions inside the error threshold but doesn't satisfy
for the three system identification methodologies. the conditions based on the heuristics.
Table 3 represents the data in Fig. 9 in terms of precision and recall. Fig. 10 shows the detailed analysis of percentage of match between
Precision, also called positive predictive value, is the fraction of inferred the inferences with the log book entries for System Identification Meth-
observation that are relevant (accurate), while recall, also known as odology C, i.e. accuracy at the operations level. The system identification
sensitivity, is the fraction of relevant observations that are inferred. Pre- accuracy is more than 80% for all the operations except for span lower-
cision can be seen as a measure of exactness or quality, whereas recall is ing. This is because of the low sample size of observations since span
a measure of completeness or quantity. In other words, high precision lowering activities will take only 10% of the total time on an average.
means that the methodology returned substantially more accurate re- Hence, the number of observations for these activities would be less
sults than inaccurate ones, while high recall means that a methodology than half when compared to rest of the operations.
returned most of the accurate results. The total accuracy of identification for all the operations is 88.13%.
When the modelling error is not considered, the System Identifica- Several factors contribute to the 11.87% error in identification. Errors
tion Methodology B has a higher correct match of 82.30% compared to in the installation of the strain gauge and the assumption of initial con-
60.98% for methodology A. In other words, System Identification Meth- ditions as well as modelling errors have an influence on the final error
odology B has a recall of 0.82 while System Identification Methodology percentage. However, it is difficult to control such errors at a construc-
A has a recall of 0.61. This is mainly because System Identification Meth- tion site.
odology B uses the trends instead of direct observations to falsify the The Type 1 error (wrong match) in this System Identification Meth-
models. However, it can be seen that the number of wrong interpreta- odology is less than 1.5% and it can be considered insignificant. Howev-
tions is higher for the methodology B. With modelling error included er, the Type 2 error or in other words zero matches contribute to 10.64%.
in the analysis, System Identification Methodology B has 26.20% of the In the context of progress monitoring, the Type 2 error might not con-
observations interpreted incorrectly whereas for System Identification tribute to a serious misinterpretation since the effect is that the schedule
Methodology A, it is 8.21%. In other words, System Identification Meth- doesn't get updated for that entry. Still, the frequency of updating is sig-
odology A has a precision of 0.97 while System Identification nificantly higher than the conventional monitoring and therefore this

Fig. 8. Sensor locations on launching girder.


88 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

Fig. 9. Comparison of accuracies represented as percentage of Correct, Wrong and Zero Matches.

error has the least impact. However, for productivity or safety monitor- The strain readings were used to infer the stages of construction. This is
ing, the presence of Type 2 error would result in ambiguous interpreta- a novel concept that has never been explored before.
tions and false alarms (in cases no alarms for an actual case). From a An existing System Identification Methodology from the literature
productivity perspective, failure to update the completion may result was evaluated and tested on the actual data from a metro rail construc-
in decreased productivity of that activity which in turn leads to incorrect tion site to infer the state of operation of a launching girder. However,
decision support. this methodology yielded an accuracy of 49.32% with 8% Type 1 error
Although the sensor data was matched to one of the 33 cases (sub- and 42.4% Type 2 error. To explore improvements, a new system identi-
activities), results are identified into one of the four operations only. fication strategy based on derived features was developed, and this
This is done so due to two reasons. The level of detail in which record- showed an accuracy of 63.26% with 26.2% Type 1 error and 10.54%
ings were made in the log book was at a macro level and hence the Type 2 error. This methodology was further improved by incorporating
evaluation of sub-activities is not possible. In addition, there are uncer- heuristics and domain knowledge which yielded an accuracy of 88.13%
tainties in the installation, errors in measurement, etc. which decrease with 1.23% Type 1 error and 10.64% Type 2 error. Based on the results, it
the resolution of measurement which is a limitation of this study. can be concluded that the System Identification Methodology C is supe-
Although the study is limited to these constraints, the System Identifica- rior to System Identification Methodologies A and B in terms of accura-
tion Methodology has the potential to identify even the sub-activities cy. Also, the results show that heuristics and domain knowledge are
provided the perfect installation of the measurement system and reduc- important for reliable system identification using noisy data from actual
tion in interference with the measurement. construction sites. The addition of the domain knowledge and heuristics
to the System Identification Methodology eliminates a significant por-
tion of Type 1 error and also filters out the negative impact of the model-
9. Summary & conclusions ling error.
The current study was able to infer the state of operation of a
This paper explored the feasibility of using structural responses from launching girder from its strain readings during different operations
a launching girder to monitor the progress of construction. Viaduct with reasonable accuracies. Since the completion of each cycle of the
construction for a metro rail project was used as a case study to evaluate launching girder operations acts as a milestone activity in the schedule
the feasibility. Strain sensors were deployed at different locations of for metro rail viaduct constructions, this monitoring system can be used
the launching girder and their values were transmitted to a base station. to monitor the progress of the project.
Although the case study used was in metro rail viaduct construction,
the outcome from this study provides valuable information related to
the practicality of implementing the proposed system for monitoring
Table 3
other equipment/temporary structures such as moving formwork,
Comparison of accuracies (precision and recall).
tower cranes etc. These data could be potentially used to monitor the
Precision Recall progress of the construction, productivity, investigate opportunities to
System Identification Methodology With 4% modelling error 0.86 0.49 automate, as a feedback response for automation controls etc. The
A Without modelling 0.97 0.61 system identification algorithms developed in this study provides
error
guidelines for inferring the state of the construction activity from
System Identification Methodology With 4% modelling error 0.71 0.63
B Without modelling 0.89 0.82
structural responses of temporary structure/equipment for these
error applications. Overall, the framework proposed can be a sub-system
System Identification Methodology With 4% modelling error 0.99 0.88 within a larger sensor-based construction automation system which
C Without modelling 0.99 0.92 can monitor, analyze and control different aspects of construction in
error
real-time.
R.K. Soman et al. / Automation in Construction 75 (2017) 79–90 89

Fig. 10. Detailed accuracy of System Identification Methodology C.

Acknowledgements [17] M. Sanayei, B.R. Brenner, E. Santini-Bell, J.D. Sipple, J.E. Phelps, Lefebvre. Base-
line model updating during bridge construction using measured strains, Pro-
ceedings of the 2010 Struct. Congr. 2010, pp. 717–728, http://dx.doi.org/10.
Authors wish to thank the help provided by Larsen and Toubro Con- 1061/41130(369)66.
struction and Hyderabad Metro Rail Ltd. in enabling the field testing of [18] B. Sevim, A. Bayraktar, A.C. Altunişik, S. Atamtürktür, F. Birinci, Finite element model cal-
ibration effects on the earthquake response of masonry arch bridges, J. Eng. Mech. 47
the methodologies. The post graduate study of the first author was sup- (2009) 621–634, http://dx.doi.org/10.1016/j.finel.2010.12.011.
ported by the scholarship from Ministry of Human Resource Develop- [19] K.J. Beven, Environmental modelling: an uncertain future?: an introduction to tech-
ment (MHRD), India niques for uncertainty estimation in environmental prediction, Routledge, London,
2009, ISBN 0415457599.
[20] Y. Robert-Nicoud, B. Raphael, I.F. Smith, System identification through model com-
References position and stochastic search, J. Comput. Civ. Eng. 19 (2005) 239–247, http://dx.
doi.org/10.1061/(ASCE)0887-3801(2005)19:3(239).
[1] G. Roumeliotis, S. Fenton, Global Construction Growth to Outpace GDP This Decade, [21] K.J. Beven, Uniqueness of place and process representations in hydrological model-
http://in.reuters.com/article/idINIndia-55293920110303 2011 (accessed July 18, 2015). ling, Hydrol. Earth Syst. Sci. 4 (2000) 203–213, http://dx.doi.org/10.5194/hess-4-
[2] Global Construction Industry Outlook 2020, http://www.prnewswire.com/news-re- 203-2000.
leases/global-construction-outlook-2020-300040813.html 2015 (accessed July 18, [22] J.-A. Goulet, S. Coutu, I.F.C.C. Smith, Model falsification diagnosis and sensor place-
2016). ment for leak detection in pressurized pipe networks, Adv. Eng. Inform. 27 (2013)
[3] R. Navon, Automated project performance control of construction projects, Autom. 261–269, http://dx.doi.org/10.1016/j.aei.2013.01.001.
Constr. 14 (2005) 467–476, http://dx.doi.org/10.1016/j.autcon.2004.09.006. [23] X. Jiang, S. Mahadevan, Bayesian validation assessment of multivariate compu-
[4] I.N. Davidson, M.J. Skibniewski, Simulation of automated data collection in buildings, tational models, J. Appl. Stat. 35 (2008) 49–65, http://dx.doi.org/10.1080/
J. Comput. Civ. Eng 9 (1995) 9–20, http://dx.doi.org/10.1061/(ASCE)0887- 02664760701683577.
3801(1995)9:1(9). [24] J. McFarland, S. Mahadevan, Multivariate significance testing and model calibration
[5] G.S. Cheok, W.C. Stone, R.R. Lipman, C. Witzgall, J. Uisod, Ladars for construction as- under uncertainty, Comput. Methods Appl. Mech. Eng. 197 (2008) 2467–2479,
sessment and update, Autom. Constr. 9 (2000) 463–477. http://dx.doi.org/10.1016/j.cma.2007.05.030.
[6] E.M. Goldratt, J. Cox, The goal: a process of ongoing improvement, North River Press, [25] R. Rebba, S. Mahadevan, Validation of models with multivariate output, Reliab. Eng.
Great Barrington, MA, 2004, ISBN 978-0-88427-178-9. Syst. Saf. 91 (2006) 861–871, http://dx.doi.org/10.1016/j.ress.2005.09.004.
[7] A. Dimitrov, M. Golparvar-Fard, Vision-based material recognition for automated [26] Y. Ben-Haim, F.M. Hemez, Robustness, fidelity and prediction-looseness of models,
monitoring of construction progress and generating building information modeling Proc. R. Soc. Lond. A Math. Phys. Eng. Sci. 468 (2012) 227–244, http://dx.doi.org/
from unordered site image collections, Adv. Eng. Inform. 28 (2014) 37–49, http://dx. 10.1098/rspa.2011.0050.
doi.org/10.1016/j.aei.2013.11.002. [27] L.S. Katafygiotis, J.L. Beck, Updating models and their uncertainties. II: model
[8] I. Brilakis, M.-W. Park, G. Jog, Automated vision tracking of project related entities, identifiability, J. Eng. Mech. 124 (1998) 463–467, http://dx.doi.org/10.1016/0375-
Adv. Eng. Inform. 25 (2011) 713–724, http://dx.doi.org/10.1016/j.aei.2011.01.003. 9474(68)90693-3.
[9] M. Memarzadeh, M. Golparvar-Fard, J.C. Niebles, Automated 2D detection of con- [28] B. Goller, J.L. Beck, G.I. Schuëller, Evidence-based identification of weighting factors
struction equipment and workers from site video streams using histograms of ori- in Bayesian model updating using modal data, J. Eng. Mech. 138 (2012) 430–440,
ented gradients and colors, Autom. Constr. 32 (2013) 24–37, http://dx.doi.org/10. http://dx.doi.org/10.1061/(ASCE)EM.1943-7889. 0000351.
1016/j.autcon.2012.12.002. [29] M. Aghagholizadeh, F.N. Catbas, A review of model updating methods for civil infra-
[10] V. Pătrăucean, I. Armeni, M. Nahangi, J. Yeung, I. Brilakis, C. Haas, State of research in structure systems, Comput. Tech. Civ. Struct. Eng 83–99 (2015), http://dx.doi.org/
automatic as-built modelling, Adv. Eng. Inform. 29 (2015) 162–171, http://dx.doi. 10.4203/csets.38.4.
org/10.1016/j.aei.2015.01.001. [30] A.P. Dempster, A generalization of Bayesian inference, J. R. Stat. Soc. Ser. B Stat
[11] E.J. Jaselskis, T. El-Misalami, Implementing radio frequency identification in the con- Methodol. 30 (1968) 205–268, http://dx.doi.org/10.1007/978-3-540-44792-4_4.
struction process, J. Constr. Eng. Manag. 129 (2003) 680–688, http://dx.doi.org/10. [31] K. Alvin, Finite Element Model Update via Bayesian Estimation and Minimization of
1061/(ASCE)0733-9364(2003)129:6(680). Dynamic Residuals, AIAA J. 35 (5) (1997) 879–886, http://dx.doi.org/10.2514/2.7462.
[12] W.-S. Jang, M.J. Skibniewski, Cost-benefit analysis of embedded sensor system [32] S.H. Cheung, J.L. Beck, Bayesian model updating using hybrid Monte Carlo simula-
for construction materials tracking, J. Constr. Eng. Manag. 135 (2009) 378–386, tion with application to structural dynamic models with many uncertain parame-
http://dx.doi.org/10.1061/(ASCE)0733-9364(2009)135:5(378). ters, J. Eng. Mech. 135 (2009) 243–255, http://dx.doi.org/10.1061/(asce)0733-
[13] X.W. Ye, Y.Q. Ni, Y.X. Xia, Distributed strain sensor networks for in-construction 9399(2009)135:4(243).
monitoring and safety evaluation of a high-rise building, Int. J. Distrib. Sens. Net- [33] S.H. Cheung, T.A. Oliver, E.E. Prudencio, S. Prudhomme, R.D. Moser, Bayesian uncertainty
works 2012 (2012) 1–13, http://dx.doi.org/10.1155/2012/685054. analysis with applications to turbulence modeling, Reliab. Eng. Syst. Saf. 96 (2011)
[14] M. Golparvar-Fard, J. Bohn, J. Teizer, S. Savarese, F. Peña-Mora, Evaluation of image- 1137–1149, http://dx.doi.org/10.1016/j.ress.2010.09.013.
based modeling and laser scanning accuracy for emerging automated performance [34] B. Goller, G.I. Schueller, Investigation of model uncertainties in Bayesian structural
monitoring techniques, Autom. Constr. 20 (2011) 1143–1155, http://dx.doi.org/10. model updating, J. Sound Vib. 330 (2011) 6122–6136, http://dx.doi.org/10.1016/j.
1016/j.autcon.2011.04.016. jsv.2011.07.036.
[15] L. Joshua, K. Varghese, Automated recognition of construction labour activity using [35] J.L. Beck, S.K. Au, M.W. Vanik, Monitoring Structural Health Using a Probabilistic
accelerometers in field situations, Int. J. Product. Perform. Manag. 63 (2014) Measure, Comput. Aided Civ. Inf. Eng. 16 (2001) 1–11, http://dx.doi.org/10.1111/
841–862, http://dx.doi.org/10.1108/IJPPM-05-2013-0099. 0885-9507.00209.
[16] N.M. Okasha, D.M. Frangopol, A.D. Orcesi, Automated finite element updating using [36] K. Christodoulou, E. Ntotsios, C. Papadimitriou, P. Panetsos, Structural model
strain data for the lifetime reliability assessment of bridges, Reliab. Eng. Syst. Saf. 99 updating and prediction variability using Pareto optimal models, Comput. Methods
(2012) 139–150, http://dx.doi.org/10.1016/j.ress.2011.11.007. Appl. Mech. Eng. 198 (2008) 138–149, http://dx.doi.org/10.1016/j.cma.2008.04.010.
90 R.K. Soman et al. / Automation in Construction 75 (2017) 79–90

[37] F.L. Zhang, S.K. Au, Fundamental two-stage formulation for Bayesian system identi- [44] K.R. Popper, The Logic of Scientific Discovery, Basic Books, Inc., New York, 1959, ISBN
fication, Part I: General theory, Mech. Syst. Signal Process. (2016) 43–61, http://dx. 0091117216.
doi.org/10.1016/j.ymssp.2015.04.024. [45] A. Tarantola, Popper, Bayes and the inverse problem, Nature Physics 2 (2006)
[38] F. Zhang, S. Au, Fundamental two-stage formulation for Bayesian system identifica- 492–494, http://dx.doi.org/10.1038/nphys375.
tion, part II: application to ambient vibration data, Mech. Syst. Signal Process. 66–67 [46] B. Raphael, I. Smith, Finding the right model for bridge diagnosis, Artificial Intelli-
(2016) 43–61, http://dx.doi.org/10.1016/j.ymssp.2015.04.024. gence in Structural Engineering Lecture Notes in Computer Science 308–319
[39] X. Jiang, Y. Yuan, X. Liu, Bayesian inference method for stochastic damage accumu- (1998). http://dx.doi.org/10.1007/bfb0030459.
lation modeling, Reliab. Eng. Syst. Saf. 111 (2013) 126–138, http://dx.doi.org/10. [47] J.-A. Goulet, I. Smith, Probabilistic Model Falsification for Infrastructure Diagnosis,
1016/j.ress.2012.11.006. thesis submitted to École polytechnique fédérale de Lausanne, 2012.
[40] E. Simoen, G. De Roeck, G. Lombaert, Dealing with uncertainty in model updating for [48] J. Goulet, I. Smith, Performance-driven measurement-system design for structural
damage assessment: a review, Mech. Syst. Signal Process. 56 (2015) 123–149, identification, J. Comput. Civ. Eng. 427–436 (2012). http://dx.doi.org/10.1061/
http://dx.doi.org/10.1016/j.ymssp.2014.11.001. (ASCE)CP.1943-5487. 0000250.
[41] A.H. Alavi, H. Hasni, N. Lajnef, K. Chatti, F. Faridazar, Automation in construction an [49] K. Beven, A manifesto for the equifinality thesis, J. Hydrol. 320 (2006) 18–36, http://
intelligent structural damage detection approach based on self-powered wireless dx.doi.org/10.1016/j.jhydrol.2005.07.007.
sensor data, Autom. Constr. 62 (2016) 24–44, http://dx.doi.org/10.1016/j.autcon. [50] J. Goulet, P. Kripakaran, I.F.C. Smith, Multimodel structural performance monitoring,
2015.10.001. J. Struct. Eng. 136 (2010) 1309–1318, http://dx.doi.org/10.1061/(ASCE)ST.1943-
[42] H. Sun, O. Büyüköztürk, Probabilistic updating of building models using incomplete 541X.0000232.
modal data, Mech. Syst. Signal Process. 75 (2016) 27–40, http://dx.doi.org/10.1016/ [51] C.E. Shannon, A Mathematical Theory of Communication, Bell Syst. Tech. J. 27
j.ymssp.2015.12.024. (1948) 379–423.
[43] J.-A. Goulet, C. Michel, I.F.C. Smith, Hybrid probabilities and error-domain structural [52] R.K. Soman, B. Raphael, K. Varghese, Sensor placement to monitor launching girder
identification using ambient vibration monitoring, Mech. Syst. Signal Process. 37 operations in segmental construction, Proceedings of 32nd Int. Symp. Autom. Robot.
(2013) 199–212, http://dx.doi.org/10.1016/j.ymssp.2012.05.017. Constr. Min., Oulu, Finland, 2015.

You might also like