J Civil Struct Health Monit

DOI 10.1007/s13349-016-0160-0

ORIGINAL PAPER

A clustering approach for structural health monitoring on bridges

Alberto Diez1 • Nguyen Lu Dang Khoa2 • Mehrisadat Makki Alamdari2 •

Yang Wang2 • Fang Chen2 • Peter Runcie2

Received: 29 September 2015 / Revised: 15 December 2015 / Accepted: 15 February 2016

 Springer-Verlag Berlin Heidelberg 2016

Abstract Structural health monitoring is a process for located. The validity of the approach was demonstrated
identifying damage in civil infrastructures using sensing using real data collected from the Sydney Harbour Bridge.
system. It has been increasingly employed due to advances The clustering results showed correlations among similarly
in sensing technologies and data analytic using machine located joints in different bridge zones. Moreover, it also
learning. A common problem within this scenario is that helped to detect a damaged joint and a joint with a faulty
limited data of real structural faults are available. There- instrumented sensor, and thus demonstrated the feasibility
fore, unsupervised and novelty detection machine learning of the proposed clustering based approach to complement
methods must be employed. This work presents a cluster- existing damage detection strategies.
ing based approach to group substructures or joints with
similar behaviour on bridge and then detect abnormal or Keywords Structural health monitoring  Damage
damaged ones, as part of efforts in applying structural detection  Novelty detection  Unsupervised learning 
health monitoring to the Sydney Harbour Bridge, one of K-means clustering
iconic structures in Australia. The approach is a combi-
nation of feature extraction, a nearest neighbor based out-
lier removal, followed by a clustering approach over both 1 Introduction
vibration events and joints representatives. Vibration sig-
nals caused by passing vehicles from different joints are Ageing and damage in transport infrastructures, such as
then classified and damaged joints can be detected and roads, bridges and tunnels, are becoming a big issue
nowadays. In order to improve safety and reduce costs
derived from ageing and damage, early identification of
& Nguyen Lu Dang Khoa
[email protected]
damage in a structure is important to avoid further risks,
both in life-safety and economic losses. Structural Health
Alberto Diez
[email protected]
Monitoring (SHM) based approaches have been increas-
ingly used during recent years to address this problem [13,
Mehrisadat Makki Alamdari
[email protected]
30].
With regard to bridges, and despite the advances in
Yang Wang
[email protected]
abstract analysis and controlled testing, failures have the
most conspicuous influence on their design, construction
Fang Chen
[email protected]
and management. Many failures are mainly caused by
inappropriate design and poor maintenance (corrosion,
Peter Runcie
[email protected]
scour, etc.) [8]. As established by Sibly and Walker [29],
each innovative bridge form is developed by trial and error
1
Tecnalia Research & Innovation, Donostia-San Sebastián, method until its limits are surpassed and spectacular failure
Gipuzkoa, Spain occurs. Only then does theory catch up with the practice
2
National ICT Australia (NICTA), Sydney, Australia and fully explains the structural behaviour of the

123
 J Civil Struct Health Monit

infrastructure. Two popular examples are Tacoma Bridge, structure is in good condition, and then quantify the con-
in 1940, due to a poor design [18], and Silver Bridge, in sistency of new feature vectors with relation to the esti-
1967, due to the use of bad material [22]. The more mated density. K-means, Gaussian Mixture Models
complex dynamic behaviour of these structures resulted in (GMM), Support Vector Clustering (SVC) and Self-Orga-
greater lapses in the designer’s knowledge, and hence, nizing Maps (SOM) algorithms are representative exam-
many failures under wind loads, as well as those caused by ples of clustering-based approaches for statistical modeling
rushing crowds. and feature classification in SHM [28]. Yu et al. [39] used a
For this reason maintenance should be considered as a fuzzy c-means (FCM) clustering approach for structural
fundamental pillar to face ageing and damage effects in damage detection. Several feature extraction techniques
bridges [15]. Maintenance strategies are essentially based were applied in their study such as data projection algo-
on nonintrusive sensing, monitoring and analysis tech- rithms, principal component analysis (PCA) and kernel
niques to provide flexible decision support, normally in the principal component analysis (KPCA). Toivola et al. [31]
form of inspection recommendations (when, where and used different dimensionality reduction methods such as
why to act). To this concern, data-driven or machine Curvilinear Component Analysis, PCA and Random Pro-
learning based analysis [35] and model-driven or finite jection to reduce high-dimensional vibration measurements
element analysis [19] are most commonly used to generate in conjunction with nearest neighbor algorithm for novelty
data models (from an data analytical perspective) and detection. In [7] a hierarchical parameter clustering tech-
physical models (from a numerical or mechanical per- nique was used to group acceleration data from a cable-
spective) that aim to represent the structure. Generated stayed bridge, based on the similarity of parameters
models are then used to identify damage in the structures involved in the process. Other clustering approaches for
[34]. SHM focused on grouping information coming from sensor
This work is an effort to apply SHM to the Sydney networks distributed throughout the structure. Yin et al.
Harbour Bridge by a data-driven machine learning [38] proposed a clustering-based multi-hop routing proto-
approach. It presents a clustering based technique to group col to group similar nodes in each span of the bridge. Nie
bridge joints with similar behaviour and then detect et al. [26] also studied the use of cluster-based data
abnormal or damaged ones. Vibration events caused by aggregation architectures for SHM to transform the raw
passing vehicles are acquired by accelerometers located in data into useful information.
several joints of the structure. Since there are not clear Several signal processing and feature extraction tech-
evidences of anomalies, data are analysed in an unsuper- niques have been applied to extract features which are
vised fashion. A combination of feature extraction and sensitive to changes in the structure. Some of these tech-
outlier removal is performed, and then similar events and niques can be found in [5, 6, 11]. The most popular tech-
joints are grouped by the clustering technique. This will niques for signal processing are wavelet transform (WT)
allow isolating possible damaged joints on the bridge. and Fourier transform (FT), as discussed by Newland et al.
[25]. Alternatively, time series analysis is commonly
applied in order to extract damage sensitive features from
2 Related work SHM data. Time series models are used to fit to the
vibration data and then damage indicators are obtained by
Statistical pattern recognition has been widely used to comparing new data against the learnt models. Gul et al.
identify damage in the structure. As can be seen in [12, 21, [17] employed time series modelling in conjunction with
37], group classification, regression analysis, neural net- Mahalanobis distance to identify structural changes on
works and density estimation were popular methods. different case studies.
Nevertheless, in most of the cases success was restricted to
simulations, laboratory studies and well-controlled exper-
iments, such as in the Z24 Bridge [23]. Thus, limited 3 Proposed approach
evidences of real structural faults are available and in most
of the cases the effectiveness of the approaches still This section describes the proposed clustering approach for
remains to be validated for operational civil structures. SHM of bridges. The proposed algorithm is based on an
In an effort of overcoming this limitation and addressing unsupervised classification of vibration events and joints.
the novelty detection challenge, many authors have pro- The main steps of the methodology can be seen in Fig. 1,
posed clustering and unsupervised learning approaches for which has been applied for the presented work on the
damage detection in SHM [16]. When there is no data Sydney Harbour Bridge.
available from damaged states, one way to proceed is to The vibration responses of the structure excited by
estimate the density of feature vectors, assuming the passing vehicles, also named as events, are measured from

123
J Civil Struct Health Monit

Fig. 1 Flowchart of the proposed clustering based approach for SHM

hundreds of joints within the structure using tri-axial 3.2 k-Nearest neighbors for outliers removal
accelerometers. The measured time history of vibration
responses, corresponding to each event, is stored for further k-Nearest Neighbors (kNN) searches for the k closest
analysis. For every event, raw acceleration data are trans- neighbors in the feature space. It can be mainly used for
formed into a unique feature in time domain which is classification [10] and regression [2], predicting the label or
explained in detail in Sect. 3.1. Outliers are then removed the value of a new data sample as the majority of labels or
based on density estimation of bridge joints and the energy values among its k closest neighbors.
of the signals, allowing reducing dimension of the data, as kNN-based approach proposed within this study per-
can be seen in Sect. 3.2. Having removed the outliers, the forms an iterative process that allows removing outliers
remaining events are then transformed into the frequency and noisy signals incrementally, until the convergence
domain using fast Fourier transform (see Sect. 3.3). criteria is met. Outliers are joints signals that are far from
Finally, clustering techniques are employed to train models their joint representative, calculated as the mean of all joint
from available data that are able to characterize normal events. This also allows resampling the number of events
structural behaviour. Joints with similar behaviour are for every joint in order to balance the whole data set.
grouped and damage joints are detected. This last phase is K Dimensional-Trees (KD-Trees) are used for optimising
described in Sect. 3.4. the (k) nearest neighbors searching process [24].
For every event, X ¼ ðx1 ; . . .; xn Þ, the sum of the energy
3.1 Data preprocessing for feature extraction in time domain is calculated as,

An event is defined as a time period in which a vehicle is X

n
EðXÞ ¼ jxi j2 : ð2Þ
driving across an instrumented joint. Vibrations caused by i¼1
passing vehicles are recorded by tri-axial accelerometers
positioned at the joints located in different parts of the Then, for each iteration, the k closest neighbors to the mean
structure. For every event sample, accelerometer data in the of the energy of the joint signals, ljoint , are taken. The
three axis, (x, y, z), are transformed in a unique feature, distance of the k closest signals, dðEðXj Þ; ljoint Þ,
Vi ¼ jAi j  jAr j, where Ai is the instantaneous acceleration j ¼ 1; . . .; k, for every joint is calculated. Xj is marked as
at ith sample and Ar is the rest vector or the average of the outlier and is removed from current joint if the condition is
three readings (x, y, z) in the first 100 samples. They are met,
collected before the event is triggered for ensuring that all
the events have a similar wave form. Events that are trig- dðEðXj Þ; ljoint Þ [ lðDÞ þ 2  rðDÞ; ð3Þ
gered within the first 100 samples are filtered out (these Pk
where lðDÞ ¼ ð1=kÞ  j¼1 dðEðXj Þ, ljoint Þ is the mean of
events are mainly caused by vehicles that were driving the distance of every event energy to the joint mean and
close to each other). Data standardisation is then applied to rðDÞ is the standard deviation of such distances.
every V as, This process is repeated until any of the previously
V  lðVÞ established stopping criteria is met. Namely, a maximum
X¼ ð1Þ
rðVÞ number of iterations and a distance threshold, calculated
during the first iteration as,
where V ¼ V1 ; . . .; Vn are extracted features from an
acceleration event. threshold ¼ lðDÞ þ ð0:5  rðDÞÞ: ð4Þ

123
 J Civil Struct Health Monit

If the maximum distance obtained from the k nearest points determined by the choice of the K value, the distance
to the mean of the joint at any iteration is below this metric employed and the initial centroids values [36].
threshold, the process is stopped. Once the FFT is applied over remaining events after the
outlier removal process, Euclidean metric is computed in
3.3 Fourier transform for vibration signals order to determine the distance between events [1]. It is
processing calculated as follows,
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X n

Fourier transform is a signal processing technique that distðX; YÞ ¼ kX  Yk ¼ ðxi  yi Þ2 ; ð7Þ

i¼1
decomposes a time domain signal into the series of fre-
quencies (amplitudes and frequencies). It was first dis- where X ¼ ðx1 ; . . .; xn Þ and Y ¼ ðy1 ; . . .; yn Þ are vibration
cussed by Joseph Fourier [14], and since then it has been signals after applying FFT.
further developed, becoming a robust frequency domain K-means clustering for damage detection allows
method in modal analysis [9]. The basic idea of spectral detecting anomalies like damage which are usually isolated
analysis is to represent the original vibration signal as a in small clusters. They can be removed and further anal-
new sequence, which determines the importance of each ysed. Therefore, remaining data set can be used for training
frequency component in the dynamics of the signal. This is accurate normality models. Classification of new events is
achieved by using the discrete version of the Fourier then performed by computing their distances to previously
transform as, learnt cluster centroids that represent behaviour of interest.
X
1
Xðf Þ ¼ xðtÞe2pift ; ð5Þ
1 4 Experimental results
where f denotes the frequency at which X(f) is evaluated.
Within proposed approach, Fast Fourier Transform 4.1 Field test measurement
(FFT) is employed. The frequency spectrum of each
vibration signal in the time domain, X ¼ x1 ; . . .; xn , is thus The Sydney Harbour Bridge was opened in 1932. It is a steel
computed as, through arch bridge operated by Roads and Maritime Services
(RMS) the road authority of the state of New South Wales,
X
j¼n
Aðfx Þ ¼ xj xnðj1Þðfx 1Þ ; ð6Þ Australia. The bridge carries lanes of 8-road traffic and two
j¼1 railway lines. Traffic lane 7 is a dedicated bus and taxi lane on
the eastern side of the bridge. Lane 7 consists of an asphalt
where fx represents frequency, jAðfx Þj is signal amplitude in road surface on a concrete deck supported by concrete and
frequency domain, xj is one of the n time domain sampling
2pi
steel jack arches as in Fig. 2. There are approximately 800
points of the vibration signal and xn ¼ e n . jack arches over a total distance of 1.2 km. The jack arches are
This signal analysis technique provides a powerful difficult to access and are inspected visually every two years
spectral based diagnostic method in stationary conditions, according to standard inspection practices. Figure 2 illustrates
when there is no transient signals involved. one single jack arch in the structure.
Acceleration data from each jack arch are collected
3.4 K-means clustering for behaviour using three tri-axial accelerometers, labeled sensors 1, 2
characterization and 3 where, respectively, are mounted to the joint in left,
middle and right positions as shown in Fig. 3. The sensi-
K-means clustering algorithm performs a partition of data tivity of the accelerometers have been set to 2 g but the
space into K clusters. Each cluster is represented by an they are capable of measuring acceleration response up to
object, named centroid or the mean point, whose initial 8 g. Each tri-axial accelerometer measures acceleration
value can be randomly set or estimated by applying some responses in the direction of passing traffic (x), in the
kind of heuristic. In an iterative process, each element is direction of gravity (z) and in the direction of (y). In this
assigned to the partition with the least distance between it study, only measured responses in (z) direction of sensor 2
and the centroid of the partition. Once all elements have is utilized since (z) direction accounts for most of the
been assigned into clusters, cluster centroids are recalcu- acceleration of the middle sensor 2. Acceleration data are
lated using the information of the elements that belong to collected after the vehicles, usually buses, drive over the
each partition. The process converges to a solution with a deck where the sensors are located. A pre-set threshold is
linear complexity O(n), which is not always a global used to trigger the recording in each sensor, which corre-
optimum. The success of K-means method is strongly sponds to an event.

123
J Civil Struct Health Monit

Fig. 2 Illustration of a the bus lane on the Sydney Harbour Bridge and b one of the concrete jack arches underneath the bus lane

Fig. 3 6 Joints experiment,

schematic of the evaluated
joints

Fig. 4 The Sydney Harbour Bridge schema

6-Joint case In the first case study, 6 joints from North For both case studies, Frequency Domain Decomposi-
pylon and North main span were monitored during the first tion technique (FDD) [3] was implemented to investigate
week of August 2012, namely Joints 1 to 6 as shown in the natural frequencies of the jack arches. For an arbitrary
Fig. 3. It is known that a crack was present in the 4th joint joint within 71-joint data, FDD technique was applied on
at that time. Vibration signals were sampled at a frequency 10 randomly selected event data and the results for the first
of 375 Hz and during 1.6 s, resulting in 600 samples per normalized singular value was presented in Fig. 7a (for
event. Frequency range was set from 0 to 187.5 Hz. each event a different line pattern was used). As shown in
71-Joint case In a separate case study, 71 joints were the figure, within the frequency range of 16 to 18 Hz, a
monitored from the first week of October 2014. Monitored significant mode can be extracted. Interestingly, a variation
joints belong to the following bridge zones: span 6, span 7, is observed between natural frequencies obtained from
span 8, North pylon and North main span (see Fig. 4 for different events, which may be caused by different envi-
locations of these monitored areas). It is known that the ronmental and traffic conditions. The same behaviour was
sensor mounted on Joint 135, located in the second bay of observed from all the other joints.
span 7, was faulty at that time. Vibration signals were FDD technique was also applied on the damaged joint in
sampled at a frequency of 250 Hz during 2 s, resulting in the 6-joint case. Compared to the healthy state (after
500 samples per event. Frequency range is set from 0 to repair), there is a considerable reduction in the dominant
125 Hz. mode. Figure 7b compares the first normalized singular
Figures 5 and 6 show plots of typical vectors V ¼ values for the damaged joint before and after repair
jAj  jAr j obtained from raw acceleration data of consid- obtained from 30 randomly selected events. In Fig. 7b, the
ered data sets. Corresponding FFT signals are also results for healthy state and damaged state are, respec-
provided. tively, shown by black line and grey line.

123
 J Civil Struct Health Monit

Fig. 5 6-Joint case: an example of vibration signal (above) and its corresponding FFT (below)

Fig. 6 71 Joint case: an example of vibration signal (above) and its corresponding FFT (below)

4.2 Event-based clustering also provided as the percentage of events belonging to each
joint that are grouped in a cluster.
Event-based clustering is performed to group similar
vibration signals, and thus capturing the structural beha- 4.2.1 6-Joints case
viour. Clusters formed are shown as graphs containing the
pattern (centroid or mean values) and the variance of In this experiment, comprising Joints 1, 2 and 3 in North
grouped events for each frequency. Clusters distribution is main span and Joints 4, 5 and 6 in North pylon, the main

123
J Civil Struct Health Monit

Fig. 7 Frequency analysis of a an arbitrary joint among 71 studied joints and b damaged joint before repair (grey line) and after repair (black
line)

Fig. 8 6-Joint case: cluster 0 showed a normal behaviour (centroid and standard deviation of cluster events above and joints distribution below)

motivation was to characterize normality from joints 4.2.2 71-Joint case

events. When there is no previous knowledge about the
presence of any kind of abnormal behaviour, the goal is to This section shows results achieved when clustering events
try to isolate outliers from high-density regions, containing from 71 joints. A quick test was first conducted only
the majority of the data. In this case, notably, with K ¼ 2 considering 5 joints located in the second bay of span 7. As
normal events of all joints were grouped in a big cluster, in previous case K-means was executed with K ¼ 2 in
containing a total of 23,849 events, whereas 4662 events order to find a small cluster defining abnormal behaviour in
related to the damage, mostly located in joint 4, were the data.
isolated in a small cluster. Clusters formed can be seen in As it can be seen in Figs. 10 and 11, several events from
Figs. 8 and 9. Events belonging to the same joint are Joint 135 were located in a small cluster (together with a
equally coloured. few events of Joint 131, located in the same bay and span),

123
 J Civil Struct Health Monit

Fig. 9 6-Joint case: cluster 1 showed a damaged behaviour

Fig. 10 Analysis of 5 joints located in the second bay of span 7: cluster 0 showed a normal behaviour

123
J Civil Struct Health Monit

Fig. 11 Analysis of 5 joints located in the second bay of span 7: cluster 1 with events from a faulty sensor

showing an abnormal pattern characterized by centroid representatives of all joints was generated. A joint repre-
values. The other cluster grouped the majority of the sentative is calculated as the mean values of all events of
events, representing a normal behaviour. each joint, after outlier removal phase. As in the case of
In the next experiment, all 71 joints were clustered. The event-based clustering, the distance metric used was
goal was to group similar joints located in similar relative Euclidean distance.
positions along the bridge. K-means clustering was exe-
cuted with K ¼ 5 since 5 different bridge areas were 4.3.1 6-Joint case
involved:
For the experiment related to the analysis of 6 joints, it can
• Span 6: 6 bays, including 33 joints.
be appreciated that the distance between the representative
• Span 7: 5 bays, including 21 joints.
of joint 4 and the others is significantly high (see Fig. 13).
• Span 8: 4 bays, including 12 joints.
The more red the colour the higher the dissimilarity
• Span North pylon: 2 bays, including 7 joints.
between joints representatives. Therefore, the damage
• Span North main span: only 1 joint.
behaviour in this joint is detected.
Figure 12 shows the clustering result in one cluster. Joints
belonging to the same bay were equally colored. It can be 4.3.2 71-Joint case
seen that no useful information can be acquired given the
fact that events from one joint can appear in several clus- Since only spans 6 and 7 had sufficient instrumented joints
ters. Therefore, it is sometime not straightforward to group covering most of the span, we only calculate the map of
events from the same joint together using event-based pairwise distances for these two spans. Missing joints in
clustering. spans 6 and 7 are:

4.3 Joint-based clustering • In span 6: 1 joint in bay 4; 3 joints in bay 5 and 2 joints
in bay 6.
To overcome the weakness of the event-based clustering • In span 7: 1 joint in bay 1; 1 joint in bay 2; 2 joints in
mentioned above, joint-based clustering was utilized. In bay 3; 2 joints in bay 4; all joints in bay 5 and 2 joints in
joint-based clustering, a map of pairwise distances among bay 6.

123
 J Civil Struct Health Monit

Fig. 12 71-Joint case: events from one joint can appear in several clusters

Fig. 13 6-Joint case: damage in

joint 4 was captured

123
J Civil Struct Health Monit

Fig. 14 71-Joint cases, span 6:

joints at middle of a bay/span
behave similarly and are
different from joints at the two
ends

Resulting maps of pairwise distances for spans 6 and 7 can similarity between signals. The correlation coefficient of
be seen in Figs. 14 and 15. The black lines within the map two signals measures the level of similarity or linear
in each figure delimit regions that correspond to different dependence [33]. Comparing two sequences X and Y, the
bays of the span. Each small cell with a color shows a Pearson product-moment correlation is defined as:
pairwise distance between two corresponding joints in the covðX; YÞ
span. The higher the distance, the less similarity between qX;Y ¼ ; ð8Þ
rX rY
two related joints.
Overall, the similarities between joints located in dif- where covðX; YÞ is the covariance of (X, Y), rX and rY are
ferent bridge parts are found out. In span 6, it is shown that the standard deviations of X and Y, respectively.
joints at middle of a bay/span behave similarly and are The correlation between two signals is thus the covari-
different from joints at the two ends. It may suggest an ance divided by the product of the standard deviations of
indication of a bridge global behaviour in a span. In span 7, both signals. Resulting coefficient will vary between 1, if
although a global view is not available due to the missing there is a total negative correlation, 0, when there is no
joints, Joint 135 located in bay 2 of span 7 had consider- correlation, and 1 for the case of total correlation. It can be
ably large distances to other joints. It is known that the expressed as a dot product between the normalized signals.
sensor in Joint 135 was faulty at the time the data were For damage detection, it has been widely used to compare
collected. Joint 131 was also behaving differently com- frequency domain signals in combination with Hilbert
pared to the others, which implies a possible problem that transform [32].
should be checked. In our approach, instead of Euclidean distance, cross
correlation can be used as a similarity measure between
4.3.3 Cross correlation signals. Results obtained in the case of joint-based clus-
tering can be seen in Figs. 16, 17 and 18.
In this section the performance of the proposed approach Concerning 6-joint experiment, the results of Euclidean
was compared with the one obtained from applying cross- distance and cross-correlation were similar. However, dif-
correlation instead of Euclidean distance to measure the ferences between healthy joints and the damaged one were

123
 J Civil Struct Health Monit

Fig. 15 71-Joint case, span 7:

faulty sensor at Joint 135 was
detected

Fig. 16 6-Joint case: map of

pairwise distances using cross
correlation

123
J Civil Struct Health Monit

Fig. 17 71-Joint case, span 6:

map of pairwise distances using
cross correlation

Fig. 18 71-Joint case, span 7:

map of pairwise distances using
cross correlation

123
 J Civil Struct Health Monit

less pronounced. Whereas in the case of 71-joint experi- 6-joint case, k was set to 5000. Regarding the 71-joint case,
ment the cross-correlation detected the faulty sensor in k was set to 500, since less events per joint were available.
Joint 135 but showed a correlation coefficient with the At every iteration and for each joint, k is the number of
other joints over 0.5, which indicated they can be consid- closest events (in terms of a distance metric) to the mean of
ered moderately correlated. Cross-correlation also esti- the energy of all joint events. In case the number of events
mated worse the location dependency among joints, as it in joint i, jeventsi j\k, k is set to the number of events in
can be appreciated in Fig. 17. Joints in all bays of span 6 joint i. The maximum number of iterations was set to 10.
showed a correlation over 0.8. It therefore seems that cross Results obtained in terms of the number of events fil-
correlation coefficient is less sensitive to small changes in tered for each experiment are summarized in Table 1.
joint behaviour than Euclidean distance. It must be also Additionally, the standard deviation of signals filtered and
noticed that due to the computational cost of covariance signals selected for training are shown in Figs. 19 and 20.
computation when dealing with high-dimensional signals, The results show that filtered events had larger standard
it is an inappropriate distance estimator in its basic form deviation compared to events used for training.
[4].

4.4 Effect of kNN-based outlier removal 4.5 Online learning and computational time

kNN-based outlier removal process was executed with a For online damage detection, previously generated models
different value of k for each experiment. This is due to the from historical data can be used. When new excitation
fact that the number of available events per joint in each events are obtained, feature extraction and signal process-
case varied significantly. k must be specified according to ing steps are applied as explained above. Then, in the case
the least amount of joint events that are being analysed. of event-based classification, distances to behaviour mod-
This allows keeping a balance between joints cardinality, els are computed and similarities are established. Events
avoiding any of them to dominate the clustering process. In far from the normality can be further studied for any

Table 1 Results obtained by

6-Joint case 36,947 events 71-Joint case 45,818 events
kNN outlier removal process
Number of events for training 28,511 27,407
Number of filtered events 8436 18,411

Fig. 19 6-Joint experiment. Standard deviation of the training events (above) and the filtered events (below) after performing the kNN-based
outlier removal step

123
J Civil Struct Health Monit

Fig. 20 71-Joint experiment: standard deviation of the training events (above) and the filtered events (below) after performing the kNN-based
outlier removal step

Table 2 Computational time (in seconds) for the proposed approaches

Event-based clustering Joint-based clustering
6 Joints 36,947 events 71 Joints 45,818 events 6 Joints 36,947 events 71 Joints 45,818 events

Feature extraction 35.45 17.18 29.79 20.88

kNN 11.1 ðk ¼ 5000Þ 10.59 ðk ¼ 500Þ 9.17 ðk ¼ 5000Þ 10.47 ðk ¼ 500Þ
FFT 12.55 0.53 12.57 3.18
K-means 10.6 ðK ¼ 2Þ 55.61 ðK ¼ 5Þ 0.12 (pairwise map) 0.63 (pairwise map)
Total time 69.7 83.91 51.65 35.16
Instant warning 0.002 0.002 0.001 0.001

abnormal behaviour such as cracking and sensor failure. constructed the clustering models, a random signal was
Instant warning can be also performed by defining mem- investigated to calculate the required computational time
bership rules based on patterns of event-based models. for online damage detection. The result of this exercise is
Regarding joint-based classification, joint-based models tabulated in the last row of Table 2. According to the
are updated so that any change in the behaviour of the obtained results, it can be seen that the required compu-
structure over time can be determined. Therefore, the tational time for online warning is considerably short
overall status of the structure can be estimated and any which demonstrates the fact that combination of off-line
deviation from the normality can be detected. learning and online testing using the proposed approach
Experiments were conducted in Python 2.7.6 64bits on can be applied for real-time damage identification.
Ubuntu-Linux, on a CORE i5 desktop PC with 4 GB of
RAM memory. We used scikit-learn and scipy packages,
available at http://scikit-learn.org/stable/ [20, 27], respec- 5 Conclusions
tively. Using mentioned software and hardware processing
specifications, computational time for conducted experi- This work presents a clustering based approach to group
ments is presented in Table 2. This includes the required joints with similar behaviour on bridge and then detect
time for feature extraction, signal processing, distance abnormal or damaged joints. Two case studies involving
calculation and the total time (in seconds). Having vibration data monitored from the Sydney Harbour Bridge

123
 J Civil Struct Health Monit

were implemented to demonstrate the validity of proposed 11. Doebling S, Farrar C, Prime M, Shevitz D (1996) Damage
approach. On the basis of an iterative kNN-based outlier identification in structures and mechanical systems based on
changes in their vibration characteristics: a detailed literature
and noisy signal removal and a Fourier transform of survey. Los Alamos National Laboratory, Rep No LA-13070-MS
resulting joint events, the K-means based clustering of both 12. Farrar CR, Duffey TA, Doebling SW, Nix DA (1999) A statistical
joint events and joint representatives were performed. The pattern recognition paradigm for vibration-based structural health
clustering results indicated similarity between joints loca- monitoring. Struct Health Monit 2000:764–773
13. Farrar CR, Worden K (2007) An introduction to structural health
ted in different bridge locations and areas and helped to monitoring. Philos Trans R Soc A Math Phys Eng Sci
group those with similar behaviour. In addition, a damaged 365(1851):303–315
joint and a joint with a faulty sensor were also captured as a 14. Fourier J (1820) Méemoire sur le refroidissement séeculaire du
result of clustering, demonstrating the robustness of the globe terrestre. Ann Chim Phys (2) 13:418–437
15. Frangopol DM, Liu M (2007) Maintenance and management of
approach when detecting a structural damage or a sensor civil infrastructure based on condition, safety, optimization, and
problem. It is thus proved the validity of this approach for life-cycle cost. Struct Infrastruct Eng 3(1):29–41
complementing existing damage detection techniques. 16. Fugate ML, Sohn H, Farrar CR (2000) Unsupervised learning
The proposed approach can also be easily deployed to methods for vibration-based damage detection. In: Proceedings of
18th international modal analysis conference—IMAC, p 18
provide a real-time health score of the structure. Once the 17. Gul M, Necati Catbas F (2009) Statistical pattern recognition for
behaviour models that best represent the structure are learnt structural health monitoring using time series modeling: theory
from historical data, they can be used to check new events and experimental verifications. Mech Syst Signal Process
using distances to normality models obtained, namely 23(7):2192–2204
18. Irwin PA, Stoyanoff S, Xie J, Hunter M (2005) Tacoma narrows
clusters centroids and joints representatives. Thus, any 50 years laterwind engineering investigations for parallel bridges.
potential damage in the structure can be detected. In order Bridg Struct 1(1):3–17
to improve the clustering result, further experiments can be 19. Jafarkhani R, Masri SF (2011) Finite element model updating
conducted focusing on vehicle tracking to reduce variance using evolutionary strategy for damage detection. Comput Aided
Civ Infrastruct Eng 26(3):207–224
among different events. 20. Jones E, Oliphant T, Peterson P et al (2001) SciPy: open source
scientific tools for Python. Retrieved from http://www.scipy.org/
Acknowledgments The main author would like to thank National 21. Lederman G, Wang Z, Bielak J, Noh H, Garrett J, Chen S et al
ICT Australia for the great support to this work and during his (2014) Damage quantification and localization algorithms for
internship, as part of the work on his diploma thesis. The authors also indirect SHM of bridges. In: Proceedings of the international
wish to thank the Road and Maritime Services (RMS) in New South conference on bridge maintenance, safety management, Shang-
Wales for provision of the support and testing facilities for this hai, China
research work. 22. Lichtenstein AG (1993) The silver bridge collapse recounted.
J Perform Constr Facil 7(4):249–261
23. Maeck J, Peeters B, De Roeck G (2001) Damage identification on
the z24 bridge using vibration monitoring. Smart Mater Struct
References 10(3):512
24. Moore AW (1991) An introductory tutorial on kd-trees. Techni-
1. Aggarwal CC, Yu PS (2013) Proximity-based outlier detection. cal Report No. 209, Computer Laboratory, University of
In: Outlier analysis. Springer, New York, pp 101–133 Cambridge
2. Altman NS (1992) An introduction to kernel and nearest-neigh- 25. Newland DE (2012) An introduction to random vibrations,
bor nonparametric regression. Am Stat 46(3):175–185 spectral & wavelet analysis, 3rd edn. Dover Publications Inc,
3. Brincker R, Zhang L, Andersen P (2000) Modal identification New York
from ambient responses using frequency domain decomposition. 26. Nie P, Li B (2011) A cluster-based data aggregation architecture
In: Proceedings of 18th international modal analysis confer- in WSN for structural health monitoring. In: 7th International
ence—IMAC, pp 625–630 wireless communications and mobile computing conference,
4. Cao G, Bouman C (2009) Covariance estimation for high 2011. IWCMC 2011, pp 546–552
dimensional data vectors using the sparse matrix transform. In: 27. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B,
Advances in neural information processing systems, vol 21, Grisel O et al (2011) Scikit-learn: machine learning in python.
pp 225–232 J Mach Learn Res 12:2825–2830
5. Carden EP, Fanning P (2004) Vibration based condition moni- 28. Santos A, Figueiredo E, Costa J (2015) Clustering studies for
toring: a review. Struct Health Monit 3(4):355–377 damage detection in bridges: a comparison study. Struct Health
6. Chang PC, Flatau A, Liu S (2003) Review paper: health moni- Monit 2015
toring of civil infrastructure. Struct Health Monit 2(3):257–267 29. Sibly P, Walker A, Stephenson R, Moisseiff L et al (1977)
7. Cho S-J (2011) Structural health monitoring of cable-stayed Structural accidents and their causes, l. Proc Inst Civ Eng
bridge using wireless smart sensors. Ph. D. Dissertation, KAIST, 62:191–208
Daejeon, Korea 30. Sohn H, Farrar CR, Hemez FM, Shunk DD, Stinemates DW,
8. Collings D (2008) Lessons from historical bridge failures. Proc Nadler BR et al (2004) A review of structural health monitoring
Inst Civ Eng-Civ Eng 161(6):20–27 literature: 1996–2001. Los Alamos National Laboratory, Los
9. Cooley JW, Tukey JW (1965) An algorithm for the machine cal- Alamos
culation of complex fourier series. Math Comput 19(90):297–301 31. Toivola J, Prada MA, Hollméen J (2010) Novelty detection in
10. Cover T, Hart P (1967) Nearest neighbor pattern classification. projected spaces for structural health monitoring. In: Advances in
IEEE Trans Inf Theory 13(1):21–27 intelligent data analysis ix. Springer, Berlin, pp 208–219

123
J Civil Struct Health Monit

32. Tsai D-M, Lin C-T, Chen J-F (2003) The evaluation of normal- 37. Yeung W, Smith J (2005) Damage detection in bridges using
ized cross correlations for defect detection. Pattern Recognit Lett neural networks for pattern recognition of vibration signatures.
24(15):2525–2535 Eng Struct 27(5):685–698
33. Wei WW-S (1994) Time series analysis. Addison-Wesley, 38. Yin A, Wang B, Hu X, Dai Z (2009) MHop-CL: a clustering
Reading protocol for bridge structure health monitoring system. In:
34. Wenzel H (2008) Health monitoring of bridges. Wiley, New York International symposium on computer network and multimedia
35. Worden K, Manson G (2007) The application of machine technology, 2009. CNMT 2009, pp 1–4
learning to structural health monitoring. Philos Trans R Soc A 39. Yu L, Zhu J-H, Yu L-L (2013) Structural damage detection in a
Math Phys Eng Sci 365(1851):515–537 truss bridge model using fuzzy clustering and measured FRF data
36. Wu W, Xiong H, Shekhar S (2004) Clustering and information reduced by principal component projection. Adv Struct Eng
retrieval, vol 11. Springer, Brelin 16(1):207–218

123