Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Developments in the Built Environment 14 (2023) 100166

Contents lists available at ScienceDirect

Developments in the Built Environment


journal homepage: www.sciencedirect.com/journal/developments-in-the-built-environment

A new damage indicator based on homogeneity and wireless


accelerometers for evaluating the structural condition of a
cable-stayed bridge
Martin Valtierra-Rodriguez a, Jose M. Machorro-Lopez b, Juan P. Amezquita-Sanchez a, *,
Aurelio Dominguez-Gonzalez c, Shaila Alvarez-Junco c, Hector M. Gasca-Zamora d
a
ENAP-RG, CA Sistemas Dinámicos y Control, Facultad de Ingeniería, Departamento de Electromecánica, Universidad Autónoma de Querétaro, Campus San Juan del
Río, 76807, San Juan del Río, Querétaro, Mexico
b
Investigador CONACYT - Instituto Mexicano del Transporte, km 12 Carretera Estatal No. 431 “El Colorado-Galindo” San Fandila, 76703, Pedro Escobedo, Querétaro,
Mexico
c
Facultad de Ingeniería, Departamento de Electromecánica. Universidad Autónoma de Querétaro, Campus San Juan del Río, 76807, San Juan del Río, Querétaro,
Mexico
d
Instituto Mexicano del Transporte, km 12 Carretera Estatal No. 431 “El Colorado-Galindo” San Fandila, 76703, Pedro Escobedo, Querétaro, Mexico

A R T I C L E I N F O A B S T R A C T

Keywords: Structural health monitoring is crucial to avoid devastating human and economic losses, mainly for large and
Cable-stayed bridge high-demanding structures as the case of vehicular bridges. In this regard, a methodology to identify a cable loss
Homogeneity index in a cable-stayed bridge from the analysis of vibration signals, which are provided by a set of wireless accel­
Multilayer perceptron neural network
erometers installed on the bridge, is here presented. Regarding the method, the damage detection is carried out
Structural health monitoring
Vibration signals
by a new proposed damage indicator, which is calculated from the homogeneity values of the vibration signals.
Wireless accelerometers The homogeneity values are also employed as inputs in a multilayer perceptron neural network to determine the
damage location when a damage condition is identified. The usefulness of the proposal is tested on the vibra­
tional responses obtained in experimental manner from the Rio Papaloapan Bridge located in Veracruz, Mexico,
resulting in 93.33% of effectiveness to detect and locate the cable loss.

1. Introduction down along with 38 vehicles from a height of 45 m, killing 43 people and
injuring 16. The main causes of the collapse were the corrosion of one of
Damage identification in bridges has been a major interest research the cables, which led to its rupture, as well as the significant decrease of
topic in civil engineering over the years (Lee et al., 2022; Cheng and Ni, the bridge useful life due to the increase of traffic volume (A.F.P, 2018).
2018; Farrar and Worden, 2007; Glišić and Inaudi, 2007). Collapses of Therefore, the development of reliable and low-cost methods to identify
this kind of structures cause huge economic losses and human fatalities, non-visible damage in bridges is crucial to avoid catastrophic collapses
as it has happened in many countries throughout history (Lee et al., and safeguard the structural integrity of these civil constructions. Thus,
2022; Cheng and Ni, 2018; Farrar and Worden, 2007; Glišić and Inaudi, structural health monitoring (SHM) methods based on the analysis of the
2007; Reimann, 2017; Swenson and Ingraffea, 1991; Davey and Wald, dynamic responses of bridges have been considered as the most
2007; A.F.P, 2018). Two recent examples of the tragic consequences of adequate options to identify damages in this type of constructions in an
not detecting damage in bridges early enough are the collapses of the efficient way (Farrar and Worden, 2007; Farrar and Lieven, 2007).
Bridge 9340 in 2007 and the Morandi Bridge in 2018, which took place Vibration-based methods for identifying damage in civil infrastruc­
in Minnesota, USA and in Genoa, Italy, respectively. The Bridge 9340 ture can be classified in two methods: parametric and non-parametric
collapsed into the Mississippi River, where unfortunately 13 people died (Yañez-Borjas et al., 2021). Parametric methods operate with data in
and 145 were injured (Davey and Wald, 2007). On the other hand, as for the modal domain (Cheng and Ni, 2018; Fan and Qiao, 2010; Kullaa
the Morandi Bridge, one of its longest sections (210 m) collapsed, falling et al., 2010; Limongelli and Zingoni, 2019), while non-parametric

* Corresponding author.
E-mail address: jamezquita@uaq.mx (J.P. Amezquita-Sanchez).

https://doi.org/10.1016/j.dibe.2023.100166
Received 26 December 2022; Received in revised form 21 April 2023; Accepted 23 April 2023
Available online 29 April 2023
2666-1659/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

methods use data in either time domain or frequency domain (Chen, some disadvantages, including a meticulous process to find the most
2009; Barbosh et al., 2020; Civera and Surace, 2021; Bueno-Lopez et al., useful mother wavelet to make evident the damage presence in each
2019; Machorro-Lopez et al., 2021; Kankanamge et al., 2020; Liew and application, as well as the need of removing the edge effects into the
Wang, 1998; Jiang and Adeli, 2007; Amezquita-Sanchez et al., 2013; CWT diagrams produced by the inherent discontinuities at the ends of
Perez-Ramirez et al., 2020; Tee, 2018; Carden and Brownjohn, 2008; any time signal, which can be interpreted as damage (Machorro-Lopez
Gul and Catbas, 2009; Gao et al., 2021). Hence, the parametric methods et al., 2021). Recent research involving the use of the WT to detect and
utilize the modal parameters of structures (e.g., mode shapes, natural locate damage in bridges is presented in (Tan et al., 2022), where a
frequencies, and damping ratios), as well as their corresponding de­ methodology was developed based on a damage index calculated from
rivatives, to determine their health states (Cheng and Ni, 2018; Fan and the wavelet energy obtained by using the vibration measurements pro­
Qiao, 2010; Kullaa et al., 2010; Limongelli and Zingoni, 2019). Some duced by moving vehicles. In the MUSIC-based methods, algorithms
examples of parametric methods used to detect damage in based on subspace techniques are used, which post-process time signals
cable-supported bridges are the ones presented in (Ni et al., 2008; embedded in high-level noise to define their corresponding frequency
Quintana-Rodriguez et al., 2014; An et al., 2016). In the first work (Ni components, being capable of detecting frequencies with
et al., 2008), a finite element method (FEM) model was developed in low-amplitudes, which is useful for damage identification in structures.
order to obtain the corresponding modal flexibility matrices of a bridge However, in order to perform reliable analyses to evaluate the integrity
and, then, a damage index was established from the change of the condition of a civil construction, these methods require a significant
relative flexibility. On the other hand, in the second research (Quinta­ computational burden and the appropriate order selection (Jiang and
na-Rodriguez et al., 2014), a FEM model of a real-life bridge was ob­ Adeli, 2007; Amezquita-Sanchez et al., 2013; Perez-Ramirez et al.,
tained and calibrated from the experimental monitoring and, then, an 2020). Lastly, the TSM-based methods operate directly with data in the
algorithm was developed in order to correlate the experimental and time domain, avoiding transformations of domain and using statistical
numerical signals and identify damage from changes of the structural concepts to extract useful features, which change with the damage
parameters of the FEM model. Finally, in (An et al., 2016), a simple presence; moreover, they do not present the leakage and truncation
change of the natural frequency of the cables of a cable-stayed bridge errors associated with other methods (Tee, 2018; Carden and Brown­
was used to detect damage in these elements but, with the particularity, john, 2008; Gul and Catbas, 2009; Gao et al., 2021). Some examples of
that a steel bar must be installed in order to amplify the frequency the TSM-based methods are the Autoregressive (AR) models. Despite
change due to the damage. Regardless that these techniques have pro­ providing promising results, these methods have also important disad­
vided interesting results to detect damage in bridges, they are signifi­ vantages, which limit their applications in analysis of real-life structures,
cantly affected by the environmental conditions and, even more, among which can be mentioned the need of the suitable order selection,
generally require a detailed numerical model of the civil structure for the difficulties to interpret dynamic data of structures with nonlinear
predicting its dynamic behavior under different damage scenarios to be behaviors, and the complications produced by noise effects (Gao et al.,
compared with the healthy case (baseline), which is a task that involves 2021). The above mentioned non-parametric methods have provided
a substantial computing time (Kullaa et al., 2010; Limongelli and Zin­ interesting results for important advances in SHM; however, more
goni, 2019). research is still needed in order to develop other methods, based on
The non-parametric methods have reduced some limitations of the vibration signals analysis, to identify damage in civil structures effi­
parametric methods and have been applied for assessing the health state ciently and at low cost in terms of computational cost.
of civil constructions, such as the bridges. Among these methods, stand In recent years, the homogeneity algorithm, a nonlinear index, has
out the ones based on the empirical mode decomposition (EMD), been applied in diverse applications (Bekaddour and Amine, 2015;
wavelet transform (WT), multiple signal classification (MUSIC), and Martinez-Herrera et al., 2022; Amezquita-Sanchez et al., 2018; Razo-­
time series models (TSM), which have been studied in detail in recent Hernandez et al., 2018; Lizarraga-Morales et al., 2017) e.g., evaluation
years (Chen, 2009; Barbosh et al., 2020; Civera and Surace, 2021; of the induction motor condition (Martinez-Herrera et al., 2022), sudden
Bueno-Lopez et al., 2019; Machorro-Lopez et al., 2021; Kankanamge cardiac death prediction (Amezquita-Sanchez et al., 2018), detection of
et al., 2020; Liew and Wang, 1998; Jiang and Adeli, 2007; Amezqui­ power quality disturbances (Razo-Hernandez et al., 2018), among
ta-Sanchez et al., 2013; Perez-Ramirez et al., 2020; Tee, 2018; Carden others. Homogeneity is characterized by measuring the uncertainty in
and Brownjohn, 2008; Gul and Catbas, 2009; Gao et al., 2021). The 1D signals; in general, it offers a uniformity or complexity measure
methods based on EMD decompose the analyzed time signal into its contained in a time signal (Lizarraga-Morales et al., 2017). Hence, it can
principal components; and then, damage spikes, generated by abrupt be an appropriate nonlinear indicator for defining the integrity state of
structural stiffness changes from the acquired data, are extracted (Chen, infrastructure, since their vibrational responses can vary in accordance
2009; Barbosh et al., 2020; Civera and Surace, 2021). Despite the ben­ with the damage level as well as its location, generating a change in the
efits of these methods, they present some drawbacks for determining the homogeneity value.
integrity condition of structures, being the main ones the significant In this work, the capabilities of the homogeneity measurement or
computing time according to the technique used and the problem of index (HI) for identifying a cable loss in the cable-stayed bridge Rio
mode mixing which is due to closely spaced frequency components of Papaloapan Bridge (RPB), subjected to dynamic vibrations, are investi­
the signal, complicating the evaluation of the structure (Civera and gated. HI is utilized to characterize changes in the acquired vibrational
Surace, 2021; Bueno-Lopez et al., 2019). An example of how the signals and, then, be employed for determining the RPB condition. In
EMD-based methods is used to identify damage in bridges can be found this regard, as main contribution, the calculated HI values are integrated
in (Meng et al., 2018), where the measured signals of a laboratory sus­ into a new damage indicator (DI) for defining the global state of the
pension bridge were decomposing by using EMD; then, useful features bridge. Additionally, the variations of the obtained HI values with a
were extracted by means of the Hilbert transform; and finally, a damage proposed multilayer perceptron neural network (MLP-NN) are
indicator was established by using the marginal spectrum. Likewise, the employed for determining the damage location. For validating the
WT-based methods can detect and locate damage by means of the ir­ proposal, the vibrational responses experimentally measured by a set of
regularities or discontinuities generated into the time signals (Cheng wireless accelerometers from the RPB located in Veracruz, Mexico are
and Ni, 2018; Machorro-Lopez et al., 2021; Kankanamge et al., 2020; studied. The results demonstrate that the method developed in this
Liew and Wang, 1998), which produce high values of the coefficients article is reliable to discover useful features into the noisy signals for
across the CWT (continuous wavelet transform) and they can be related efficiently detecting and locating a damage in the RPB.
to defects in the structure (Machorro-Lopez et al., 2021; Kankanamge
et al., 2020; Liew and Wang, 1998). Nevertheless, these methods have

2
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

2. Methodology Secondly, a histogram, hj(m), of Dj(n) is computed, which varies from


− K+1 to K− 1.
Civil structures are daily subjected to undesirable dynamic excita­ Thirdly, a probability function, PD(m), of hj(m) is calculated by:
tions e.g., traffic, earthquakes, wind, among others, which can produce
hj (m)
an inevitable deterioration of structural elements (Park et al., 2019). In PD (m) = (2)
A
this regard, the vibrational response generated by a civil structure is a
suitable alternative for determining its state since it (1) contains the where A represents the total number of differences in Dj(n).
dynamic alterations produced by a damage and (2) can be acquired in a Finally, the homogeneity value of a 1D signal is calculated by:
constant manner, permitting the online evaluation of the structural

condition (Mousavi et al., 2021). HI =
1
PD (m) (3)
A representative diagram of the proposal for establishing the cable- m
1 + m2
stayed bridge condition is shown in Fig. 1. As the first step, the vibra­
It is worth noting that HI requires a time-invariant integer, j, for
tional responses of the cable-stayed bridge subjected to dynamic vi­
estimating the complexity or uniformity of a time signal. Liz­
brations, generated mainly by wind and traffic, are acquired for both
arraga-Morales et al. (2017) noted that for time-varying or nonsta­
conditions: (1) healthy and (2) damage produced by a cable loss. Later,
tionary signals the HI value varies according to the j selected. In this
the acquired signals for both conditions are analyzed by HI in order to
regard, a lower number of j implies a compact histogram and thus a
characterize patterns with the potential of evaluating correctly the
higher HI value is obtained; on the contrary, a higher j value implies an
cable-stayed bridge condition. As the second step, the Kruskal–Wallis
expanded histogram and thus a lower HI value is estimated. Hence, in
method (KWM) is utilized to identify the most reliable HI values in order
this work different values of j are investigated in order to identify more
to determine the overall condition of the bridge. Then, a feature
suitable features to evaluate the integrity condition of the studied
reduction is carried out according to the KWM results. Later, based on
cable-stayed bridge.
the selected features in previous step, a new damage indicator (DI) is
introduced to estimate the global condition of cable-stayed bridge,
which is formulated by integration of the estimated HI values from all 2.2. Feature selection
sensors. Once the bridge condition is determined by DI, the damaged
zone is located by a proposed MLP-NN, which is also based on the KWM, a nonparametric method, is employed for comparing the dif­
estimated HI values. The main steps of the proposed method are detailly ferences among datasets with non-normal or normal distribution
described in the following sub-sections. (Kruskal and Wallis, 1952). For performing this task, a probability value
(p-value), based on a chi-square distribution, must be computed, which
denotes the probability of rejecting a null hypothesis (NH). This value is
2.1. Homogeneity feature generally set in 0.01 (Pakyari and Balakrishnan, 2012), where the me­
dians of datasets are analyzed for determining if they present similar or
Homogeneity is a texture index initially employed for characterizing equal values (Kutner, 2005). In this regard, if p-value is lower than NH
patterns into images (Unser, 1986); however, it has recently proved to value, the NH is rejected; therefore, it is established that the datasets
be a suitable index for examining time signals (Lizarraga-Morales et al., contain patterns with the capability for being associated to the phe­
2017). In particular, it is characterized by computing the uniformity or nomena studied. In contrast, if p-value is higher than NH value, the NH
complexity of time signal, which varies between 0 and 1, indicating a is accepted, indicating that the datasets present similar values and they
high, i.e., a time-varying signal, and low, i.e., a quasi-stationary signal, cannot be useful for distinguishing among the diverse conditions eval­
uniformity of the patterns inherent in a time signal (Lizarraga-Morales uated. Therefore, KWM is utilized for discovering the most convenient
et al., 2017). In this sense, the HI can characterize an alteration into time patterns for evaluating the health state of the cable-stayed bridge.
signals associated to a damage encountered in a civil structure.
HI is calculated as follows:
2.3. Condition assessment of the bridge
Firstly, a vector of differences for a 1D signal, x, with N samples, is
computed by:
Civil infrastructure is characterized by changing their physical
Dj (n) = x(n) − x(n − j) (1) properties such as stiffness, damping ratio, mass, among others, when
they experience a damage. These changes can be monitored in their
where j represents a time-invariant integer and n varies from 0 to N-1 dynamic response (Ruocci et al., 2014). Taking in consideration the
samples, which are quantized in K different levels. This process consists aforementioned affirmation, the authors have the research ideology that
of transforming the input time signal in a discrete time signal defined by the HI values obtained by all sensors located along the structure can
the analog-digital converter (ADC) resolution (number of bits of ADC) exhibit a similar value when the bridge has a healthy condition; on the
employed for acquiring the time signal. For instance, if ADC is of 10 bits, contrary, when the bridge has a damage condition, the HI values will
it offers 210 = 1024 possible quantizing levels in which the analyzed present some type of variability in their values.
time signal can be transformed. Therefore, it is required to suggest a DI capable of determining the

Fig. 1. Block diagram of the proposal.

3
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

healthy or damage condition of the bridge by integrating in a quanti­ on the contrary, if a damage occurs in a specific zone (i.e., a damaged
tative way the variations of the HI values provided by all sensors. In this zone, DZi for i = 1, 2, 3, …, n-1, n), the closest sensor will measure the
sense, for determining if the bridge is damaged, the following new DI vibration signal with the strongest changes (e.g., the blue square in
based on HI values is proposed. Fig. 2a for each DZi). It is important to note that in this work only the loss
( ⃒) of one cable is considered, i.e., there are no simultaneous losses of cables
Ns ⃒
1 ∑ ⃒HIi − std(HI)⃒
that could generate, for example, two different blue squares in Fig. 2a. It
DI = 1 − ⃒ ⃒ (4)
Ns i=1 ⃒ min(HI) ⃒ is also worth mentioning that the analyzed damage (i.e., the loss of a
cable) is a real damage in a real-life bridge; therefore, different damage
where HIi indicates the HI value at the position i, min(HI) and std(HI) are configurations, as the case of a laboratory experiment where simulta­
the minimum and standard deviation values of HI at all the positions, neous damages at different locations can be carried out, are not possible,
respectively, and Ns represents the number of sensors. limiting the available data. Yet, the proposed strategy can be applied
It should be noted that the DI is introduced in this work to measure under this laboratory conditions.
the variability of HI values at all locations in order to determinate the From Fig. 2a, it is evident that the damage location can be carried out
bridge condition. In this regard, firstly, the difference, HIi − std(HI), is through a pattern recognition algorithm, i.e., an algorithm that recog­
proposed because it allows evaluating the variability/dispersion of each nizes where the blue square is. In this regard, the MLP-NN is proposed by
HI value (i.e., if a bridge in healthy or damaged, the HIs will maintain its easy implementation and low complexity (Abiodun et al., 2019) (see
similar or different values due to the low and high standard deviation Fig. 2b). In general, the MLP-NN consists of an input layer, a hidden
values, respectively). Secondly, each difference i is divided by the layer, and an output layer, where each layer is compound of neurons. In
minimum value of HI, min(HI), of all sensors, which allows obtaining a the proposed architecture, the quantity of input neurons is determined
value from 1 (if the obtained minimum value is similar to the difference) by the number of available sensors, since each input corresponds to the
to 0 (if the obtained minimum value is larger than the difference, which HI value obtained from the vibration signal measured at the sensor
occurs when the damage produces a variation in the HI values, mainly in location. On the other hand, the amount of output neurons is equal to the
the zone affected). The resulting summation is divided by the number of number of inputs since each output represents the place of the damaged
sensors, 1/Ns , in order to determinate a proportional variation of the HI zone. Thus, the damage location will be given by the activated output
values, generating a value between 0 and 1. Finally, the resulting value neuron. If none output neuron is activated, the condition of the bridge is
is subtracted from 1 in order to obtain a final DI value, where a closer healthy; however, this condition is firstly assessed by the proposed DI.
value to 0 indicates the healthy condition and a farther value to 0 in­ Finally, the quantity of neurons in the hidden layer is estimated by a
dicates the damage condition. To differentiate between the two condi­ trial-and-error process, i.e., different amounts of neurons are tested and
tions, a threshold value has to be established. Throughout the analysis of the number with the best accuracy result is chosen. In order to map the
the existing data and the experimentation carried out, this value can be desired outputs according to the given inputs, the MLP-NN is trained by
determined. For the bridge treated in this work, a value of 0.5 is iden­ the backpropagation algorithm. This algorithm is a supervised learning
tified and discussed later in the article. This value can change according method which minimizes the error between the desired and the calcu­
to the in-test structure, lated outputs through the adjustment of weights and bias. The output of
each neuron is given by (Abiodun et al., 2019):

2.4. Damage location ∑


n
yj = wij xi + θj (5)
i=1
Damage location is a fundamental task to apply fast and proper so­
lutions, mainly considering large structures as the case of cable-stayed where y, wi, xi, and θ represents the output, i-weights, i-inputs, and bias
bridges. In this regard, the MLP-NN is proposed to analyze the HI for the node j, respectively. The output value is also evaluated by an
values obtained from the sensors which are located along the bridge and, activation function.
from them, recognize where the damage is. In other words, the For the MLP-NN, a confusion matrix can be used to summarize its
assumption is that the closest sensor to the damage will measure a vi­ performance according to the number of correct and incorrect pre­
bration signal with the strongest changes associated to the damage. dictions. The accuracy, A, precision, P, recall, R, and F1 values of the
Fig. 2a shows the previous idea. That is, if the bridge is healthy (HLT), all suggested MLP-NN can be computed as follows:
the sensors will capture a similar vibration signal (e.g., empty squares);

Fig. 2. (a) Damage location strategy and (b) proposed MLP-NN for damage location.

4
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

TP + TN studied bridge (Yañez-Borjas et al., 2021)), producing 115,200 samples


A= (6)
TP + TN + FP + FN by each sensor for generating its health condition. On the other hand,
the cable No. 6 from SH 3 was removed in order to generate a damage
P=
TP
(7) condition as shown in Fig. 7, where the red line indicates the cable
TP + FP removed. In order to maintain the consistence between both conditions,
the bridge was monitored during 30 min with the same sampling fre­
TP
R= (8) quency. For both conditions, only a dataset is obtained by each scenario
TP + FN
of bridge; therefore, the monitored vibratory response is segmented in
P⋅R time windows of 1 min each one, allowing obtaining somehow 30 tests
F1 = 2 (9) for each condition.
P+R

where TP, TN, FP, and FN indicate true positives, true negatives, false 4. Results
positives, and false negatives, respectively.
Following the flowchart steps of the proposed methodology, the
3. Application to the RIO Papaloapan brige accelerations acquired by both conditions of the bridge, healthy and
cable loss, are evaluated through the HI method. Fig. 8 shows the
A cable-stayed bridge named the Rio Papaloapan Bridge which is vibrational responses acquired by the twelve wireless accelerometers
located in Veracruz, Mexico, is employed to develop and test the pro­ when the RPB condition is healthy and damaged (cable loss). Based on
posed methodology. The bridge, built in 1994 and shown in Fig. 3, is this figure, it should be pointed out that substantial alterations or vari­
407.2 m of length and 203 m of the main span. It consists of 8 semi-harps ations between both conditions are not visually identified because the
(represented by the word SH in Fig. 4), where each one of them has 14 vibrational responses of a civil structure are nonstationary and
cables, resulting in 112 cables in total. In order to anchor each cable on a nonlinear, restricting a direct and reliable recognition of features or
pylon, an upper anchoring system is employed,. It is based on a steel patterns with the potential of evaluating the structure condition
plate attached to the anchoring element by welding. The shape of the adequately. Therefore, this study offers a new computer-aided assess­
anchoring element is cylindrical on one side and flat on the welded side ment methodology for identifying features or patterns with the ability of
as shown in Fig. 5. estimating/evaluating the health state of the cable-stayed bridge.
During the years 2000 and 2015, the bridge suffered fractures in its It is important to mention that HI requires a time-invariant integer j
elements of anchoring, resulting in the loss of cables. For this reason, it (a distance of separation between two data points of the analyzed signal)
was instrumented with a remote monitoring system and complementary for estimating the complexity or uniformity of the vibration signal
instrumentation based on wireless accelerometers, in order to analyze its correctly. This value can change according to the application. To select
vibratory response and structural behavior. In particular, the bridge, in this value, j is varied from 1 to 10 (this range is investigated since it
the year 2019, was monitored during a maintenance service by means of allows capturing/identifying suitable patterns into the analyzed signal
twelve LORD MicroStrain® wireless sensors model G-Link®-LXRS®, (Lizarraga-Morales et al., 2017)), resulting that the j = 7 value presents
which are high-speed and low-cost tri-axial MEMS accelerometers with a the best results. It is worth mentioning that this task is performed by
measurement range of ±2 g (bridge was exposed to ambient vibrations means of KMW, which allows estimating the data distribution of HI
produced by wind and random traffic). These accelerometers have an values as well as its p-value for determining the most discriminant
internal temperature sensor and are suitable for remote monitoring and patterns. Fig. 9 illustrates the data distribution of the estimated HI
vibration data acquisition of structures. Moreover, these sensors have a values for diverse values of j, where it is possible to appreciate that the
sampling frequency up to 4096 Hz, allow up to 2 km of wireless range, HI values obtained by the twelve sensors for both conditions, respec­
and have low power consumption for extended use. The gateways are tively, are joined in order to identify the most reliable j value for
the fundamental components of the wireless sensing system since they assessing the bridge condition, i.e., for separating the two conditions.
coordinate the wireless transmissions among a network of nodes (sen­ According to this figure, it is possible to observe that the j = 7 presents a
sors). The wireless communication protocol between nodes and gate­ larger separation or variation and lesser overlap between both condi­
ways, named LORD MicroStrain® LXRS® communication protocol, tions than other values of j, which is also confirmed by the lowest
allows lossless data throughput, ±32 μs node-to-node synchronization, p-value (value estimated by KWM) of 7.4 × 10− 47, which indicates that
and high-speed sampling. The nodes can be easily programed and this j value is the most reliable value for analyzing the vibration signals
configured for data logging by using the corresponding software (Node with the aim of identifying the bridge condition.
Commander®). Lastly, the web-based interface called SensorCloud™ Despite that the j value of 7 can be useful for estimating the health
makes efficient the analysis of large amounts of data (gigabytes) from condition of the cable-stayed bridge, an overlap between both condi­
remote networks. tions is obtained; hence, a new DI denoted by Eq. (4) is formulated/
The accelerometers were oriented in vertical direction for the RPB generated in order to get maximum discrimination between both con­
monitoring (see Fig. 6a) and positioned in the SHs 3, 4, 5, and 6, which ditions of the bridge. Fig. 10 illustrates the results obtained by applying
are indicated by blue circles in Fig. 6b. In this regard, the bridge was the proposed DI. Based on this figure, it can be observed that the new DI
firstly monitored during a time window of 1800 s with 64 Hz of sampling allows a more distinct separation between two classes, healthy (HLT)
frequency (this range allows measuring the relevant information of the and damaged (cable loss - DMG). In particular, when the bridge is
healthy, the DI value is less than 0.5. This value is selected as the
threshold value to separate both classes as shown in Fig. 10 through a
red line. The obtained results indicate that HI values of the measured
vibration signals exhibit a close range for the HLT condition; on the
contrary, if the bridge is damaged (cable loss) the HI values decrease and
the estimated DI values become larger than the threshold. Table 1 pre­
sents the confusion matrix obtained for the established threshold ac­
cording to an if-else rule, where it is possible to observe that 93.33% of
accuracy for determining the bridge condition is obtained (28 tests for
each condition are correctly detected). If a damage scenario is deter­
Fig. 3. Analyzed cable-stayed bridge. mined, the stage corresponding to the damage location is carried out as

5
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Fig. 4. Rio Papaloapan Bridge layout for general dimensions and identification of semi-harps.

Fig. 5. (a) General design of cables’ anchoring system and (b) its assembly before installation.

Fig. 6. (a) Mounted wireless sensor on the RPB and (b) location of accelerometers (accelerometers: S1, S2, …, S12) employed for monitoring the vibration responses
of the bridge.

Fig. 7. Removed cable (red line) to create the condition of damaged bridge, (a) real condition and (b) schematic. (For interpretation of the references to colour in this
figure legend, the reader is referred to the Web version of this article.)

6
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Fig. 8. Vibrational responses acquired by twelve wireless accelerometers for (a) healthy bridge and (b) damaged bridge (cable loss).

mentioned in section 2.4. that the HI values decrease with a bigger change in sensor 5, which is
In order to have a trained MLP-NN and due to the limited data, expected since it corresponds with the damage location. It is worth
synthetic data are generated from the real data distribution depicted in noting that the data presented in Fig. 11 are an ungrouped version of
Fig. 11, i.e., random HI values that satisfy the boxplot for each sensor are data presented in Fig. 9d. For synthetic damage condition data, two
generated for both HLT condition and DMG condition, where is observed steps, i.e., generation of random HI values and circular shifting, are

7
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Fig. 9. HI values estimated by a j-value of (a) 1, (b) 3, (c) 7, and (d) 10.

values. For the validation of both the HLT condition and the DMG
condition in sensor 5, the 30 real HI values for each condition are used
(see the dataset presented in Fig. 12).
Once the dataset is built, the following MLP-NN architecture is
proposed: 12 input neurons (i.e., 12 HI values obtained from the vi­
bration signals measured by the 12 sensors), 50 hidden neurons, and 12
output neurons (i.e., one neuron for each possible damage location). The
quantity of neurons in the hidden layer is obtained by a process of trial-
and-error, where the goal is to obtain the best accuracy. In all the neu­
rons, the sigmoid symmetric function is utilized as activation function.
As previously mentioned, the training is carried out with the 70% of the
available data (i.e., 910 HI values) and a k-fold cross validation of 10.
The validation results are shown in Table 2 as a confusion matrix. In
quantitative terms, the metrics for DZ5 (i.e., damage location) are: 1 of
accuracy, 1 of precision, 1 of recall, and 1 of F1-score, i.e., perfect
metrics. These high metrics demonstrate the effectiveness of the pro­
posal. It is worth noting that the MLP-NN has a perfect score for damage
location, but its global effectiveness is determined for the damage
detection stage, i.e., if the damage scenario is not detected, the stage of
Fig. 10. Distribution of DI values estimated according to the HI values using j
damage location will not be carried out; therefore, the 28 tests detected
= 7 for evaluating the bridge condition.
with damage (93.33% of accuracy according to Table 1) present 100% of
accuracy to locate the damage on the bridge since in all the cases its
Table 1 location is correctly determined.
Results obtained by evaluating the bridge condition (confusion matrix).
Healthy Damaged Accuracy 4.1. Discussion
Healthy 28 2 93.33%
Damaged 2 28 93.33% In the literature, different works related to detection and location of
damage in bridges have been proposed. Table 3 summaries the main
features of similar works and the proposal. As can be observed, all the
carried out in order to represent the scenarios where the cable loss ap­
works presented carry out the detection and location of a damage con­
pears in other location; thus, the structure of the dataset presented in
dition; however, in some of them, i.e., (Ni et al., 2008; An et al., 2016;
Fig. 2a is obtained. 100 HI values for each condition (i.e., DZ1, DZ2, DZ3,
Tan et al., 2022), the accuracy is not reported or analyzed, which can
DZ4, DZ5, DZ6, DZ7, DZ8, DZ9, DZ10, DZ11 DZ12) are synthetically
compromise negatively their reliability. Although a strict comparison in
generated for training (70%) and validation (30%), resulting in 1200
terms of accuracy cannot be carried out because the conditions are

8
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Fig. 11. Real data distribution for the HI values.

degraded under the scenarios previously mentioned since its accuracy


for both detection and location of damage depends strongly on the
number and position of the accelerometers. A small number of accel­
erometers reduces both the capability of sensing vibrations associated to
the damage throughout the bridge and the resolution to locate the
damage, mainly considering, for instance, a bridge of 407.2 m of length
as the one analyzed in this work.
It is evident that real experimentation is fundamental to validate and
improve structural health monitoring schemes; however, the analysis of
real-life bridges under healthy and damaged conditions is not an easy
task in terms of human, infrastructure, and economical resources. In this
regard, the development and application of FEM-based models have to
Fig. 12. Dataset for training and validation.
be explored as in (Ni et al., 2008)- (Quintana-Rodriguez et al., 2014) in
order to study and generalize the methodology in other structure
totally different, it is important to note that the highest value, i.e.,
conditions.
97.00%, is presented in (Meng et al., 2018), whereas the proposal and
(Quintana-Rodriguez et al., 2014) provide 93.33% and 90.00% of
5. Conclusions
effectiveness, respectively, where the primary advantage of the proposal
is that none domain transformation is required, i.e., the HI is computed
A novel approach based on the homogeneity method to detect and
in time domain, implying a lower computational cost. This fact is
locate lost cables in a cable-stayed bridge subjected to dynamic excita­
fundamental for this work since in future research the goal is to integrate
tions (ambient vibrations) is introduced in this article. In particular, the
to the accelerometers the low-cost hardware solution for implementing
proposal analyses/examines the bridge vibrational responses through
the in-situ HI estimation before the wireless transmission, thus reducing
the HI in order to characterize vibrational changes with the capability of
transmission and storage costs. In addition, the proposal has the
evaluating the structural condition of the bridge, where the KWM is
advantage of locating the damage by means of a conventional and
utilized to encounter which HI j-value calculated for all sensors gener­
low-complex MLP-NN in an automatic way, making possible the com­
ates the most useful information to evaluate the bridge state. The
plete methodology implementation in a future hardware solution based
selected HI j-value is employed to develop a new DI which considers the
on a low-end processor.
HI values obtained by all sensors for establishing, in an automatic way,
Although promising experimental results have been obtained for this
the bridge state. Likewise, the MLP-NN is utilized for determining which
particular scenario, further experimental and analytic research has to be
zone presents the damage. For defining the effectiveness of the proposal,
explored. Regarding experimental research, study cases that consider
the vibrational responses obtained experimentally from the RPB located
scenarios where i) the location of an accelerometer does not coincide or
in Veracruz, Mexico, are studied. The results prove that HI coupled with
is far away from the damage location, ii) the cable is not completely
the new DI is efficient to determine the bridge condition (damage
removed (i.e., stiffness reduction of the cable), and iii) there is more than
detection) because 93.33% of accuracy is reached, where if a bridge a
one cable removed, are necessary to improve the robustness of the
damage condition is determined, its location is defined with 100% of
proposal. At this moment, the performance of the proposal will be
accuracy, i.e., all detected damage conditions are correctly located on

Table 2
Confusion matrix for damage location.
True class

Predicted class DZ1 DZ2 DZ3 DZ4 DZ5 DZ6 DZ7 DZ8 DZ9 DZ10 DZ11 DZ12
DZ1 28 0 0 0 0 0 0 0 0 0 0 0
DZ2 0 28 0 0 0 0 0 0 0 0 0 0
DZ3 0 0 28 0 0 0 0 0 0 0 0 0
DZ4 0 0 0 28 0 0 0 0 0 0 0 0
DZ5 0 0 0 0 28 0 0 0 0 0 0 0
DZ6 0 0 0 0 0 28 0 0 0 0 0 0
DZ7 0 0 0 0 0 0 28 0 0 0 0 0
DZ8 0 0 0 0 0 0 0 28 0 0 0 0
DZ9 0 0 0 0 0 0 0 0 28 0 0 0
DZ10 0 0 0 0 0 0 0 0 0 28 0 0
DZ11 0 0 0 0 0 0 0 0 0 0 28 0
DZ12 0 0 0 0 0 0 0 0 0 0 0 28

9
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Table 3 with or is far from the damage location, 2) the cable is gradually losing
Features of the proposed method and other works for structural monitoring of its stiffness, rather than being entirely removed, 3) multiple cables have
bridges. been removed, and 4) other kind of damages are present. These studies
Work Proposed methodology Structure Accuracy will allow for a better understanding of the proposed methodology’s
condition (%) efficacy under different conditions, especially when the availability of
Ni et al., (Ni et al., 2008) * FEM is utilized to Detection/ Not experimental research is limited.
obtain the bridge modal Location reported
flexibility matrices.
Declaration of competing interest
* Changes in the
flexibility are
considered to define the The authors declare that they have no known competing financial
damage condition and interests or personal relationships that could have appeared to influence
its location. the work reported in this paper.
Quintana-Rodriguez * FEM is used to obtain Detection/ 90.00
et al., ( the bridge model. Location
Quintana-Rodriguez * The dynamic response Data availability
et al., 2014) is experimentally
acquired. Data will be made available on request.
* Function minimization
to determine structural
changes between the Acknowledgments
experimental and FEM
data. This work was supported by the “Facultad de Ingeniería-Problemas
An et al., (An et al., * Frequency variation is Detection/ Not Nacionales 2021 No. project 202112” and the National Council of Sci­
2016) used as damage Location reported
ence and Technology (CONACYT - Mexico) through the project 34/2018
indicator to assess the
cable condition. of the Program “Investigadoras e Investigadores por México” del CON­
* A strategy based on the ACYT (Cátedras CONACYT).
implementation of a
steel bar to amplify the
References
change of frequency
due to the damage
Abiodun, O.I., Jantan, A., Omolara, A.E., Dada, K.V., Umar, A.M., Linus, O.U.,
condition is proposed.
Kazaure, A.A., Gana, U., Kiru, M.U., 2019. Comprehensive review of artificial neural
Meng et al., (Meng et al., * EMD is used to Detection/ 97.00
network applications to pattern recognition. IEEE Access 7, 158820–158846.
2018) decompose the Location
A.F.P, 2018. Why the Morandi Bridge Collapsed in Genoa, Italy? El Universal (Mexico).
acquired signals. Amezquita-Sanchez, J.P., Garcia-Perez, A., Romero-Troncoso, R., Osornio-Rios, R.,
* Hilbert transform is Herrera-Ruiz, G., 2013. High-resolution spectral-analysis for identifying the natural
applied to identify modes of a truss-type structure by means of vibrations. J. Vib. Control 19 (16),
relevant features. 2347–2356.
* Marginal spectra of Amezquita-Sanchez, J.P., Valtierra-Rodriguez, M., Adeli, H., Perez-Ramirez, C.A., 2018.
both conditions A novel wavelet transform-homogeneity model for sudden cardiac death prediction
(damage and healthy) using ECG signals. J. Med. Syst. 42 (10), 1–15.
are utilized to assess the An, Y., Zhong, Y., Tan, Y., Ou, J., 2016. Experimental and numerical studies on a test
damage. method for damage diagnosis of stay cables. Adv. Struct. Eng. 20 (2), 1–12.
Tan et al., (Tan et al., * Numerical model of a Detection/ Not Barbosh, M., Singh, P., Sadhu, A., 2020. Empirical mode decomposition and its variants:
a review with applications in structural health monitoring. Smart Mater. Struct. 29
2022) bridge. Location reported
(1), 1–45.
* Vibration responses
Bekaddour, F., Amine, C.M., 2015. IHBA: an improved homogeneity-based algorithm for
obtained from a moving
data classification. In: Proceedings of the 5th International Conference on Computer
vehicle. Science and its Applications, pp. 129–140.
* Wavelet energy is Bueno-Lopez, M., Giraldo, E., Molinas, M., Bjarte-Fosso, O., 2019. The mode mixing
calculated. problem and its influence in the neural activity reconstruction. Int. J. Comput. Sci.
* Identification of Appl. 46 (3), 1–11.
damage based on a Carden, E.P., Brownjohn, J.M., 2008. ARMA modelled time-series classification for
proposed index. structural health monitoring of civil infrastructure. Mech. Syst. Signal Process. 22
This work * A novel damage Detection/ 93.33 (2), 295–314.
indicator based on the Location Chen, J., 2009. Application of empirical mode decomposition in structural health
homogeneity value for monitoring: some experience. Adv. Adapt. Data Anal. 1 (4), 601–621.
damage detection. Cheng, H.P., Ni, Y.Q., 2018. Structural Health Monitoring for Large Civil Engineering
* MLP-NN used for Structures, first ed. Wiley, USA.
Civera, M., Surace, C.A., 2021. Comparative analysis of signal decomposition techniques
damage location on the
for structural health monitoring on an experimental benchmark. Sensors 21 (5),
bridge.
1–33.
Davey, M., Wald, M.L., 2007. Potential Flaw Is Found in Design of Fallen Bridge. The
New York Times, USA.
the bridge. Fan, W., Qiao, P., 2010. Vibration-based damage identification methods: a review and
In summary, the proposal demonstrated to be a reliable tool to detect comparative study. Struct. Health Monit. 9 (3), 83–111.
Farrar, C.R., Lieven, N.A., 2007. Damage prognosis: the future of structural health
and locate a cable loss into the RPB through wireless accelerometers, monitoring. Philos. Trans. Royal Soc. 365 (1851), 623–632.
which is a desired feature in a SHM scheme since it permits decreasing Farrar, C.R., Worden, K., 2007. An introduction to structural health monitoring. Philos.
examination time, mainly in large civil infrastructures as the examined Trans. Royal Soc. 365 (1851), 303–315.
Gao, Y., Mosalam, K.M., Chen, Y., Wang, W., Chen, Y., 2021. Auto-regressive integrated
one. Furthermore, it should be noted that the proposal can be deemed as
moving-average machine learning for damage identification of steel frames. Appl.
a solution of low complexity, which will permit the low-end processor Sci. 11, 1–27.
implementation for evaluating the bridge state of constant manner. Glišić, B., Inaudi, D., 2007. Fibre Optic Methods for Structural Health Monitoring, first
ed. Wiley, USA.
Although the experimental results appear promising, it is imperative to
Gul, M., Catbas, F.N., 2009. Statistical pattern recognition for structural health
conduct additional experimental and analytical research (e.g., devel­ monitoring using time series modeling: theory and experimental verifications. Mech.
opment of FEM models) to strengthen the proposed methodology’s Syst. Signal Process. 23 (7), 2192–2204.
robustness. For example, it is essential to investigate various scenarios, Jiang, X., Adeli, H., 2007. Pseudospectra, MUSIC, and dynamic wavelet neural network
for damage detection of highrise buildings. Int. J. Numer. Methods Eng. 71 (5),
such as those where: 1) the accelerometer’s placement does not coincide 606–629.

10
M. Valtierra-Rodriguez et al. Developments in the Built Environment 14 (2023) 100166

Kankanamge, Y., Hu, Y., Shao, X., 2020. Application of wavelet transform in structural Pakyari, R., Balakrishnan, N., 2012. A general purpose approximate goodness-of-fit test
health monitoring. Earthq. Eng. Vib. 19 (1), 515–532. for progressively type-II censored data. IEEE Trans. Reliab. 61 (1), 238–244.
Kruskal, W.H., Wallis, W.A., 1952. Use of ranks in one-criterion variance analysis. J. Am. Park, H.S., Kim, J., Oh, B.K., 2019. Model updating method for damage detection of
Stat. Assoc. 47 (260), 583–621. building structures under ambient excitation using modal participation ratio.
Kullaa, J., 2010. Vibration-based structural health monitoring under variable Measurement 133 (1), 251–261.
environmental or operational conditions. In: Deraemaeker, A., Worden, K. (Eds.), Perez-Ramirez, C.A., Machorro-Lopez, J.M., Valtierra-Rodriguez, M., Amezquita-
New Trends in Vibration Based Structural Health Monitoring. Springer, New York, Sanchez, J.P., Garcia-Perez, A., Camarena-Martinez, D., Romero-Troncoso, R.J.,
pp. 107–181. 2020. Location of multiple damage types in a truss-type structure using multiple
Kutner, M.H., 2005. Applied Linear Statistical Models, fifth ed. McGraw-Hill Irwin, USA. signal classification method and vibration signals. Mathematics 8 (932), 1–16.
Lee, S., Oh, H.J., Cho, B.H., 2022. Decision-making process for maintenance of concrete Quintana-Rodriguez, J.A., Carrion-Viramontes, F.J., Crespo-Sanchez, S., 2014. Damage
bridge deck with asphalt overlays using automated digital road scanner. Dev. Built detection on a cable stayed bridge using wave propagation analysis. In: Proceedings
Env. 12, 100103. of the 7th European Workshop on Structural Health Monitoring, pp. 2052–2059.
Liew, K.M., Wang, Q., 1998. Application of wavelet theory for crack identification in Razo-Hernandez, J.R., Valtierra-Rodriguez, M., Amezquita-Sanchez, J.P., Granados-
structures. J. Eng. Mech. 124 (2), 152–157. Lieberman, D., Gomez-Aguilar, J.F., Rangel-Magdaleno, J.J., 2018. Homogeneity-
Limongelli, M.P., 2019. Vibration-based structural health monitoring: challenges and PMU-based method for detection and classification of power quality disturbances.
opportunities. In: Zingoni, A. (Ed.), Advances in Engineering Materials, Structures Electronics 7 (12), 1–13.
and Systems: Innovations, Mechanics and Applications. Taylor & Francis, London, Reimann, M., 2017. The Deadliest Bridge Disaster in US History Was Caused by a Tiny
pp. 1999–2004. Crack Just Three Millimeters Deep. Timeline, USA.
Lizarraga-Morales, R.A., Rodriguez-Donate, C., Cabal-Yepez, E., Lopez-Ramirez, M., Ruocci, G., Cumunel, G., Le, T., Argoul, P., Point, N., Dieng, L., 2014. Damage assessment
Ledesma-Carrillo, L.M., Ferrucho-Alvarez, E.R., 2017. Novel FPGA-based of pre-stressed structures: a SVD-based approach to deal with time-varying loading.
methodology for early broken rotor bar detection and classification through Mech. Syst. Signal Process. 47 (1–2), 50–65.
homogeneity estimation. IEEE Trans. Instrum. Meas. 66 (7), 1760–1769. Swenson, D.V., Ingraffea, A.R., 1991. The collapse of the Schoharie Creek Bridge: a case
Machorro-Lopez, J.M., Amezquita-Sanchez, J.P., Valtierra-Rodriguez, M., Carrion- study in concrete fracture mechanics. Int. J. Fract. 51 (1), 73–92.
Viramontes, F.J., Quintana-Rodriguez, J.A., Valenzuela-Delgado, J.I., 2021. Wavelet Tan, C., Zhao, H., Uddin, N., Yan, B.A., 2022. Fast wavelet-based bridge condition
energy accumulation method applied on the Rio Papaloapan Bridge for damage assessment approach using only moving vehicle measurements. Appl. Sci. 12 (1),
identification. Mathematics 9 (422), 1–30. 1–18.
Martinez-Herrera, A.L., Ferrucho-Alvarez, E.R., Ledesma-Carrillo, L.M., Mata-Chavez, R. Tee, K.F., 2018. Time series analysis for vibration-based structural health monitoring: a
I., Lopez-Ramirez, M., Cabal-Yepez, E., 2022. Multiple fault detection in induction review. Symmetry 1 (1), 1–19.
motors through homogeneity and kurtosis computation. Electronics 15 (1), 1–12. Unser, M., 1986. Sum and difference histograms for texture classification. IEEE Trans.
Meng, F., Mokrani, B., Alaluf, D., Yu, J., Preumont, A., 2018. Damage detection in active Pattern Anal. Mach. Intell. 1, 118–125.
suspension bridges: an experimental investigation. Sensors 18 (3002), 1–20. Yañez-Borjas, J.J., Machorro-Lopez, J.M., Camarena-Martinez, D., Valtierra-
Mousavi, M., Holloway, D., Olivier, J.C., Gandomi, A.H., 2021. Beam damage detection Rodriguez, M., Amezquita-Sanchez, J.P., Carrion-Viramontes, F.J., Quintana-
using synchronisation of peaks in instantaneous frequency and amplitude of Rodriguez, J.A., 2021. A new damage index based on statistical features, PCA, and
vibration data. Measurement 168 (1), 1–40. Mahalanobis distance for detecting and locating cables loss in a cable-stayed bridge.
Ni, Y.Q., Zhou, H.F., Chan, K.C., Ko, J.M., 2008. Modal flexibility analysis of cable-stayed Int. J. Struct. Stabil. Dynam. 21 (9), 1–29.
Ting Kau bridge for damage identification. Comput. Aided Civ. Infrastruct. Eng. 23
(1), 223–236.

11

You might also like