Feature Extraction Methods for Neural Networks in the

Classification of Structural Health Anomalies

Natasha Hamilton

, Jim Harkin

, Liam McDaid

, Junxiu Liu

and Eoghan Furey

School of Computing, Engineering and Intelligent Systems, Ulster University, Northland Road, Derry, U.K.

Department of Computing, Atlantic Technological University, Port Road, Donegal, Ireland

Keywords: Structural Health Monitoring, Feature Extraction, Spiking Neural Networks, Classification.

Abstract: Failure of large complex structures such as buildings and bridges can have monumental repercussions such

as human mortality, environmental destruction and economic consequences. It is therefore paramount that

detection of structural damage or anomalies are identified and managed early. This highlights the need to

develop automated Structural Health Monitoring (SHM) systems that can continuously allow the safety status

of structures to be determined, even in the worst and most isolated conditions, to ultimately help prevent

destruction and save lives. Signal processing is a crucial step to detecting structural anomalies and recent

work demonstrates the opportunities for neural networks, however the encoding of data for SHM requires the

extraction of features due to often, noisy data. This paper focuses on feature extraction methods for artificial

neural networks (ANNs) and spiking neural networks (SNNs) and aims to identify bespoke features which

enable SNNs to encode data and perform the classification of anomalies. Results show that extraction of

particular features in large real-world applications improve the classification accuracy of SNNs.

1 INTRODUCTION

Large man-made civil infrastructures exercise an

important role in both the societal and economical

evolution of the modern world (Khemapech,

Sansrimahachai & Toahchoodee, 2016; Song et al,

2020). Structures such as bridges, tunnels and

buildings are used on a daily basis by billions of

people worldwide, to complete day-to-day activities

(Khemapech, Sansrimahachai & Toahchoodee,

2016). With this in mind it is critical that complex

structures such as these, are continually fit for their

intended purpose and are safe for human use (Ibrahim

et al, 2020). This is a challenging task as throughout

their operational lifetime, artificial structures are

highly vulnerable to damage (Li et al, 2015).

Exposure to a number of environmental,

anthropogenic and operational factors can all

contribute to causing structural deterioration

(Abdeljaber et al, 2017). There are many different

https://orcid.org/0000-0001-7699-1216

https://orcid.org/0000-0001-7484-8205

https://orcid.org/0000-0002-1197-4375

https://orcid.org/0000-0002-9790-1571

https://orcid.org/0000-0002-6697-1462

types of damage that can surface, for example in the

forms of corrosion, erosion, degradation or decay, all

of which have the potential to cause structural

collapse and require continuous monitoring

(Abdeljaber et al, 2017). Areas that are incredibly

difficult to access or that are susceptive to natural

disasters like landslides, earthquakes or forest fires

are often affected by such catastrophic devastation

(Moaveni et al, 2011). Disasters such as these can

occur without warning so preparation is crucial,

having functional and well-maintained infrastructure

is extremely important, as it will reduce the potential

aftermath of future disasters (Pang et al, 2020).

Traditionally, the severity of damage to a structure

is visually assessed by experienced human inspectors,

who physically examine any structurally unsound

sites (Pang et al, 2020). Visual analysis, despite the

extensive efforts of inspectors, experience a number

of challenges; restricted access to damaged locations,

lengthy inspection completion times and regular

514

Hamilton, N., Harkin, J., McDaid, L., Liu, J. and Furey, E.

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies.

DOI: 10.5220/0012184800003595

In Proceedings of the 15th International Joint Conference on Computational Intelligence (IJCCI 2023), pages 514-523

ISBN: 978-989-758-674-3; ISSN: 2184-3236

manual structural maintenance assessments

(Hernandez, Roohi, & Rosowsky, 2018).

Additionally, localized experimental fault

detection techniques such as radiographs, thermal

field methods and acoustic or ultrasonic approaches

have also, been used to identify structural damage

(Doebling et al, 1996). However, the issue with these

methods is that the damaged areas must be known and

accessible for inspection prior to experimental

analysis (Doebling et al, 1996). These limitations

highlight the need to computerise this monitoring

process to make identifying, locating and determining

damage more efficient and accurate (Song et al,

2020).

This has led to the need to develop automated

SHM and damage identification systems that can

detect and monitor infrastructural damage, without

human interaction (Anton, Inman & Park, 2009;

Moaveni et al, 2010). These physical SHM systems

need to be accurate and efficient whilst remaining

operational for extended periods, for example, in

buildings, and concealed in concrete infrastructures

(Abdo, 2014; Yu, Wang, & Meng, 2005). A key

challenge that needs to be considered is how to

effectively manage and process large amounts of raw

data obtained from these systems whilst still being

able to classify structural damage correctly and

efficiently. This, therefore, establishes the focus for

this paper; to investigate the extraction of specific

features from large real-world datasets, in order to

achieve the highest degree of accuracy possible when

applied to brain-inspired solutions.

The remainder of this paper is organised with

section 2 outlining an overview of SHM, neural

networks and the key challenges. Section 3 defines

the selected SHM dataset/application and the analysis

of various feature extraction techniques. Section 4

reports on the accuracy evaluation of both ANN and

SNN networks based on bespoke extracted features.

Finally, section 5 discusses future work and provides

a conclusion.

2 BACKGROUND

It is inevitable that structures will degrade over time

due to a number of factors, including frequent use and

environmental causes like soil erosion, flooding or

unexpected anomalies like earthquakes, landslides or

forest fires (De La Torre et al, 2020). It is therefore

paramount, for both safety and financial reasons to

monitor large complex infrastructures such as

buildings, bridges, dams and railroads, on a regular

basis (Nuhu et al, 2020). SHM is an engineering field

that focuses on developing damage identification

systems that can monitor and evaluate the condition

and stability of man-made structures (Crémona, 2016;

Semperlotti, 2009). The techniques used are designed

to enable early damage detection, allowing preventive

measures to be implemented to avoid structural

failure, such as required maintenance and structural

reinforcement (Couture, 2013).

SHM has progressed rapidly in recent years, due

to the evolution of sensor networks, data processing

and information management (Li et al, 2015). This

automation has led to the development of increased

precision and financially feasible data acquisition

systems, as well as rapid growth in dataset size

(Crémona, 2016). There are, however, still challenges

that need to be addressed.

2.1 Structural Health Monitoring

(SHM) Techniques

To achieve a high level of accuracy and reliability,

SHM systems need to have a well-designed damage

classification framework, that enables structural

damage to be detected (Ying et al, 2013). Figure 1

shows the process of damage identification is

comprised of four core stages. The stages include: 1)

signal monitoring, 2) signal processing, 3) feature

extraction and 4) classification (Amezquita-Sanchez

& Adeli, 2015; Goyal & Pabla, 2015).

Data is obtained from a sensor network and

digitised during the signal monitoring stage. Signal

processing methods such as Fourier transforms,

Hilbert-Huang transforms, statistical time series

models, Wavelet transforms and Cohen’s class are

then used, to examine the data in order to extract,

determine and categorise core features (Goyal &

Pabla, 2015). Feature extraction, for example the

orthogonal decomposition technique, is then carried

out using measured data, to detect anomalous

information with the goal of revealing non-obvious

damage states (Eftekhar Azam, Rageh & Linzell,

2018; Overbey, 2008). Feature extraction, is

therefore, a key step in the damage identification

process. A number of techniques have been used

previously during the classification stage to identify

structural damage accurately and correctly. These

methods include clustering algorithms, specifically

K-means (KM) clustering, Support Vector Machines

(SVM), Artificial Neural Networks (ANN), Spiking

Neural Networks (SNN) and Hybrid Classifiers

(Goyal & Pabla, 2015).

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies

515

Figure 1: Four main damage identification stages within a

SHM system.

2.2 Artificial and Spiking Neural

Networks

ANNs consist of a number of vastly connected

processing nodes (neurons) that operate concurrently

(Keller, Liu & Fogel, 2016). These networks learn as

a result of training which is performed using datasets

where the input and output data is known (Keller, Liu

& Fogel, 2016). This data is used to train the neural

network and modulate the synaptic weights between

neurons (Keller, Liu & Fogel, 2016). ANNs are a very

beneficial tool as they have the functionality to

extract trends from input data (Notley & Magdon-

Ismail, 2018). However, despite this, factors such as

power utilisation and the expense of implementing

them in hardware as edge computing devices,

presently do not meet practicality requirements for

anomaly detection in real world, always-on

applications (Pang et al, 2020).

The original concept of ANNs has progressed

rapidly to develop generations of ANNs, which mimic

more closely the biological principles for learning and

fault tolerance (Pang et al, 2020). SNNs are considered

to be the third generation of ANNs (Paugam-Moisy &

Bohte, 2012). Derived from neuroscience

advancements and brain inspired natural computing,

these networks use an adapted version of the spike

timing of neurons to encode and process information

(Liu et al, 2017). Similar to biological NN, SNNs

enable communication through incorporating electrical

pulses (spikes) (Zhang, Gu, & Pan, 2018) with the

concept of time illustrated in Figure 2 In SNNs,

information is communicated via the timing between

spike events or frequencies. The integration of multiple

frequencies enables a spiking neuron to aggregate the

frequencies to reflect a membrane voltage increase

within the neuron. When a threshold is exceeded, the

neuron produces a single spike output. This process

enables temporal patterns to be identified via training

of synaptic weights which impact on the contribution

to the neuron’s membrane voltage.

There are a number of models that have been

developed to determine the impact of action potential

spikes on selected neurons, these include the

Hodgkin-Huxley (HH) model, the leaky integrate-

and-fire (LIF) model and the adaptive exponential

integrate-and-fire (AdExIF) model (Paugam-Moisy

& Bohte, 2012). The LIF model is the least

computationally expensive, in comparison to the HH

model which is deemed the most expensive (Paugam-

Moisy & Bohte, 2012).

Figure 2: Communicating information in SNNs.

Research to date has established the ultra-low

power capability of SNNs in hardware due to the fact

that energy is only consumed when an input spike is

received and processed, resulting in an overall saving

of power (Zhang, Gu, & Pan, 2018). Currently SNNs

have shown benefit in SHM system development, as

compared to ANNs, as the hardware expense is more

cost-effective and power efficiency is improved

(Pang et al, 2020).

The key challenge is the extraction of features

from the SHM data and the encoding of identified

features which improve the accuracy of the network.

This is challenging as the data from sensors (e.g.

accelerometers) is highly variable.

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

516

3 DATASET & FEATURE

ANALYSIS

The identification of suitable benchmark data was a

key step, and the Qatar University Grandstand

Simulator (QUGS) (Avci, 2018) was selected due to

the availability of a full range of structural failures

across a known structure with labelled data points.

This dataset has also been used in several other works

(Abdeljaber et al, 2016; Abdeljaber et al, 2017; Avci

et al, 2018; Kiranyaz et al, 2021) and will form the

basis for benchmarking of performance.

The data was originally used by Abdeljaber et al.

(Abdeljaber et al, 2017) to develop a 1D

convolutional neural network (CNN) for vibration

based SHM, with the primary intention of developing

damage identification approaches that can efficiently

monitor present-day infrastructure (Kiranyaz et al,

2021). The simulator is situated in a laboratory

environment and is reported as the biggest stadium

framework constructed in a controlled environment

(Abdeljaber et al, 2016), shown in Figure 3.

Figure 3: QUGS Sensor Point Locations Identified. Source:

Adapted from (Abdeljaber et al, 2017).

Devised to hold 30 observers, the main hot-rolled

steel shell is 4.2m x 4.2m in size (Abdeljaber et al,

2017). The QUGS has a total of 30 structural joints

(shown as 1A to 6E in Figure 3), in which 30

accelerometers are used to measure the structural

vibrational response (Kiranyaz et al, 2021). The steel

frame is equipped with 27 PCB model 393B04

accelerometers and 3 B&K model 8344

accelerometers (Abdeljaber et al, 2017). Vibration

was applied to the structure through the use of a

modal shaker, that used a SmartAmp power amplifier,

to implement the signal to the shaker (Abdeljaber et

al, 2017). Finally, the production of the shaker input

and recording of the acceleration output are achieved

through using two 16-channel data acquisition

instruments (Abdeljaber et al, 2017).

Structural damage is injected by slackening the

bolts at a specific joint, which is a very slight

alteration to the structure’s rotational stiffness

(Kiranyaz et al, 2021), as displayed in Figure 4.

Figure 4: Demonstration of How Structural Damage is

Artificially Applied in the QUGS. Source: Adapted from

(Kiranyaz et al, 2021).

There were 31 damage tests implemented; 1

undamaged (healthy) case for benchmarking

purposes and 30 damaged cases to simulate structural

anomalies (Avci et al, 2018). Each scenario was

recorded for 256 seconds, at a sampling frequency of

1,024Hz, resulting in a total of 262,144 samples per

joint per test (Avci et al, 2018). This, therefore, results

in a total of 243,793,920 samples for the entire

dataset.

The particular reasons outlined demonstrate that

the QUGS dataset provides an ideal range of

anomalies for data training and evaluation purposes

of the neural networks.

3.1 Raw Data

The QUGS dataset has a significantly large number

of raw data samples, 243,793,920 in total. This data

when graphically displayed is extremely noisy and

difficult to discern any visual trend or features, due to

high frequency sampling. It is therefore very difficult

to distinguish whether a sample is of a damaged or

undamaged state as seen in Figure 5.

Using this data in its raw form will make it

tremendously challenging for any classification

technique to determine the structural state accurately.

Therefore, feature extraction was required to ensure

that any potentially masked damage states reported in

the sensor data are identified and to consolidate to key

element of interest (Amezquita-Sanchez & Adeli,

2015).

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies

517

Hence, feature extraction is a critical step in the

damage identification process.

3.2 Feature Selection

It is important to select features that best represent the

data. Certain features may suit specific real-world

datasets better than others. The choice of damage-

sensitive parameters for the QUGS dataset is based

on multiple different factors such as the data type and

which features will best identify the health status best.

There may be several features that could determine

structural health accurately whilst avoiding the effect

of various environmental and structural conditions

(Pang et al, 2020).

Figure 5: Displays the comparison of the raw noisy data at

joint 1 between a) an undamaged sample and b) a damaged

sample.

There two different types of sensors used in the

data acquisition process that measure bolt vibration

over a period of 256 seconds. This means that time

and frequency domains, due to the nature of the data,

can be used to extract specific features. These features

include mean, standard deviation, variance, energy,

Zero-crossing rate, and Fourier Transforms (Toivola

& Hollmén, 2009). Some features proved better than

others for example, zero-crossing rate only showed

very minor differences between the damaged and

undamaged data, when extracted from the noisy raw

data, as displayed in Figure 6 and therefore was not

the best feature choice.

Figure 6: Displays the comparison of zero-crossing rate at

joint 7 between a) an undamaged sample and b) a damaged

sample.

A number of MATLAB scripts were created to

firstly, extract a selection of features. These features

were then displayed graphically and analysed to

determine if there were any significant differences in

the damaged and undamaged data. This was to

ultimately determine which feature was the best

choice, in aiding damage identification. After this

extensive analysis process, several features: absolute

mean, variance, standard deviation and Fast Fourier

Transforms (FFT) were chosen, as they showed

distinguished profiles between undamaged and

damage data.

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

518

Figure 7: Illustrates the comparison of variance at joint 7

between a) an undamaged sample and b) a damaged sample.

3.3 Fast-Fourier Transform (FFT)

The Fast Fourier Transforms (FFT) proved to be the

superior feature for this particular dataset, as it

showed a considerable difference between the

damaged and undamaged data samples, as illustrated

in Figure 8.

This was identified through developing a

MATLAB script that was able to graphically display

a double-sided magnitude spectrum for each sample

and determine the top three highest magnitudes of

each one. Interestingly the frequency associated with

the third highest magnitude in each comparative

graph showed the largest difference in frequency

between the damaged and undamaged data.

This analysis also uncovered that the frequency

associated with third largest magnitude in the

undamaged joints, surrounding a damaged joint,

showed significant variation. In addition, looking at

these as a collective instead of individually could

also, prove as another technique to aid data

classification, as illustrated in Table 1.

Figure 8: Displays the comparison between Fast-Fourier

Transforms at joint 7 for a) an undamaged sample and b) a

damaged sample.

Table 1: Displays the frequencies associated with the three

largest magnitudes for joints 1B, 2A, 2B, 2C and 3B, for

both the benchmark data (2B-B) and the damaged data (2B-

D) when only the single joint 2B is damaged.

4 NEURAL NETWORK

APPLICATION

ANNs are a well-established technique making an

excellent benchmark for all other future designed

networks. This research aims to validate that an SNN

can make a relatively accurate prediction on real

world data. However, the dataset is extremely noisy

due a to high frequency sampling rate, making it very

challenging to classify. Therefore, it requires pre-

processing in extracting and analysing several

a a

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies

519

features; this establishes the contribution from the

research.

Each stage of this research is depicted in Figure 9,

where the main stages are: 1) Raw data obtained from

the 30 accelerometers, 2) Extraction of features from

the raw data, 3) Feature extracted data is inputted into

neural networks and 4) Output from neural networks

is determined.

Figure 9: Illustrates a summary of the workflow for this

research.

4.1 ANN

An ANN was established, using the feature extracted

data as the input data. This was created to classify

when an anomaly has occurred, i.e. detect if the input

data reflects a healthy or unhealthy structural state.

The Neural Network Toolbox in MATLAB was used

to develop the ANN. The network had 30 neurons in

the input layer and three output states: the three

outputs were predetermined prior to classification;

undamaged, damaged or unclassified.

Figure 10: Presents the confusion matrixes for the ANN for

a) training and b) testing.

The network was trained using 75 percent of the

feature extracted samples and tested on 25 percent of

the feature extracted samples. Both phases achieved

100 percent accuracy, correctly identifying all of the

damaged and undamaged samples, as shown in the

confusion matrix of Figure 10.

These results provide a benchmark to compare

with the SNN accuracy.

4.2 SNN

The QUGS feature extracted data is represented as

numerical numbers. To enable the development of an

SNN the data must first be encoded into spike trains

which represent a frequency, in order to be used as

input data.

Using PyCharm with a python package called

BindsNet, the selection of a new encoding scheme

was required to best reflect the varied data. The data

spanned over a large range of frequencies between

approximately 15Hz – 475Hz. To make this range

smaller a banding system was created, dividing the

data into 10Hz wide bands comprising of 47 in total.

As the band intervals increase in size, so do the length

of the spike trains, adding additional spikes to each

band, making each one larger than previous. The total

spike trains are repeated to provide a 1-second

duration of stimulus, i.e. to achieve an appropriate

length of input stimulus for the SNN.

The SNN consisted of 30 neurons in the input

layer (one neuron per joint in the dataset) and 2 output

LIF neurons: undamaged, damaged. The SNN is a

fully connected network and uses the Spike-Timing

Dependant Plasticity (STDP) learning algorithm.

The network was able to achieve an accuracy level

of 87.5 percent and was able to identify all of the

undamaged samples and majority of the damaged

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

520

samples, when trained on 75 percent and tested on 25

percent of the feature extracted data. An equal amount

of damaged and undamaged samples were used in

both the training/testing groups. Comparative work

has been conducted by Zanatta et al, achieving an 88

percent accuracy level in comparison to the 87.5

percent accuracy level from this research (Zanatta et

al, 2021). However, the comparative network is a

Long Short-Term SNN (LSNN) and is significantly

more complex with recurrent neurons, and also in

neuron density with between 50 to 500 input neurons,

20 recurrent neurons and 2 output neurons (Zanatta et

al, 2021).

The proposed SNN was created to determine if the

QUGS data could be classified correctly, as the vision

for this work is to incorporate a self-repair element,

in the form of an artificial astrocyte cell into the SNN

network, providing the additional capability to

tolerate failure. Rapid decision-making was not the

aim of this research but is ultimately a focus of future

work.

4.3 Comparing SNN Against ANN

The ANN proved to have overall, a better level of

accuracy of 100 percent identifying all of the

damaged and undamaged structural health states

correctly, in comparison to the SNN which classified

87.5 percent of the samples correctly. This is

demonstrated in Table 2.

Table 2: Displays the results comparing the accuracy of the

ANN and SNN, when trained on 75 percent and tested on

25 percent of the data.

The ANN results provide good benchmark data

and was expected to have a higher level of accuracy

compared to the SNN. This is because ANNs are well

established and not overly complex.

However, when considering the application of

SHM require low power and compact edge

computing capabilities, SNNs can achieve must lower

area/power performances than ANN equivalents

(Yang et al. 2021). There is a trade-off between high

accuracy/high-compute overheads and meeting

lower-power budgets but with a reduction in

accuracy.

SNNs are more complex but possess the ability to

incorporate a self-repair element into the network;

ANNs do not have this capability. This network sets

the foundation for future research.

5 CONCLUSION AND FUTURE

WORK

Feature extraction is a key step when developing

structural health applications and working with large

datasets. Based on this work, it is evident that there

are particular features that suit bespoke datasets better

than others. FFT demonstrated to be the superior

feature in the QUGS dataset. This contributed to the

overall accuracy results achieved by the ANN and

SNN, as the input data was more discernible between

damaged and undamaged samples than displayed in

the raw data. This, therefore, aided data classification.

Future work could involve incorporating the use

of ensembles to further improve the accuracy and

performance of the SNN. However, the main goal is

to develop the SNN further by creating an astrocyte-

neuron network (SANN), that can monitor and

classify structural damage as well as realising self-

repairing capabilities (Liu et al 2017). Upon

achieving a good level of accuracy, the intention is to

implement the network in FPGA hardware, where the

systems performance will be benchmarked against

conventional methods and evaluated, in terms of

reliability and accuracy. This should enable large

man-made structures to be monitored for long periods

of time, without human intervention.

REFERENCES

Abdeljaber, O., Avci, O., Kiranyaz, S., Gabbouj, M., &

Inman, D. J. (2017). Real-time vibration-based

structural damage detection using one-dimensional

convolutional neural networks. Journal of Sound and

Vibration, 388, 154–170. https://doi.org/10.1016/j.jsv.

2016.10.043

Abdeljaber, O., Younis, A., Avci, O., Catbas, N., Gul, M.,

Celik, O., & Zhang, H. (2016). Dynamic testing of a

laboratory stadium structure. Geotechnical and

Structural Engineering Congress 2016. https://doi.org/

10.1061/9780784479742.147

Abdo, M. (2014). Structural Health Monitoring, History,

Applications and Future. A Review Book. New York,

NY, USA.

Amezquita-Sanchez, J., & Adeli, H. (2015). Feature

extraction and classification techniques for health

monitoring of structures. Scientia Iranica, 1931–1940.

Anton, S. R., Inman, D. J., & Park, G. (2009). Reference-

free damage detection using instantaneous baseline

measurements. AIAA Journal, 47(8), 1952–1964.

https://doi.org/10.2514/1.43252

Avci, O. (2018). Qatar University Grandstand Simulator

(QUGS). Onur Avci. http://onur-avci.com/benchmark/

qugs/

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies

521

Avci, O., Abdeljaber, O., Kiranyaz, S., Hussein, M., &

Inman, D. J. (2018). Wireless and real-time structural

damage detection: A novel decentralized method for

wireless sensor networks. Journal of Sound and

Vibration, 424, 158–172. https://doi.org/10.1016/

j.jsv.2018.03.008

Couture, Z. (2013). Structural Health Monitoring

(dissertation).

Crémona, C. (2016). Big Data and Structural Health

Monitoring. IABSE Congress, Stockholm 2016:

Challenges in Design and Construction of an

Innovative and Sustainable Built Environment.

https://doi.org/10.2749/stockholm.2016.1793

De La Torre, R. Dll., Pasobillo, G. A., Rebueno, M. F.,

Sunga, D. P., Esguerra, B. J., & Concepcion, R. (2020).

Vibration-based structural health monitoring system for

bridges using ADXL345 accelerometer with MATLAB

standalone application. 2020 IEEE 12th International

Conference on Humanoid, Nanotechnology,

Information Technology, Communication and Control,

Environment, and Management (HNICEM).

https://doi.org/10.1109/hnicem51456.2020.9400068

Doebling, S. W., Farrar, C. R., Prime, M. B., & Shevitz, D.

W. (1996). Damage Identification and Health

Monitoring of Structural and Mechanical Systems from

Changes in Their Vibration Characteristics: A

Literature Review. https://doi.org/10.2172/249299

Eftekhar Azam, S., Rageh, A., & Linzell, D. (2018).

Damage detection in structural systems utilizing

artificial neural networks and proper orthogonal

decomposition. Structural Control and Health

Monitoring, 26(2). https://doi.org/10.1002/stc.2288

Goyal, D., & Pabla, B. S. (2015). The vibration monitoring

methods and signal processing techniques for Structural

Health Monitoring: A Review. Archives of

Computational Methods in Engineering, 23(4), 585–

594. https://doi.org/10.1007/s11831-015-9145-0

Hernandez, E., Roohi, M., & Rosowsky, D. (2018).

Estimation of elementbyelement demandto

capacity ratios in instrumented SMRF buildings using

measured seismic response. Earthquake Engineering &

Structural Dynamics, 47(12), 2561–2578.

https://doi.org/10.1002/eqe.3099

Ibrahim, A., Eltawil, A., Na, Y., & El-Tawil, S. (2020). A

machine learning approach for structural health

monitoring using noisy data sets. IEEE Transactions on

Automation Science and Engineering, 17(2), 900–908.

https://doi.org/10.1109/tase.2019.2950958

Keller, J. M., Liu, D., & Fogel, D. B. (2016). Fundamentals

of Computational Intelligence: Neural Networks, Fuzzy

Systems, and evolutionary computation. Wiley.

Khemapech, I., Sansrimahachai, W., & Toahchoodee, M.

(2016). A real-time health monitoring and warning

system for bridge structures. 2016 IEEE Region 10

Conference (TENCON). https://doi.org/10.1109/ten

con.2016.7848598

Kiranyaz, S., Avci, O., Abdeljaber, O., Ince, T., Gabbouj,

M., & Inman, D. J. (2021). 1D convolutional neural

networks and applications: A survey. Mechanical

Systems and Signal Processing, 151, 107398.

https://doi.org/10.1016/j.ymssp.2020.107398

Li, H.-N., Ren, L., Jia, Z.-G., Yi, T.-H., & Li, D.-S. (2015).

State-of-the-art in structural health monitoring of large

and complex civil infrastructures. Journal of Civil

Structural Health Monitoring, 6(1), 3–16.

https://doi.org/10.1007/s13349-015-0108-9

Liu, J., McDaid, L. J., Harkin, J., Wade, J. J., Karim, S.,

Johnson, A. P., Millard, A. G., Halliday, D. M., Tyrrell,

A. M., & Timmis, J. (2017). Self-repairing learning rule

for spiking astrocyte-neuron networks. Neural

Information Processing, 384–392. https://doi.org/10.10

07/978-3-319-70136-3_41

Moaveni, B., He, X., Conte, J. P., & Restrepo, J. I. (2010).

Damage identification study of a seven-story full-scale

building slice tested on the UCSD-Nees Shake Table.

Structural Safety, 32(5), 347–356. https://doi.org/10.10

16/j.strusafe.2010.03.006

Moaveni, B., He, X., Conte, J. P., Restrepo, J. I., &

Panagiotou, M. (2011). System identification study of a

7-story full-scale building slice tested on the UCSD-

Nees Shake Table. Journal of Structural Engineering,

137(6), 705–717. https://doi.org/10.1061/(asce)st.19

43-541x.0000300

Notley, S., & Magdon-Ismail, M. (2018). Examining the

Use of Neural Networks for Feature Extraction: A

Comparative Analysis Using Deep Learning, Support

Vector Machines and K-Nearest Neighbor Classifier.

Nuhu, B. K., Aliyu, I., Adegboye, M. A., Ryu, J. K.,

Olaniyi, O. M., & Lim, C. G. (2020). Distributed

network-based Structural Health Monitoring Expert

System. Building Research & Information, 49(1), 144–

159. https://doi.org/10.1080/09613218.2020.1854083

Overbey, L. A. (2008). Time series analysis and feature

extraction techniques for Structural Health Monitoring

Applications (dissertation).

Pang, L., Liu, J., Harkin, J., Martin, G., McElholm, M.,

Javed, A., & McDaid, L. (2020). Case study-spiking

neural network hardware system for Structural Health

Monitoring. Sensors, 20(18), 5126. https://doi.org/

10.3390/s20185126

Paugam-Moisy, H., & Bohte, S. (2012). Computing with

spiking neuron networks. Handbook of Natural

Computing, 1–47.

Semperlotti, F. (2009). Structural damage detection via

nonlinear system identification and structural intensity

methods (dissertation).

Song, L., Li, S., Wang, J., Wang, Z., & Zhao, G. (2020).

Research progress on structural damage identification

in civil engineering. 2020 International Conference on

Intelligent Transportation, Big Data & Smart City

(ICITBS)

. https://doi.org/10.1109/icitbs49701.2020.00

076

Toivola, J., & Hollmén, J. (2009). Feature extraction and

selection from vibration measurements for structural

health monitoring. Advances in Intelligent Data

Analysis VIII, 213–224. https://doi.org/10.1007/978-3-

642-03915-7_19

Ying, Y., Garrett, J. H., Oppenheim, I. J., Soibelman, L.,

Harley, J. B., Shi, J., & Jin, Y. (2013). Toward data-

NCTA 2023 - 15th International Conference on Neural Computation Theory and Applications

522

driven structural health monitoring: Application of

machine learning and signal processing to damage

detection. Journal of Computing in Civil Engineering,

27(6), 667–680. https://doi.org/10.1061/(asce)cp.1943-

5487.0000258

Yu, L., Wang, N., & Meng, X. (2005). Real-time forest fire

detection with wireless sensor networks. Proceedings.

2005 International Conference on Wireless

Communications, Networking and Mobile Computing,

2005. https://doi.org/10.1109/wcnm.2005.1544272

Z. Yang, Z. Han, Y. Huang and T. T. Ye (2021). 55nm

CMOS Analog Circuit Implementation of LIF and

STDP Functions for Low-Power SNNs. IEEE/ACM

International Symposium on Low Power Electronics

and Design. pp. 1-6

Zanatta, L., Barchi, F., Burrello, A., Bartolini, A., Brunelli,

D., & Acquaviva, A. (2021). Damage detection in

structural health monitoring with spiking neural

networks. 2021 IEEE International Workshop on

Metrology for Industry 4.0 & IoT

(MetroInd4.0&IoT). https://doi:10.1109/metroind

4.0iot51437.2021.9488476

Zhang, M., Gu, Z., & Pan, G. (2018). A survey of

neuromorphic computing based on spiking neural

networks. Chinese Journal of Electronics, 27(4), 667–

674. https://doi.org/10.1049/cje.2018.05.006

Feature Extraction Methods for Neural Networks in the Classiﬁcation of Structural Health Anomalies

523