A Percussion Method to Detect Erosion of Elbow Using Machine

Learning Algorithms

By

Lan Cao

A thesis submitted to the Department of Mechanical Engineering,

Cullen College of Engineering

in partial fulfillment of the requirements for the degree of

Master of Science

in Mechanical Engineering

Chair of Committee: Gangbing Song

Committee Member: Zheng Chen

Committee Member: Xuemin Chen

Committee Member: Matthew A. Franchek

University of Houston
December 2022
Copyright 2022, Lan Cao
 ACKNOWLEGEMENTS

I would like to express my great appreciation to Dr. Gangbing Song for all of his

supporting during my graduate study at the University of Houston (UH). When I decided

to apply for graduate school back in 2020, I planned to attend an online Master degree

program for Mechanical Engineering at a different university. Dr. Song successfully

persuaded me to apply to the one at the University of Houston and do the thesis option

instead of the full courses option. During the two and half years of my academic and

research study at UH, Dr. Song presented me his previous research on smart materials

and ongoing research on solving mechanical engineering problems with machine learning

methods, which I assumed had only belonged to the computer science applications, such

as YouTube video watching, Google translation, and video gaming. Dr. Song's guidance

really expanded my view on integration of machine learning into many fields, not limited

to mechanical engineering, but also to civil, chemical and medical industry as well. I

graduated from UH with my Bachelor's degree of Mechanical Engineering in 2014, and I

was astonished that the technology had made such a great progress since then.

Throughout my thesis work, Dr. Song gave me close guidance and many suggestions

when I was facing numerous challenges to improve machine learning results.

Also I would like to offer my special thanks to my defense committee members

Dr. Zheng Chen, Dr. Xuemin Chen and Dr. Matthew A. Franchek for taking their time to

read my thesis and serve on my defense committee.

iii
 Last but not least, thanks to Mr. Ji’An Chen, a Ph.D. student in Dr. Song’s Smart

Material and Structure Lab. We did experiment together and he helped to train me on

computer science programming.

iv
 ABSTRACT

Elbows are widely used in many industries, especially in oil and gas industry. The

purpose of elbow is to change the flow direction in pipeline systems. In some severe

applications, elbows are employed to transport abrasive high-pressure multiphase flow

medium. With the increase of the service time, the wall thickness of the elbow will

become thinner due to erosion and wear, which may lead to piercing or bursting of the

high-pressure piping system and cause negative impacts on both the economy and the

environment.

A novel method of using percussion and machine learning to detect the rate of

elbow’s erosion was developed and discussed in this thesis. Three sets of elbow and pipe

assembly were used as test specimens. Then, six different erosion levels were simulated

by grinding off mass from the internal wall of the elbows. The elbow bottom location,

where the simulated erosion was, was tapped to generate the percussion sound, which

was recorded by a smart phone. The power spectral density (PSD) and mel-frequency

cepstral coefficient (MFCC) were employed to extract features from the percussion

sound.

The k-nearest neighbor (KNN), the decision tree (DT), and the support vector

machine (SVM) were implemented with PSD features to learn the training samples and

predict test samples. By using the above three basic machine learning methods, the

experiment achieved an average of 90% accuracy on training data and 80% on testing

data. Then, the recurrent neural network (RNN), a deep learning method, was

v
implemented with MFCC features to learn and train the data. This method achieved

100% accuracy on training data and 97% on testing data. Finally, the unsupervised

clustering algorithms, k-means and Gaussian mixture model (GMM), were implemented

with transformed MFCC features. The accuracy of k-means algorithm varied in a range

from 49% to 68%, while the GMM clustering method achieved an accuracy of 76%.

The results of this work have demonstrated the feasibility of the novel method of

percussion and machine learning to detect the level of erosion of elbow in pipeline.

Compared with the conventional method, the proposed method does not require

installation of sensors or extra signal acquisition instruments. The erosion detection using

percussion and machine learning brings great potential contribution to pipeline operating

safety assurance.

vi
Table Contents

ACKNOWLEGEMENTS .................................................................................................. iii

ABSTRACT .........................................................................................................................v

List of Tables ..................................................................................................................... ix

List of Figures ......................................................................................................................x

1 Introduction ..................................................................................................................1

2 An Overview of Related Work ....................................................................................4

2.1 The basic understanding of elbow erosion and relevant prevailing detection

methods ........................................................................................................................... 4

2.2 Percussion-based detection method ..................................................................... 6

2.3 Sound Feature Selection and Machine learning algorithms ................................. 8

3 Research Methodology and Methods.........................................................................10

3.1 Power spectrum density ..................................................................................... 11

3.2 K-nearest neighbor algorithm............................................................................. 13

3.3 Decision tree algorithm ...................................................................................... 13

3.4 Support vector machine algorithm ..................................................................... 14

vii
 3.5 Mel-frequency cepstral coefficients ................................................................... 15

3.6 Recurrent neural network ................................................................................... 18

3.7 K-means clustering ............................................................................................. 22

3.8 Clustering using Gaussian mixture model ......................................................... 23

4 Experiment Setup and Data Collection ......................................................................26

5 Data Analyses and Results Discussion ......................................................................29

5.1 Erosion detection using basic machine learning algorithms .............................. 29

5.2 Erosion detection using deep learning method .................................................. 33

5.3 Severe erosion detection using clustering method ............................................. 39

6 Conclusion and Future Work .....................................................................................43

6.1 Conclusion .......................................................................................................... 43

6.2 Recommendation for future work ...................................................................... 43

References ..........................................................................................................................45

viii
List of Tables

Table 1: Components to control the cell state and hidden state .........................................20

Table 2: Formulas of components......................................................................................22

Table 3: Arrangement of six classes ..................................................................................28

Table 4: Test Results of basic learning methods ...............................................................32

Table 5: Confusion matrix of RNN method for scenario 1 ...............................................36

Table 6: Confusion matrix of RNN method for scenario 2 ...............................................37

Table 7: Confusion matrix of RNN method for scenario 3 ...............................................37

Table 8: Confusion matrix of RNN method for scenario 4 ...............................................38

Table 9: Test results of deep learning method ..................................................................38

Table 10: Classification results of unsupervised learning methods ...................................42

ix
List of Figures

Figure 1: Solid particle erosion causing leakage in an elbow ..............................................1

Figure 2: Photographs illustrating the elbow section of the pipe where the leak was

found ....................................................................................................................................2

Figure 3: Schematic diagram of elbow erosion by liquid-solid two phase flow .................4

Figure 4: Flow chart of proposed methods in this thesis ...................................................11

Figure 5: Decision tree algorithm chart .............................................................................14

Figure 6: Maximum-margin hyperplane and margins for an SVM [45] ...........................15

Figure 7: Flow chart of recurrent neural network ..............................................................19

Figure 8: Diagram of flow of time series through LSTM layer .........................................19

Figure 9: Diagram of an individual cell .............................................................................21

Figure 10: K-mean clustering data [53] .............................................................................23

Figure 12: Percussion performed on elbow and three sets of elbow and pipe assembly ...26

Figure 13: Four scenarios of experiment ...........................................................................27

Figure 14: The elbows after grinding on the last step ........................................................28

Figure 15: Plot of six erosion rates of power spectrum of signals in frequency domain ...29
x
Figure 16: 3D plot showing scattering data on PSD features (2nd, 5th and 7th)…………..30

Figure 17: 3D plot showing scattering data on PSD features (3rd, 5th, and 9th) .................31

Figure 18: MFCC features extracted from single frame sound signal ...............................33

Figure 19: The training process of RNN for scenario 1.....................................................34

Figure 20: The training process of RNN for scenario 2.....................................................35

Figure 21: The training process of RNN for scenario 3.....................................................35

Figure 22: The training process of RNN for scenario 4.....................................................36

Figure 23: K-means clustering method’s diagram .............................................................39

Figure 24: 3D plot showing results of k-means clustering on selected transformed MFCC

features ...............................................................................................................................40

Figure 25: 3D plot showing results of clustering with GMM on selected transformed

MFCC features ...................................................................................................................41

xi
1 Introduction

Elbows are widely used in many industries, especially in oil and gas industry. The

purpose of elbow is to change the flow direction in pipeline systems. In some severe

applications, elbows employed to transport abrasive high-pressure multiphase flow

medium, such as fracturing fluid [1]. Studies show that the erosion of the elbow is about

50 times more serious than that of straight pipe [2]. With the increase of the service time,

the wall thickness of the elbow will become thinner due to erosion and wear, which may

lead to piercing or bursting of the high-pressure piping system [1]. Figure 1 shows an

example of the failure of an elbow, causing leakage from pipeline [3].

Figure 1: Solid particle erosion causing leakage in an elbow

1
Figure 2: Photographs illustrating the elbow section of the pipe where the leak was found

As another example, Kusmono et al. investigated in his research paper the failure

of an elbow, which had a leakage after two months in service. The details of leakage are

shown in Figure 2 [4]. Figure 2a shows the leak area, which is an elbow of a pipeline.

Figure 2b shows the elbow being cut from pipeline and leakage area was determined.

Figure 2c shows the inner area of elbow and wall thinning area caused by erosion and

corrosion. Any leakage of pipeline would cause an unscheduled shutdown and result in

economic losses. Therefore, it is necessary to find an efficient erosion monitoring method

since elbow erosion monitoring is of great significance to the safety of pipeline systems

and maintenance personnel. Often, the elbow erosion monitoring is based on the

2
detection of the wall thickness and the detailed literature review on the detection method

will be provided in the next section.

This work proposes a new elbow erosion monitoring method, which is the

percussion method with the help of machine learning algorithms. To simulate erosion

degree of elbow, a grinder was used to remove mass from the elbow at the location of the

curved bottom where the erosion would be most severe. The mass reduction of the elbow

was measured by a high-precision electronic scale and recorded. By tapping the curved

bottom of elbow, the sound are collected by a smart phone and pre-processed by power

spectral density (PSD) and mel-frequency cepstral coefficient (MFCC) to extract

features. Machine learning algorithms, including the k-nearest neighbor and recurrent

neural network, are employed to classify the different mass loss. The results show that the

newly proposed method for elbow wall thickness erosion detection is effective.

3
2 An Overview of Related Work

This chapter reviews the related researches on percussion-based method, machine

learning algorithm and their applications. Basic knowledge of elbow erosion and

prevailing detection methods are introduced first.

2.1 The basic understanding of elbow erosion and relevant prevailing detection

methods

When the fluid passes through the elbow, the flow direction changes sharply, and

solid particles impact intensively on local areas of the inner wall at a certain speed and

angle, causing local mass loss and wall thickness thinning [5], as shown in Figure 3.

Figure 3: Schematic diagram of elbow erosion by liquid-solid two phase flow

There are various nondestructive testing methods for pipeline defect detection,

including optical testing method, radiography technology, magnetic flux leakage testing

4
technology and ultrasonic testing technology [6–8]. However, these methods have some

limitations for the real-time monitoring of erosion of the elbow. Optical principle

detection technology need to send the test probe into the pipeline and, through the

acquisition of images, directly display the defect status. This method is not effective in

the environment of high pressure, multiphase medium and pipeline vibration [1]. The X-

ray digital real-time imaging testing method and infrared nondestructive testing method

are both radiographic technologies [9], which can be used to detect local corrosion of

pipelines and measure wall thickness with the help of standard image characteristics

display instrument. However, the X-ray method involves the radiation hazard, and the

thickness of pipe wall cannot be read directly. The main ultrasonic methods used in

nondestructive testing of pipeline defects include ultrasonic pulse reflection method and

ultrasonic guided wave method [10]. Whereas, these methods are mainly used for multi-

point discontinuous measurement, which requires the surface of the tested part to be flat

and uncoated. The irregular structure of elbow causes the modal transformation of guided

waves, which affects the propagation guided waves and the defect detection accuracy.

The aforementioned methods are mostly non-destructive testing based and may not be

suitable for real time monitoring [1]. In order to realize real-time monitoring, structural

health monitoring (SHM) that utilizes permanently installed sensors [11,12] integrated

with communication and signal processing algorithms to report the structure’s health

status in real time, has achieved great progress [13,14]. In SHM, Lead Zirconate Titanate

(PZT), with a strong piezoelectric effect, is a commonly used transducer that can be

integrated with a structure for real time monitoring [15–17]. Li et al. proposed a new

elbow erosion monitoring method, which combines the PZT (Lead Zirconate Titanate)
5
enabled active sensing with the fractional Fourier transform (FrRT), and takes the

fractional-order energy peak of the stress wave signal as the damage index. The results

show that there is a one-to-one relationship between mass of erosion loss and damage

index. Compared with the traditional time domain signal energy method, this method has

the advantage of eliminating the saturation phenomenon. However the above mentioned

methods all require installing sensors on the structure.

2.2 Percussion-based detection method

In recent years, the percussion method, that does not require the installation of

sensors on the structure, has gathered attentions in the application of detecting looseness

of various mechanical joints, such as spatial bolt-ball joint [18], bolted flange [19], and

cup-lock scaffolds [20]. The percussion method uses a structure’s sound response

subjected to a low-level impact or tapping for further analysis to detect the structure’s

health status [21–24]. The percussion method has a great potential for integration with

robotics technology to enable fully automated structural inspection. Even the percussion

method can be used to detect bolt looseness of a subsea flange using speech recognition

technology and least square support vector machine [25]. Also researches have been

conducted in detection of shear loading on bolts in bridge structures [26], sand deposition

in pipeline [27] and bolt head corrosion [28]. Next, selected research work using the

percussion methods are discussed in details.

Kong et al. proposed a percussion method to detect or monitor the bolt looseness.

The bolted connection was set at different tightening levels with a torque wrench

6
followed by percussions, and sound signals were collected. With selected ranges in the

PSD as features, a basic machine learning algorithm, decision tree, can successfully

classify the different tightening levels of the bolted connection [19]. Wang et al.

developed a novel robotic-assisted percussion method for spatial bolt-ball joint looseness

detection. In this research, mel-frequency cepstral coefficient (MFCC) based features

plus CNN-based deep machine learning method were used. In this experiment, a robotic

arm instead of manual tapping was used, which explored the potential implementation in

real industry. The experimental results showed that the proposed method was effective in

monitoring the connection looseness [18]. Wang et al. conducted an exploratory study to

attempt to research one issue that can affect the stability of scaffolding systems, namely

the looseness of the cup-lock joint. In this paper, to detect looseness of cup-lock scaffold,

a new percussion-based method was developed. A convolutional neural network (CNN)

was used to craft characteristics from MFCC features, and a bi-directional long short-

term memory architecture (BLSTM) was used to improve classification accuracy [20].

Wang et al. developed a percussion method to detect shear loading of through bolts in

bridge structures. This paper proposed a one-dimensional, memory-augmented

convolutional neural network (1D-MACNN) to classify the different types of shear

loading. The paper also compared the effectiveness of using two technologies, which

were detection by PZT-enabled active sensing and by percussion-based method. The

experimental results showed that the percussion method by using machine leaning

techniques can significantly achieve better classification accuracy [26].

7
 Percussion method is also used in applications other than structural looseness

detecting. Cheng et al. developed a new non-destructive approach using the percussion

method and voice recognition with support vector machine to detect the sand deposits in

the steel pipeline. The proposed mel-frequency cepstrum coefficients–based pipeline

deposit monitoring model can estimate the deposits in the pipeline with high accuracy

[27]. Wang et al. developed a novel entropy-enhanced acoustic emission (AE) method

with the help of machine learning to detect bolt head corrosion [28].

2.3 Sound Feature Selection and Machine learning algorithms

From literature review in Section 2.2, we can see that the percussion-based

detection methods are often used with machine learning (ML) algorithms. ML algorithms

at both basic learning and deep learning levels have seen great advances in the past few

decades [29–31]. ML techniques have also been used for voice and sound processing

[32–34] and normally consist of the following steps: data processing after collecting

sound signals, extraction of features that are sensitive to target, and the feature

classification via a certain classifier (e.g., KNN, DT, SVM or CNN/RNN) [26]. Even

though the classification accuracy through percussion and ML methods is relatively high,

the hand-crafted features require professional knowledge and extensive labor hours [26].

That means one method used in a previously-developed-research may not be efficient in

the other experiment on different structures. Researchers often need to test different

combination of features and ML algorithms to find the best fit for the particular scenario

being explored.

8
 Among these machine learning algorithms, neural networks, especially the deep

learning methods, are often used in application of sound signal classification. MFCC

features combined with neural network algorithms has demonstrated its superiority in

speech recognition [35,36]. For instance, Liu et al. proposed a MFCC-CNN hybrid

method for short utterance speaker recognition [37], and Rejaibi et al. proposed to use a

deep recurrent neural network-based framework (MFCC+RNN) to detect depression and

to predict its severity level from speech [38]. Similar researches are made on detecting

bowel sound by LSTM neural network using MFCC features [39] proposed by Liu et al.

In this thesis, a deep learning method based on RNN with help from MFCC is

proposed to effectively detect erosion levels of an elbow after exploring basic learning

methods including KNN, DT, and SVM.

9
3 Research Methodology and Methods

The general research methodology of the thesis is to use percussion induced sound

on pipeline elbows with machine learning methods to estimate the level of erosions if the

data are labeled (supervised learning), i.e., we have the baseline data. In case of absence

of the baseline data (unsupervised learning), we use clustering methods to detect the

severe erosion.

This chapter introduces the feature selection methods and the machine learning

methods used in this thesis. The power spectrum density (PSD) of the recorded signal

with selected frequency segment was used as signal features, and these PSD features

were applied with basic learning algorithms, including KNN, DT, and SVM, to classify

the different levels of elbow erosion. In addition, mel-frequency cepstral coefficients

(MFCCs) were used as features for the deep learning algorithm, recurrent neural network

(RNN). Finally, transformed MFCC features were used in clustering, including k-mean

and Gaussian mixture model (GMM), which are unsupervised learning algorithms. Figure

4 summarizes the methods used in this thesis [40].

10
 Figure 4: Flow chart of proposed methods in this thesis

3.1 Power spectrum density

The Power spectrum density, or simply power spectrum, of the input signal can be

estimated using the fast Fourier transform (FFT). It is a discrete Fourier transform

algorithm which reduces the number of computations needed for N points from 2N2 to

2N*log2N, where log2 is the base-2 logarithm. For each feature, the total energy of PSD

11
segment was computed by the summation of the PSD sampling values at the

corresponding frequency range [19].

The discrete Fourier transform (DFT) can be used to decompose any signal into a

sum of simple sine and cosine waves that we can easily measure the frequency, amplitude

and phase. DFT is defined as

/
2 2
cos sin (1)

where

 N = number of samples
 n = current sample
 k = current frequency, where k∈[0,N−1]
 xn = the sine value at sample n
 Xk = The DFT which include information of both amplitude and phase

The amplitude and phase of the signal can be calculated as

| |
amp
(2)

2 , ))

where Im(Xk) and Re(Xk) are the imagery and real part of the complex number, atan2 is

the two-argument form of the arctan function [41].

12
3.2 K-nearest neighbor algorithm

The k-nearest neighbor algorithm, also known as KNN, is a non-parametric,

supervised learning classifier, which uses proximity to make classifications or predictions

about the grouping of an individual data point. The goal of KNN algorithm is to identify

the nearest neighbors of a given query point, so that we can assign a class label to that

point [42].

Using the below formula, it measures a straight line between the query point and

the other point being measured [40]

, . (3)

The k value in the KNN algorithm defines how many neighbors will be checked to

determine the classification of a specific query point [42].

3.3 Decision tree algorithm

A decision tree is a non-parametric supervised learning algorithm, which is utilized

for both classification and regression tasks [43]. It has a hierarchical, tree structure, which

consists of a root node, branches, internal nodes and leaf nodes. A final output is based

on a series of decisions made from multiple conditions. The flow chart is shown in Figure

5 [43].

13
 Figure 5: Decision tree algorithm chart

3.4 Support vector machine algorithm

A support vector machine is a supervised learning algorithm used for many

classification and regression. The objective of the SVM algorithm is to find a hyperplane

that, to the best degree possible, separates data points of one class from those of another

class [44]. The “Best” is defined as the hyperplane with the largest margin between the

two classes, represented by plus versus minus in the Figure 6 below. The margin means

the maximal width of the slab parallel to the hyperplane that has no interior data points.

Only for linearly separable problems can the algorithm find such a hyperplane, and for

most practical problems the algorithm maximizes the soft margin allowing a small

number of misclassifications.

14
 Figure 6: Maximum-margin hyperplane and margins for an SVM [45]

MATLAB function “ClassificationECOC” is an error-correcting output codes

(ECOC) classifier for multiclass learning, where the classifier consists of multiple binary

learners such as SVMs [46]. This function was used in the experiment in this thesis.

3.5 Mel-frequency cepstral coefficients

Mel-frequency cepstral coefficients is a feature selection method widely used in

automatic speech and speaker recognition. The method was introduced by Davis and

Mermelstein in the 1980's for processing of voice data [47]. Hereafter are implementation

steps [48].

Step 1 is to frame an audio signal into 20-40ms frames.

15
Step 2 is to apply the discrete Fourier transform to obtain the power spectrum in the

frequency domain. The formula is given by

(4)

where h(n) is an N sample long analysis window (e.g. hamming window), and k is the

length of the DFT. The periodogram-based power spectral estimate for the speech frame

is given by

| | . (5)

This is called the periodogram estimate of the power spectrum.

Step 3 is to convert the frequency (Hertz scale) to the mel scale through a filter bank. The

first filter is very narrow and gives an indication of how much energy exists near 0 Hertz.

As the frequencies get higher our filters get wider as we become less concerned about

variations. The mel scale tells us exactly how to space our filter banks and how wide to

make them. The formula for converting from frequency to mel scale is

1125 ln 1 . (6)
700

To go from mel scale back to frequency, the formula is given by

16
 700 exp 1
1125

0, 1
1 (7)
, 1
1
1
, 1
1
0, 1

where m is the number of filters we want, and f( ) is the list of m+2 mel-spaced

frequencies.

Step 4 is to take the logarithm of the filter bank energies once we have them. The

logarithm allows us to use cepstral mean subtraction, which is a channel normalization

technique.

Step 5, the final step, is to compute the discrete cosine transform (DCT) of the log filter

bank energies. The formula is given by

.
∑ cos
. (8)
∑ | |

There are two main reasons why this is performed. Since the filter banks are all

overlapping, the filter bank energies are quite correlated with each other. The DCT

decorrelates the energies, which means diagonal covariance matrices can be used to

model the features in an HMM (hidden Markov model) classifier. Only 12-14 of the 26

17
DCT coefficients are kept. This is because the higher DCT coefficients represent fast

changes in the filter bank energies and it turns out that these fast changes actually

degrade automatic speech recognition performance, therefore we obtain a small

improvement by dropping them.

In order to use clustering methods we developed a new feature based on MFCC,

i.e. transformed MFCC. The MFCC in this thesis has 14 lines of features and 24 columns

of time steps. The transformed MFCC is a vector which has size of 1x28. All data in each

row are calculated with a mean and a standard deviation, which formed the new

transformed MFCC vector.

3.6 Recurrent neural network

Recurrent neural network is a class of neural networks where connections between

nodes form a directed graph along a temporal sequence. This allows it to exhibit temporal

dynamic behavior [49]. Derived from feedforward neural networks, RNNs can use their

internal state (memory) to process variable length sequences of inputs. This makes them

applicable to tasks of “sequence translation,” such as handwriting recognition and speech

recognition [49]. RNNs are distinguished by their “memory” as they take information

from prior inputs to influence the current input and output. RNN is a type of deep

learning method [49].

An LSTM (Long Short Term Memory) network is a type of recurrent neural

network that can learn long-term dependencies between time steps of sequence data [50].

This diagram illustrates the architecture of a simple LSTM network for classification.
18
The network, as shown in figure 7, starts with a sequence input layer followed by an

LSTM layer. To predict class labels, the network ends with a fully connected layer, a

softmax layer, and a classification output layer [51].

Figure 7: Flow chart of recurrent neural network

This diagram in Figure 8 illustrates the flow of a time series X with C features

(channels) of length S through an LSTM layer. In the diagram, ht and ct denote the output

(also known as the hidden state) and the cell state at time step t, respectively [51].

Figure 8: Diagram of flow of time series through LSTM layer

19
 The first LSTM block uses the initial state of the network and the first time step of

the sequence to compute the first output and the updated cell state. At time step t, the

block uses the current state of the network (ct−1, ht−1) and the next time step of the

sequence to compute the output and the updated cell state ct. The state of the layer

consists of the hidden state (also known as the output state) and the cell state. The hidden

state at time step t contains the output of the LSTM layer for this time step. The cell state

contains information learned from the previous time steps. At each time step, the layer

adds information to or removes information from the cell state. The layer controls these

updates using gates. The following components, as illustrated in Table 1, control the cell

state and hidden state of the layer [51].

Table 1: Components to control the cell state and hidden state

Component Purpose

Input gate (i) Control level of cell state update

Forget gate (f) Control level of cell state reset (forget)

Cell candidate (g) Add information to cell state

Output gate (o) Control level of cell state added to hidden state

This diagram in Figure 9 illustrates the flow of data at time step t. The diagram highlights

how the gates forget, update, and output the cell and hidden states [51].

20
 Figure 9: Diagram of an individual cell

The learnable weights of an LSTM layer are the input weights W (input weights),

the recurrent weights R (recurrent weights), and the bias b (bias). The matrices W, R, and

b are concatenations of the input weights, the recurrent weights, and the bias of each

component, respectively. These matrices are concatenated as

, , (9)

where i, f, g, and o denote the input gate, forget gate, cell candidate, and output gate,

respectively. The cell state at time step t is given by

(10)
⊙ 1 ⊙

where ⊙ denotes the Hadamard product (element-wise multiplication of vectors).

The hidden state at time step t is given by

(11)
⊙

21
where σc denotes the state activation function. The ‘lstmLayer’ function, by default, uses

the hyperbolic tangent function (tanh) to compute the state activation function. The

following formulas in Table 2 describe the components at time step t.

Table 2: Formulas of components

Component Formula

Input gate

Forget gate

Cell candidate

Output gate

In these calculations, σg denotes the gate activation function. The ‘lstmLayer’

−1
function, by default, uses the sigmoid function given by σ(x)=(1+e−x) to compute the

gate activation function.

3.7 K-means clustering

K-means clustering, shown in Figure 10, is a partitioning method which is one of

unsupervised machine learning method. The algorithm calculates the best category based

on the similarity of the distance between sample points. And the closer the two objects

are, the more similar they are. Its core idea is to divide samples into different categories

through an iterative process. This algorithm needs to specify the number of clusters, and

it can scale up to a large number of samples [40].

22
 The MATLAB function ‘kmeans’ partitions data into k mutually exclusive

clusters and returns the index of the cluster to which it assigns each observation, and it

also treats each observation in the data as an object that has a location in space [52]. The

function finds a partition in which objects within each cluster are as close to each other as

possible, and as far from objects in other clusters as possible. Like many clustering

methods, k-means clustering requires to specify the number of clusters k before clustering

[52].

Figure 10: K-mean clustering data [53]

3.8 Clustering using Gaussian mixture model

Gaussian mixture model (GMM), shown in Figure 11, is often used for data

clustering. To perform clustering, the GMM assigns query data points to the multivariate

normal components that maximize the component posterior probability. That is, given a

23
fitted GMM, the cluster assigns query data to the component yielding the highest

posterior probability.

Figure 11: Gaussian Mixture Models [54]

GMM clustering can accommodate clusters that have different sizes and

correlation structures within them. Therefore, in certain applications, GMM clustering

can be more appropriate than methods such as k-means clustering [55]. GMM can also be

integrated with other methods, such as HMM (Hidden Markov Model), to achieve more

comprehensive tasks [56,57]. Like many clustering methods, GMM clustering requires

you to specify the number of clusters before fitting the model. The number of clusters

specifies the number of components in the GMM. For GMMs, there are below best

practices:

 Consider the component covariance structure. You can specify diagonal or full

covariance matrices, and whether all components have the same covariance

matrix.

24
  Specify initial conditions. The Expectation-Maximization (EM) algorithm fits the

GMM.

 Implement regularization.

In general, the Gaussian density function is given below [54] by

1 1 (12)
| ,Σ exp Σ
2
2 |Σ|

where x represents our data points, and D is the number of dimensions of each data point.

μ and Σ are the mean and covariance, respectively. If we have a dataset comprised of N =

1000 three-dimensional points (D = 3), then x will be a 1000 × 3 matrix. μ will be a 1 × 3

vector, and Σ will be a 3 × 3 matrix. For later purposes, we will also find it useful to take

the log of this equation, which is given by [54]

1 1 (13)
| ,Σ 2 Σ Σ .
2 2 2

Below is the Expectation-Maximization algorithm:

 Compute soft membership probabilities, “responsibility” r(ic)

 Update mixture component parameters given soft memberships

 Ascent on log-likelihood: convergent, but local optima

25
4 Experiment Setup and Data Collection

Percussion was performed on the bottom corner of elbows, as shown in Figure 9,

to detect the different erosion rate of the elbow with the assistance of machine learning

algorithms. After impacts, signals were input into MATLAB and pre-processed, there

were three strategies doing data analysis. The first strategy was to process the input signal

by PSD and train and test the processed data with basic machine learning methods, which

were KNN, DT and SVM. The second phase was to process the input signal by MFCC

and use deep learning method, RNN, to train and test the data. The third phase was to use

transformed MFCC as features and unsupervised machine learning methods, i.e. K-mean

clustering and clustering with GMM to classify the data.

In the experiment, there were three identical sets of elbow and pipe assembly, as

shown in Figure 12.

Figure 12: Percussion performed on elbow and three sets of elbow and pipe assembly

26
 An impact hammer with a metal tip was used to repeatedly tap on the bottom of

the elbow. Meanwhile, the hit-sound was recorded by a smart phone. The percussion-

induced sound signals were pre-processed and trained by different machine learning

algorithms. Also cross validation data was used to test the trained model.

In the experiment, we considered four scenarios at which sound signals were

collected as shown in Figure 13. Scenario one is an elbow only and it is hanged with a

string; Scenario two is an assembly of an elbow with one pipe being connected on one

end of the elbow; Scenario three is an assembly of an elbow with two pipes being

connected on both ends of the elbow; Scenario four was is the same assembly as Scenario

three, but both pipes are fixed on a wooden board.

Figure 13: Four scenarios of experiment

The erosion rate for each scenario has six classes. Class 1 uses original elbow

with no mass loss, and class 2 to class 6 involve 0.4g mass loss on each step. The details

27
of erosion rates corresponding to the class numbers are listed in below table 3. Also

Figure 14 shows the elbows after grinding on the last step and one example of zoomed-in

elbow’s picture. The tools included a hammer with metal tip and a smart phone which

was used to record signals. The tapping point is on the bottom of elbow. The smart phone

was placed on the table right beneath the wooden frame.

Table 3: Arrangement of six classes

Mass of pipeline elbow (g)

Elbow Class Mass loss (g)
A B C

C1 0.0 404.4 406 406.7

C2 0.4 404 405.6 406.3

C3 0.8 403.6 405.2 405.9

A/B/C
C4 1.2 403.2 404.8 405.5

C5 1.6 402.8 404.4 405.1

C6 2.0 402.4 404 404.7

Figure 14: The elbows after grinding on the last step

28
5 Data Analyses and Results Discussion

5.1 Erosion detection using basic machine learning algorithms

The signals were input and processed by PSD, from which the most sensitive nine

frequency segments were manually selected as features. They form a vector that is

edges

1000,1500,1600,2150,5750,6200,6250,7070,7200,7700, …
(14)

9000,9300,10000,11000,13000, 14000,17000,18000 .

The PSD plots for the 6 different erosion levels are printed in Figure 15.

Figure 15: Plot of six erosion rates of power spectrum of signals in frequency domain
29
 All processed signals of six classes were labelled. Three random PSD features

from elbow A, the 2nd, 5th and 7th features, were selected to plot a 3D figure showing the

scattering of all labelled data as in below Figure 16. From the plot we can see that from

the scenario 1 to the scenario 4, the data were scattered more and more easily being

classified. Also three other features, i.e. the 3rd, 5th, and 9th features, were selected and

plotted to see if data were scattered better. Figure 17 shows that scenario 2 and 3 had

better results when using the other set of features. From the above observation the

scenario 4 had the best classified scattering data.

Figure 16: 3D plot showing scattering data on PSD features (2nd, 5th and 7th)

30
 Figure 17: 3D plot showing scattering data on PSD features (3rd, 5th, and 9th)

By doing so, we obtained the data matrix which had size of 48 x 10 with last

column as y-label. For the training data, we had data from elbow A, B, and C forming a

864 x 10 matrix, i.e. total 864 data. Then we partitioned the whole data into 80/20

percentage. 80% of the data were used as training data.

Firstly, we use KNN classification to train the combined data set of elbow A, B

and C. By setting K value as 3, we obtain the highest accuracy. Then we use the trained

model to test the cross validated data, which is 20% of partitioned data. Secondly, we

31
use DT classification to perform the same training and testing as stated above. Thirdly,

we use SVM classification to train the model. Since the standard SVM classification

learner only deals with binary model, we used a MATLAB built-in function, fitcecoc, to

train a multiclass SVM model. Test results are shown as in Table 4.

Table 4: Test Results of basic learning methods

Scenario 1 Scenario 2 Scenario 3 Scenario 4

Elbow only Elbow + pipe Elbow + two Fixed Elbow +

pipes two pipes

ML Train Test Train Test Train Test Train Test

Algorithm Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy

PSD+KNN 81.60 73.84 83.80 73.64 88.66 77.33 94.79 84.88

PSD+DT 86.57 64.53 89.00 62.79 91.90 64.53 93.98 75.00

PSD+SVM 94.44 69.19 96.53 67.44 98.61 77.91 98.50 86.05

We observe that the accuracy on the training data are above 80% for KNN and

DT methods while SVM method has the best result. We also observe that the accuracy on

the cross validates testing data from these three classification methods are above 60%.

32
Also we can conclude that the scenario 4 has the best train and test results. From these

basic machine learning methods, SVM has the best overall accuracy.

5.2 Erosion detection using deep learning method

In order to improve the testing accuracy, deep learning algorithm was

implemented for data classification. Here we adopted MFCC + RNN method. After

collecting sound via percussion of elbows, we employed the MFCC to process the signals

and obtained a time sequence feature. An example of MFCC feature is displayed below

in Figure 18, which shows the data, the 20th signal, from all six classes of elbow A and

scenario 4.

Figure 18: MFCC features extracted from single frame sound signal

33
 Data from each class is processed by MFCC, generating a 48 x 1 cell array, which

contains 48 MFCC feature matrixes. Each feature matrix has 14 lines of time steps and 24

columns of features. MFCC matrixes of all classes are labelled correspondently. To train

a deep neural network to classify each time step of sequence data, a sequence-to-

sequence LSTM network was used. The input size is 14 since there are 14 time steps for

each sample. The number of hidden units are set to 200. The number of classes are 6 as

there are six classes of each scenario. The data are also divided into 80% as training part

and 20% as testing part. The training accuracy reaches 95% on the first 12 iterations, and

quickly reached 100% in short time of computation. The accuracy on testing data reached

95% and higher on all four scenarios. The training and testing process for scenarios 1 to 4

are shown in Figures 19 to 22.

Figure 19: The training process of RNN for scenario 1

34
Figure 20: The training process of RNN for scenario 2

Figure 21: The training process of RNN for scenario 3

35
 Figure 22: The training process of RNN for scenario 4

Also the confusion matrix for all four scenarios are displaced in Tables 5 to 8.

Table 5: Confusion matrix of RNN method for scenario 1

confusion_matrix_test_set_scenario 1
True
Class Accuracy Error
C1 34 0 0 0 0 0 100.00
C2 2 34 0 0 0 0 94.44 5.56
C3 0 1 31 0 0 0 96.88 3.13
C4 0 0 3 16 0 0 84.21 15.79
C5 0 0 0 0 22 0 100.00
C6 0 0 0 0 0 29 100.00

94.44 97.14 91.18 100.00 100.00 100.00

5.56 2.86 8.82
Predicted
Class C1 C2 C3 C4 C5 C6

36
Table 6: Confusion matrix of RNN method for scenario 2

confusion_matrix_test_set_scenario 2
True
Class Accuracy Error
C1 30 0 1 0 0 0 96.77 3.23
C2 0 27 0 0 0 0 100.00
C3 0 0 27 0 0 0 100.00
C4 0 0 0 31 0 0 100.00
C5 0 0 0 0 24 3 88.89 11.11
C6 0 0 0 0 0 29 100.00

100.00 100.00 96.43 100.00 100.00 90.63

3.57 9.38
Predicted
Class C1 C2 C3 C4 C5 C6

Table 7: Confusion matrix of RNN method for scenario 3

confusion_matrix_test_set_scenario 3
True
Class Accuracy Error
C1 29 0 0 0 0 1 96.67 3.33
C2 0 25 1 0 0 0 96.15 3.85
C3 0 0 30 0 0 0 100.00
C4 0 0 0 22 2 1 88.00 12.00
C5 0 0 0 0 25 0 100.00
C6 0 1 0 0 0 35 97.22 2.78

100.00 96.15 96.77 100.00 92.59 94.59

3.85 3.23 7.41 5.41
Predicted
Class C1 C2 C3 C4 C5 C6

37
Table 8: Confusion matrix of RNN method for scenario 4

confusion_matrix_test_set_scenario 4
True
Class Accuracy Error
C1 33 0 0 0 0 0 100.00
C2 1 28 0 0 0 0 96.55 3.45
C3 0 0 28 0 0 0 100.00
C4 0 0 0 24 0 0 100.00
C5 0 0 0 1 24 0 96.00 4.00
C6 0 0 0 1 0 32 96.97 3.03

97.06 100.00 100.00 92.31 100.00 100.00

2.94 7.69
Predicted
Class C1 C2 C3 C4 C5 C6

Through experiments in this phase, better results are achieved by applying the

deep learning method. The detailed training and testing results are shown in Table 9.

Table 9: Test results of deep learning method

Scenario 1 Scenario 2 Scenario 3 Scenario 4

Elbow only Elbow + pipe Elbow + two Fixed Elbow +

pipes two pipes

Train Test Train Test Train Test Train Test

Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy

MFCC+RNN 100 96.51 100 97.67 100 96.51 100 98.26

38
5.3 Severe erosion detection using clustering method

The last phase of experiments is to find out what degree of accuracy we can get if

we use supervised learning methods (data are labeled). Here we use two clustering

methods. One is the k-means method, and the other is Gaussian mixture model. K-means

clustering is an unsupervised clustering algorithm. The algorithm, shown in Figure 23,

calculates the best category based on the similarity of the distance between sample points.

And the closer the two objects are, the more similar they are [40], while Gaussian

mixture models account for covariance. For example in two dimensions, covariance

determines the shape of the distribution [54].

Figure 23: K-means clustering method’s diagram

In order to use these two clustering methods, we choose data from three classes,

i.e. the first class, the fourth class and the sixth class. Then we modified MFCC feature

by taking mean and standard deviation for all time steps, i.e. the data in 24 columns in the

39
MFCC matrix. All data in each row are calculated with a mean and a standard deviation.

Each MFCC line has 14 mean values and 14 standard deviation values, which form a

vector feature with dimension 1x28. With these transformed MFCC features, both

clustering methods are implemented. The results of k-means clustering method is

displayed in Figure 24. The data are combined from elbow A, B and C, and three random

features, the 3rd, 15th and 24th features, were selected. Also the results of clustering with

GMM is displayed in Figure 25. The selected three random features were the 3rd, 11th,

and 21st. Both figures successfully showed that three classes of data were clustered.

However, we only worked to classify data in this thesis, but we couldn’t tell which one is

the most erosive data. This work will be in included in future researches.

Figure 24: 3D plot showing results of K-means clustering on selected transformed MFCC
features

40
Figure 25: 3D plot showing results of clustering with GMM on selected transformed
MFCC features

From the experimental results we can see that both clustering methods

successfully classified the elbow at severe erosion rates, however, the GMM method has

a better accuracy than k-means method. K-means method reached an overall accuracy

from 49% to 68%, while the GMM method improved the accuracy to around 77% and a

very high accuracy with 90% on scenario 1. Clustering results are listed in Table 10.

41
Table 10: Classification results of unsupervised learning methods

K‐MEANS GMM‐CLUSTERING
single single
class overall class overall
Class1 Class2 Class3 accuracy accuracy Class1 Class2 Class3 accuracy accuracy
Scenario 1 Scenario 1
99 44 1 68.75 144 0 0 100.00
30 103 11 71.53 68.06 17 103 24 71.53 90.28
19 33 92 63.89 1 0 143 99.31
Scenario 2 Scenario 2
53 84 7 36.81 92 50 2 63.89
3 63 78 43.75 49.31 46 94 4 65.28 69.21
33 14 97 67.36 31 0 113 78.47
Scenario 3 Scenario 3
61 56 27 42.36 92 46 6 63.89
9 103 32 71.53 67.36 3 127 14 88.19 77.78
8 9 127 88.19 0 27 117 81.25
Scenario 4 Scenario 4
58 86 0 40.28 95 49 0 65.97
15 98 31 68.06 66.44 49 95 0 65.97 77.31
7 6 131 90.97 0 0 144 100.00

42
6 Conclusion and Future Work

6.1 Conclusion

In this study, we used percussion method with machine learning algorithms to

classify different degrees of elbow erosion. We simulated the degree of erosion by

grinding off mass on elbow’s internal wall and applied different machine learning

algorithms, which were KNN, DT, SVM, RNN and clustering. For KNN, DT and SVM,

we obtained the training and testing accuracies in the range of 70% to 80%. Then, by

using MFCC matrix with time sequence as the feature, we applied one deep neural

network, RNN, and achieved very high accuracy which is above 95%. Finally

unsupervised learning, clustering, was explored. Mean and standard deviation of all

features on each MFCC time step were calculated as input features, and k-means

clustering and clustering with Gaussian mixture model were used. For k-means

clustering, the final accuracy was not stable since it varied widely from 49% to 68%. To

improve the results, the GMM clustering, with the same features, was used and the better

performance was achieved with the accuracy ranging from 70% to 90%.

To conclude, among these machine learning algorithms, the MFCC plus deep

learning algorithm, RNN, had the best accuracy on both training and testing data.

6.2 Recommendation for future work

During the experiment, it was observed that there was one limitation of percussion

method in detecting erosion rate of elbow’s wall thickness. The percussion method is

very sensitive to mass loss from the inner wall of elbow. Any tiny difference on mass or
43
grinding location would affect PSD of the collected sound signals. This limitation

decreases the accuracy, especially on independent testing data. Based on this limitation,

we would recommend that this percussion method could be used to detect critical erosion

rate or minimum wall thickness of elbow.

Besides the limitation mentioned in the above paragraph, there are also other

factors, such as corrosion, line pressure and sediment in pipe, to affect the detection

results of elbow’s wall thickness. Among these factors, corrosion and erosion normally

happen together affecting elbow’s wall thickness. Static and dynamic line pressure also

can affect the energy of sound by percussion, and this factor needs to be considered when

evaluating the rate of erosion and corrosion combined situation. Sediment in pipeline is

also a factor to be considered to affect the accuracy of detection of residual wall

thickness. All the above factors need to be examined and compensated when determining

the real wall thickness of elbows in service. Research along these areas can be performed

in the future.

Comparing to SVM classification method, linear discriminant analysis (LDA) is

also widely used in data classification. LDA has an analytical solution while SVM has a

numerical solution. Using LDA to analyze and train multi-class percussion data and

comparing of both classification methods can be explored in future work.

44
References

[1] Li, N., Wang, F., and Song, G. “A Feasibility Study on Elbow Erosion Monitoring

Using Active Sensing Approach and Fractional Fourier Transform.” Journal of

Intelligent Material Systems and Structures, Vol. 32, No. 5, 2021, pp. 584–596.

https://doi.org/10.1177/1045389X20963172.

[2] Peng, W., and Cao, X. “Numerical Prediction of Erosion Distributions and Solid

Particle Trajectories in Elbows for Gas–Solid Flow.” Journal of Natural Gas

Science and Engineering, Vol. 30, 2016, pp. 455–470.

https://doi.org/10.1016/j.jngse.2016.02.008.

[3] Hassani, S. Solid Particle Erosion, Sand Monitoring and Transport in Oil and Gas

Production. https://www.linkedin.com/pulse/solid-particle-erosion-sand-

monitoring-transport-oil-gas-hassani-1. Accessed Sep. 20, 2022.

[4] Kusmono, and Khasani. “Analysis of a Failed Pipe Elbow in Geothermal

Production Facility.” Case Studies in Engineering Failure Analysis, Vol. 9, 2017,

pp. 71–77. https://doi.org/10.1016/J.CSEFA.2017.08.001.

[5] Zhang, J., Kang, J., Fan, J., and Gao, J. “Study on Erosion Wear of Fracturing

Pipeline under the Action of Multiphase Flow in Oil & Gas Industry.”

Journal of Natural Gas Science and Engineering, Vol. 32, 2016, pp. 334–346.

https://doi.org/10.1016/j.jngse.2016.04.056.

45
[6] Wang, Z. D., Gu, Y., and Wang, Y. S. “A Review of Three Magnetic NDT

Technologies.” Journal of Magnetism and Magnetic Materials, Vol. 324, No. 4,

2012, pp. 382–388. https://doi.org/10.1016/j.jmmm.2011.08.048.

[7] Datta, S., and Sarkar, S. “A Review on Different Pipeline Fault Detection

Methods.” Journal of Loss Prevention in the Process Industries, Vol. 41, 2016, pp.

97–106. https://doi.org/10.1016/j.jlp.2016.03.010.

[8] Coramik, M., and Ege, Y. “Discontinuity Inspection in Pipelines: A Comparison

Review.” Measurement, Vol. 111, 2017, pp. 359–373.

https://doi.org/10.1016/j.measurement.2017.07.058.

[9] Udod, V. A., Van, Ya., Osipov, S. P., Chakhlov, S. v., Usachev, E. Yu., Lebedev,

M. B., and Temnik, A. K. “State-of-the Art and Development Prospects of Digital

Radiography Systems for Nondestructive Testing, Evaluation, and Inspection of

Objects: A Review.” Russian Journal of Nondestructive Testing, Vol. 52, No. 9,

2016, pp. 492–503. https://doi.org/10.1134/S1061830916090072.

[10] Alobaidi, W. M., Alkuam, E. A., Al-Rizzo, H. M., and Sandgren, E. “Applications

of Ultrasonic Techniques in Oil and Gas Pipeline Industries: A Review.” American

Journal of Operations Research, Vol. 05, No. 04, 2015, pp. 274–287.

https://doi.org/10.4236/ajor.2015.54021.

[11] Tsangouri, E., Karaiskos, G., Aggelis, D. G., Deraemaeker, A., and van

Hemelrijck, D. “Crack Sealing and Damage Recovery Monitoring of a Concrete

46
 Healing System Using Embedded Piezoelectric Transducers.” Structural Health

Monitoring, Vol. 14, No. 5, 2015, pp. 462–474.

https://doi.org/10.1177/1475921715596219.

[12] Liu, T., Zou, D., Du, C., and Wang, Y. “Influence of Axial Loads on the Health

Monitoring of Concrete Structures Using Embedded Piezoelectric Transducers.”

Structural Health Monitoring, Vol. 16, No. 2, 2017, pp. 202–214.

https://doi.org/10.1177/1475921716670573.

[13] Peairs, D. M., Park, G., and Inman, D. J. “Improving Accessibility of the

Impedance-Based Structural Health Monitoring Method.” Journal of Intelligent

Material Systems and Structures, Vol. 15, No. 2, 2004, pp. 129–139.

https://doi.org/10.1177/1045389X04039914.

[14] Zhang, T., Biswal, S., and Wang, Y. “SHMnet: Condition Assessment of Bolted

Connection with beyond Human-Level Performance.” Structural Health

Monitoring, Vol. 19, No. 4, 2020, pp. 1188–1201.

https://doi.org/10.1177/1475921719881237.

[15] Farinholt, K. M., Miller, N., Sifuentes, W., MacDonald, J., Park, G., and Farrar, C.

R. “Energy Harvesting and Wireless Energy Transmission for Embedded SHM

Sensor Nodes.” Structural Health Monitoring, Vol. 9, No. 3, 2010, pp. 269–280.

https://doi.org/10.1177/1475921710366647.

47
[16] Talakokula, V., Bhalla, S., and Gupta, A. “Monitoring Early Hydration of

Reinforced Concrete Structures Using Structural Parameters Identified by Piezo

Sensors via Electromechanical Impedance Technique.” Mechanical Systems and

Signal Processing, Vol. 99, 2018, pp. 129–141.

https://doi.org/10.1016/j.ymssp.2017.05.042.

[17] Li, W., Fan, S., Ho, S. C. M., Wu, J., and Song, G. “Interfacial Debonding

Detection in Fiber-Reinforced Polymer Rebar–Reinforced Concrete Using Electro-

Mechanical Impedance Technique.” Structural Health Monitoring, Vol. 17, No. 3,

2018, pp. 461–471. https://doi.org/10.1177/1475921717703053.

[18] Wang, F., Mobiny, A., van Nguyen, H., and Song, G. “If Structure Can Exclaim:

A Novel Robotic-Assisted Percussion Method for Spatial Bolt-Ball Joint

Looseness Detection.” Structural Health Monitoring, Vol. 20, No. 4, 2021, pp.

1597–1608. https://doi.org/10.1177/1475921720923147.

[19] Kong, Q., Zhu, J., Ho, S. C. M., and Song, G. “Tapping and Listening: A New

Approach to Bolt Looseness Monitoring.” Smart Materials and Structures, Vol.

27, No. 7, 2018, p. 07LT02. https://doi.org/10.1088/1361-665X/aac962.

[20] Wang, F., and Song, G. “Looseness Detection in Cup-Lock Scaffolds Using

Percussion-Based Method.” Automation in Construction, Vol. 118, 2020, p.

103266. https://doi.org/10.1016/j.autcon.2020.103266.

48
[21] Yuan, R., Lv, Y., Kong, Q., and Song, G. “Percussion-Based Bolt Looseness

Monitoring Using Intrinsic Multiscale Entropy Analysis and BP Neural Network.”

Smart Materials and Structures, Vol. 28, No. 12, 2019, p. 125001.

https://doi.org/10.1088/1361-665X/ab3b39.

[22] Zhou, Y., Wang, S., Zhou, M., Chen, H., Yuan, C., and Kong, Q. “Percussion

based Bolt Looseness Identification Using Vibration guided Sound

Reconstruction.” Structural Control and Health Monitoring, Vol. 29, No. 2, 2022.

https://doi.org/10.1002/stc.2876.

[23] Chen, L., Xiong, H., Sang, X., Yuan, C., Li, X., and Kong, Q. “An Innovative

Deep Neural Network–Based Approach for Internal Cavity Detection of Timber

Columns Using Percussion Sound.” Structural Health Monitoring, Vol. 21, No. 3,

2022, pp. 1251–1265. https://doi.org/10.1177/14759217211028524.

[24] Kong, Q., Ji, K., Gu, J., Chen, L., and Yuan, C. “A CNN-Integrated Percussion

Method for Detection of FRP–Concrete Interfacial Damage with FEM

Reconstruction.” Structural Health Monitoring, 2022, p. 147592172210820.

https://doi.org/10.1177/14759217221082007.

[25] Wang, F., Chen, X., and Song, G. Percussion-Based Detection of Bolt Looseness

Using Speech Recognition Technology and Least Square Support Vector Machine.

2020.

49
[26] Wang, F., Song, G., and Mo, Y. “Shear Loading Detection of through Bolts in

Bridge Structures Using a Percussion based One dimensional Memory

augmented Convolutional Neural Network.” Computer-Aided Civil and

Infrastructure Engineering, Vol. 36, No. 3, 2021, pp. 289–301.

https://doi.org/10.1111/mice.12602.

[27] Cheng, H., Wang, F., Huo, L., and Song, G. “Detection of Sand Deposition in

Pipeline Using Percussion, Voice Recognition, and Support Vector Machine.”

Structural Health Monitoring, Vol. 19, No. 6, 2020, pp. 2075–2090.

https://doi.org/10.1177/1475921720918890.

[28] Wang, F., and Zhu, R. “Detection of Bolt Head Corrosion under External

Vibration Using a Novel Entropy-Enhanced Acoustic Emission Method.”

Nonlinear Dynamics, Vol. 108, No. 4, 2022, pp. 3807–3816.

https://doi.org/10.1007/s11071-022-07390-x.

[29] Yuan, F.-G., Zargar, S. A., Chen, Q., and Wang, S. Machine Learning for

Structural Health Monitoring: Challenges and Opportunities. 2020.

[30] Malekloo, A., Ozer, E., AlHamaydeh, M., and Girolami, M. “Machine Learning

and Structural Health Monitoring Overview with Emerging Technology and High-

Dimensional Data Source Highlights.” Structural Health Monitoring, Vol. 21, No.

4, 2022, pp. 1906–1955. https://doi.org/10.1177/14759217211036880.

50
[31] Bao, Y., and Li, H. “Machine Learning Paradigm for Structural Health

Monitoring.” Structural Health Monitoring, Vol. 20, No. 4, 2021, pp. 1353–1372.

https://doi.org/10.1177/1475921720972416.

[32] Purwins, H., Li, B., Virtanen, T., Schlüter, J., Chang, S., and Sainath, T. “Deep

Learning for Audio Signal Processing.” 2019.

https://doi.org/10.1109/JSTSP.2019.2908700.

[33] Latif, S., Cuayáhuitl, H., Pervez, F., Shamshad, F., Ali, H. S., and Cambria, E. “A

Survey on Deep Reinforcement Learning for Audio-Based Applications.”

Artificial Intelligence Review, 2022. https://doi.org/10.1007/s10462-022-10224-2.

[34] Lane, N. D., Georgiev, P., and Qendro, L. DeepEar. 2015.

[35] Tiwari, V. “MFCC and Its Applications in Speaker Recognition.” International

journal on emerging technologies, Vol. 1, No. 1, 2010, pp. 19–22.

[36] Hossan, Md. A., Memon, S., and Gregory, M. A. A Novel Approach for MFCC

Feature Extraction. 2010.

[37] Liu, Z., Wu, Z., Li, T., Li, J., and Shen, C. “GMM and CNN Hybrid Method for

Short Utterance Speaker Recognition.” IEEE Transactions on Industrial

Informatics, Vol. 14, No. 7, 2018, pp. 3244–3252.

https://doi.org/10.1109/TII.2018.2799928.

51
[38] Rejaibi, E., Komaty, A., Meriaudeau, F., Agrebi, S., and Othmani, A. “MFCC-

Based Recurrent Neural Network for Automatic Clinical Depression Recognition

and Assessment from Speech.” Biomedical Signal Processing and Control, Vol.

71, 2022, p. 103107. https://doi.org/10.1016/j.bspc.2021.103107.

[39] Liu, J., Yin, Y., Jiang, H., Kan, H., Zhang, Z., Chen, P., Zhu, B., and Wang, Z.

Bowel Sound Detection Based on MFCC Feature and LSTM Neural Network.

2018.

[40] Song, G. Dr.Song’s Machine Learning Course Notes.

[41] Discrete Fourier Transform.

https://en.wikipedia.org/wiki/Discrete_Fourier_transform. Accessed Nov. 20,

2022.

[42] K-Nearest Neighbors Algorithm. https://www.ibm.com/topics/knn. Accessed Nov.

20, 2022.

[43] Decision Trees Explained With a Practical Example.

https://towardsai.net/p/programming/decision-trees-explained-with-a-practical-

example-fe47872d3b53. Accessed Nov. 20, 2022.

[44] Support Vector Machine Algorithm. https://www.geeksforgeeks.org/support-

vector-machine-algorithm/. Accessed Nov. 20, 2022.

52
[45] Support Vector Machine. https://en.wikipedia.org/wiki/Support_vector_machine.

Accessed Nov. 20, 2022.

[46] Multiclass Model for Support Vector Machines (SVMs) and Other Classifiers -

MATLAB.

https://www.mathworks.com/help/stats/classificationecoc.html;jsessionid=c634e18

5f8f51f48dce05015b079. Accessed Sep. 20, 2022.

[47] Ayvaz, U., Gürüler, H., Khan, F., Ahmed, N., Whangbo, T., and Akmalbek

Bobomirzaevich, A. “Automatic Speaker Recognition Using Mel-Frequency

Cepstral Coefficients Through Machine Learning.” Computers, Materials &

Continua, Vol. 71, No. 3, 2022, pp. 5511–5521.

https://doi.org/10.32604/cmc.2022.023278.

[48] Mel Frequency Cepstral Coefficient (MFCC) Tutorial.

http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-

frequency-cepstral-coefficients-mfccs/. Accessed Nov. 20, 2022.

[49] Recurrent Neural Network.

https://en.wikipedia.org/wiki/Recurrent_neural_network. Accessed Nov. 20, 2022.

[50] Long Short-Term Memory. https://en.wikipedia.org/wiki/Long_short-

term_memory. Accessed Nov. 20, 2022.

53
[51] Long Short-Term Memory Networks - MATLAB & Simulink.

https://www.mathworks.com/help/deeplearning/ug/long-short-term-memory-

networks.html. Accessed Sep. 20, 2022.

[52] K-Means Clustering - MATLAB & Simulink.

https://www.mathworks.com/help/stats/k-means-clustering.html. Accessed Sep.

20, 2022.

[53] Clustering Data. https://i.stack.imgur.com/cIDB3.png. Accessed Nov. 20, 2022.

[54] Carrasco, O. C. Gaussian Mixture Models Explained | by Oscar Contreras

Carrasco | Towards Data Science. https://towardsdatascience.com/gaussian-

mixture-models-explained-6986aaf5a95. Accessed Sep. 20, 2022.

[55] Cluster Using Gaussian Mixture Model - MATLAB & Simulink.

https://www.mathworks.com/help/stats/clustering-using-gaussian-mixture-

models.html. Accessed Sep. 20, 2022.

[56] Zhang, M., Chen, X., and Li, W. “A Hybrid Hidden Markov Model for Pipeline

Leakage Detection.” Applied Sciences, Vol. 11, No. 7, 2021, p. 3138.

https://doi.org/10.3390/app11073138.

[57] Zhang, M., Chen, X., and Li, W. “Hidden Markov Models for Pipeline Damage

Detection Using Piezoelectric Transducers.” 2020. https://doi.org/10.1007/s13349-

021-00481-0.

54
55