Research on Detection and Location of Fluid-Filled Pipeline ... - MDPI

sensors Article

Research on Detection and Location of Fluid-Filled Pipeline Leakage Based on Acoustic Emission Technology Shengshan Pan 1,2 , Zhengdan Xu 1,2 , Dongsheng Li 1,2, * and Dang Lu 1,2 1 2

*

State Key Laboratory of Coastal and Offshore Engineering, Dalian University of Technology, Dalian 116024, China; [email protected] (S.P.); [email protected] (Z.X.); [email protected] (D.L.) School of Civil Engineering, Dalian University of Technology, Dalian 116024, China Correspondence: [email protected]

Received: 17 August 2018; Accepted: 20 October 2018; Published: 25 October 2018







Abstract: Because of the inconvenience of installing sensors in a buried pipeline, an acoustic emission sensor is initially proposed for collecting and analyzing leakage signals inside the pipeline. Four operating conditions of a fluid-filled pipeline are established and a support vector machine (SVM) method is used to accurately classify the leakage condition of the pipeline. Wavelet decomposition and empirical mode decomposition (EMD) methods are initially used in denoising these signals to address the problem in which original leakage acoustic emission signals contain too much noise. Signals with more information and energy are then reconstructed. The time-delay estimation method is finally used to accurately locate the leakage source in the pipeline. The results show that by using SVM, wavelet decomposition and EMD methods, leakage detection in a liquid-filled pipe with built-in acoustic emission sensors is effective and accurate and provides a reference value for real-time online monitoring of pipeline operational status with broad application prospects. Keywords: fluid-filled pipe; acoustic emission; support vector machine; leak location

1. Introduction Pipeline transport has become the fifth largest transportation tool in modern society and offers great advantages in transporting fluid media (e.g., oil, natural gas, water). Pipelines in China have spread across large areas of land and sea and are an important approach to boosting regional economic ties. However, existing pipelines are often vulnerable to damage due to their long service life, corrosion due to aging and external load impact, leading to major accidents. Therefore, an effective structural health monitoring is important for pipeline engineering. At present, pipeline damage monitoring methods include remote field eddy current testing, ultrasonic guided waves, vibration sensors and modified ML pre-filters, distributed fiber Bragg grating sensors and others [1–4]. When a liquid-filled pipe is damaged, the fluid generates a stress wave near the leak that rapidly propagates along the wall. Hence, an acoustic emission (AE) technology is used to collect and analyze the signal and to detect the leak in the pipeline. The AE detection method for pipelines can achieve real-time, dynamic monitoring. It has been applied to detect leakage in oil and gas pipelines. Local and international scholars have conducted extensive research in this field. To improve the location accuracy of pipeline leakage sources, Wang Yongqing et al. [5] used the wavelet packet and cross correlation theories to investigate location detection. On the basis of the propagation theory of guided waves in pipes, Jingpin Jiao et al. [6] proposed a specific model of AE signal propagation in pipelines and introduced a new method to determine the leakage source. Li Zhenlin [7] proposed a novel leakage detection scheme based on kernel principal component

Sensors 2018, 18, 3628; doi:10.3390/s18113628

www.mdpi.com/journal/sensors

Sensors 2018, 18, 3628

2 of 17

analysis (kernel PCA) and the support vector machine (SVM) classifier for the recognition of the leakage level. Uncontrolled noise interference is an important factor that affects the accuracy of leakage location. Therefore, numerous scholars have started to focus on the data analysis and noise processing of AE signals from pipeline leakage and some research results have been obtained. For instance, through empirical mode decomposition (EMD) [8–10], the AE signal is initially decomposed and then a useful signal with noticeable signal features and high energy is reconstructed. Thereafter, leakage location is detected based on the reconstructed signal and accuracy is significantly improved. In addition, analysis methods based on wavelet theory [11,12] and HHT transform [13] are used in the decomposition and extraction of AE signals, which have certain effects on denoising. Gao Y et al. [14] used the generalized cross correlation (GCC) methods to find the leak location. In current research, Shama et al. [15] simulated a section of underwater oil pipeline to study the possibility of using AE technology to monitor leakage in a submarine oil pipeline. The researchers set up various defects and leaks with different flow rates in the pipelines. The results show that AE can monitor and identify leakage in underwater oil pipelines. To address the inconvenience of leak measurement in long-distance buried pipelines, Xu et al. [16] designed a guide rod that can be inserted into the soil. This rod extends directly to the surface of the pipeline. The AE information transmitted by the guide rod is used to locate the damaged area. For pipeline leakage with high-pressure gas, Mostafapour et al. [17,18] deduced the radial displacement of pipeline vibration in AE wave propagation and compared the measured signal spectrum with numerical simulation calculation, proving that the AE stress wave can theoretically be used in pipeline leakage detection. This paper attempts to solve two problems related to leakage monitoring in buried liquid-filled pipes. The first one is installing AE sensors. This study attempts to locate the AE sensors installed inside a pipeline to collect leakage signals and to verify the effectiveness of the method used. This provides a basis for subsequent built-in self-capacitive AE sensors to monitor the damage of fluid-filled pipelines. The second one is identifying leakage types and locations. A support vector machine (SVM) method is used for leakage detection in a liquid-filled pipe with a built-in AE sensor and then methods based on wavelet decomposition and EMD are utilized to effectively denoise the leakage signal. Moreover, the time-delay estimation method is employed to accurately locate the leakage source in the pipeline. 2. SVM An SVM is an algorithm developed on the basis of a statistical theory. As a powerful and practical learning machine, it is widely used in the field of pattern recognition, including face recognition, speech recognition, text classification, regression prediction and accident detection. The SVM mainly aims to obtain a maximal category boundary by minimizing structural risk. Therefore, an optimal decision boundary function is established [19]. 2.1. Optimal Hyperplane In Figure 1, the main idea of an SVM is explained in two dimensions. The hollow # and the solid are used to simulate two different samples and H is the classification plane. H1 and H2 pass through the sample data closest to H and parallel to the H plane. Their span is regarded as the classification margin. As shown in the figure, the optimal hyperplane has two main functions, namely (1) accurately classifying the training samples to ensure that the classification achieves the highest accuracy and (2) maximizing the classification interval and minimizing the confidence range of the extension community to ensure the lowest risk rate.

Sensors 2018, 18, 3628

3 of 17

Sensors 2018, 18, x FOR PEER REVIEW

3 of 17

y H1 H H2

margin=2/||ω||

x Figure Figure 1. 1. Optimal Optimal separating separating hyperplane hyperplane in in the the support vector machine (SVM).

2.2. Selection Selection of of Kernel Kernel Functions Functions 2.2. In the the SVM, SVM, according according to to specific specific engineering engineering problems, problems, the the corresponding corresponding kernel kernel functions functions can can In be selected, which not only increases the computing power of the algorithm but also improves its be selected, which not only increases the computing power of the algorithm but also improves its accuracy. Four types of most frequently used kernel functions are as follows: accuracy. Four types of most frequently used kernel functions are as follows: (1) Gaussian Gaussian radial radial basis basis kernel kernel (RBF) (RBF) kernel kernel function function (1)   2 X)l , = X exp  exp Xk2 ,, gg> 00 K ( XKl ,X − g·gk XXl l− X (1)





g parameter where norm in Euclidean space, is the parameter representing thewidth, kernelK (width, where kk is is thethe norm in Euclidean space, g is the representing the kernel Xl , X ) is the kernel function, X represents the l training sample, and X is the characteristic values the K X l , X is the kernell function, X l threpresents the lth training sample, and X isof the test sample. characteristic the test sample. (2) Linearvalues kerneloffunction (2) Linear kernel function K ( Xl , X ) = X · Xl (2)





K  Xl , X   X  Xl

(3) Polynomial kernel function

(2)

(3) Polynomial kernel function K ( Xl , X ) = [k ( X · Xl ) + m]d

(3)

, X intercept  k  X  Xand mis  the power. l  d where k is the slope of the polynomial,Km X isl the (4) Sigmoid kernel function where k is the slope of the polynomial, m is the intercept and d is the power. (4) Sigmoid kernel function K ( Xl , X ) = tanh[κ ( X · Xl ) − δ] d

(3)

(4)

K  Xthe X l     of the AE signal for pipeline   tanh   X parameters (4) l , Xcharacteristic According to the SVM principle, leakage are numerous and unordered and their difference is not noticeable. When a hyperplane cannot According to the SVM principle, the characteristic parameters of the AE signal for pipeline effectively separate the samples, the Gaussian RBF kernel function, which is the most widely-used leakage are numerous and unordered and their difference is not noticeable. When a hyperplane function, is feasible. The variable g represents the kernel width of the RBF kernel function, which cannot effectively separate the samples, the Gaussian RBF kernel function, which is the most widelydetermines the complexity of the algorithm. A penalty factor c determines the performance of the g represents used function, is feasible. The variable the kernel width which of the can RBFbekernel function, SVM classifier. Therefore, (c, g) is used as a search optimization variable, obtained using which determines the complexity of the algorithm. A penalty factor determines the performance c the grid search method [20] and corresponds to the maximal accuracy of the classification.





of the SVM classifier. Therefore, c, g is used as a search optimization variable, which can be obtained 2.3. Acoustic Emission Technology using the grid search method [20] and corresponds to the maximal accuracy of the classification. As shown in Figure 2, the AE analysis system includes three parts: a dedicated sensor for AE, a signal amplifier and a signal analysis system. The AE detection process includes: engineering 2.3. Acoustic Emission Technology material acting as a propagation medium to propagate the instantaneous elastic wave generated by As shown in Figure 2, the AE analysis system includes three parts: a dedicated sensor for AE, a the AE source to the surface of the material; an AE sensor coupled to the material’s surface receives signal amplifier and a signal analysis system. The AE detection process includes: engineering a signal source and deforms; in order to obtain a digital signal that can be computed by a computer, material acting as a propagation medium to propagate the instantaneous elastic wave generated by the received mechanical signal is converted and amplified by a piezoelectric effect mechanism and is the AE source to the surface of the material; an AE sensor coupled to the material’s surface receives finally recorded in the form of a waveform; considering the material's characteristics, the collected AE a signal source and deforms; in order to obtain a digital signal that can be computed by a computer, the received mechanical signal is converted and amplified by a piezoelectric effect mechanism and is

Sensors 2018, 18, xx FOR Sensors 2018, 2018, 18, 18, 3628 FOR PEER PEER REVIEW REVIEW Sensors

44 of of 17 17 4 of 17

finally finally recorded recorded in in the the form form of of aa waveform; waveform; considering considering the the material's material's characteristics, characteristics, the the collected collected AE parameters information of of characteristic parameters andand the the waveform information of the signal, the the purpose of inferring the AE characteristic characteristic parameters and the waveform waveform information of the the signal, signal, the purpose purpose of inferring inferring the of the AE of the object is damage statestate of the of the object is achieved. the damage damage state of AE the source AE source source of tested the tested tested object is achieved. achieved.

Figure Figure 2. 2. Schematic Schematic diagram diagram of of acoustic acoustic emission. emission.

3. Leakage Source Localization Based on Cross-Correlation Analysis 3. 3. Leakage Leakage Source Source Localization Localization Based Based on on Cross-Correlation Cross-Correlation Analysis Analysis Leakage from a liquid-filled pipe produces a continuous AE signal. Leakage detection can be Leakage Leakage from from aa liquid-filled liquid-filled pipe pipe produces produces aa continuous continuous AE AE signal. signal. Leakage Leakage detection detection can can be be performed by collecting the leakage signal and analyzing related characteristics through AE equipment. performed by collecting the leakage signal and analyzing related characteristics through AE performed by collecting the leakage signal and analyzing related characteristics through AE The leakage location can be determined by installing a pair of sensors inside a pipe. Figure 3 shows equipment. equipment. The The leakage leakage location location can can be be determined determined by by installing installing aa pair pair of of sensors sensors inside inside aa pipe. pipe. the two sensors arranged on both sides of the leak. Figure Figure 33 shows shows the the two two sensors sensors arranged arranged on on both both sides sides of of the the leak. leak.

Figure Figure 3. 3. Schematic Schematic diagram diagram for for detection. detection. Figure 3. diagram for detection.

Therefore, source is as follows: Therefore, the the positioning positioning formula formula of of the the pipeline pipeline leakage leakage source sourceis isas asfollows: follows:

tt Dν·∆t D −D ,,  tt1t11 d =dd  , ∆t =tt t2tt− 22  2 22

(5) (5)

relevant parameters parametersin inthe theformula formulaare areshown shownin thefigure, figure,where whereν isis The the propagation The relevant in the formula are shown ininthe the figure, where isthe thepropagation propagation speed speed the AE waveon onthe theliquid-filled liquid-filledsteel steel pipe. Therefore, obtaining unknown quantities, namely, of pipe. Therefore, obtaining two unknown quantities, namely, the of the AE wave wave on the liquid-filled steel pipe. Therefore, obtaining twotwo unknown quantities, namely, the the speed ν and the time difference ∆t, location. In this study, the propagation speed   t speed and the time difference , is the key to leak study, the propagation speed and the time difference , is the key to leak location. In this study, the propagation speed of the AE signal signalis calculatedvia viathe the famous Nielsen–Hsu lead-cut experiment [21–23]. famous Nielsen–Hsu lead-cut experiment [21–23]. of the the AE AE signal isiscalculated calculated via the famous Nielsen–Hsu lead-cut experiment [21–23]. To t we To obtain ,, we use time-delay estimation method based on cross correlation analysis, which To obtain obtain ∆t, weuse useaaatime-delay time-delayestimation estimationmethod methodbased basedon oncross crosscorrelation correlationanalysis, analysis, which which is widely used in processing continuous AE signals. Assuming that the AE signals collected by the is is widely widely used used in in processing processing continuous continuous AE AE signals. signals. Assuming Assuming that that the the AE AE signals signals collected collected by by the the sensors on both sides are x (t) and y(t), the corresponding mathematical model is as follows: sensors sensors on on both both sides sides are are x t and and y t ,, the the corresponding corresponding mathematical mathematical model model is is as as follows: follows:





x (t) = s(t) + n1 (t), y(t) = αs(t − τ ) + n2 (t)

(6) (6) (6) where s(t) is the leakage source signal, α is the attenuation factor, τ is the delay time and n1 (t), n2 (t) s t isis noise. are the ambient where the is where the leakage leakage source source signal, signal,  is is the the attenuation attenuation factor, factor, is the the delay delay time time and and Within a certain integral time T, the cross correlation formula is as follows: n11 t 、n22 t are are the the ambient ambient noise. noise. Z T 1 correlation Within T, Within aa certain certain integral integral time time R T, the cross formula is as as follows: follows: ˆ xythe = correlation αs(t)s(t formula − τ )dt is (7) (τ )cross T 0

x  t   s  t   n11t  , y t    s t     n22 t 

  





R Rxyxy   

11 TT  ss ttss tt   dt dt TT 00

(7) (7)

Sensors 2018, 18, 18, 3628 Sensors 2018, x FOR PEER REVIEW

5 of5 17 of 17



As shown in the cross correlation formula, the two AE signals are a function of the delay time . As shown in the cross correlation formula, the two AE signals are a function of the delay time τ. In the cross correlation diagram, the position of the peak point implies the appearance of the maximal In the cross correlation diagram, the position of the peak point implies the appearance of the maximal correlation coefficient two signals and corresponding the signal difference. correlation coefficient of of thethe two signals and thethe corresponding τ0 is the time time difference. 0 issignal

4. 4. Wavelet Decomposition and EMD Wavelet Decomposition and EMD ByBy directly using the detected original signal waveform to calculate cross correlation directly using the detected original signal waveform to the calculate the crosscoefficient, correlation a delay time with larger deviation is always obtained. Therefore, choosing an appropriate method to coefficient, a delay time with larger deviation is always obtained. Therefore, choosing an appropriate effectively denoise and optimize the original signal waveform is also key to the positioning accuracy. method to effectively denoise and optimize the original signal waveform is also key to the positioning In accuracy. the denoising of continuous AEcontinuous signals, twoAE typical methods are introduced thisintroduced section, namely, In the denoising of signals, two typical methodsinare in this wavelet decomposition and EMD. section, namely, wavelet decomposition and EMD. 4.1. Wavelet Decomposition and Threshold Denoising 4.1. Wavelet Decomposition and Threshold Denoising Wavelet analysis can be used to characterize the local time–frequency of nonstationary random Wavelet analysis can be used to characterize the local time–frequency of nonstationary random signals with noise. This analysis not only realizes the function of removing a large amount of signals with noise. This analysis not only realizes the function of removing a large amount of interference but also retains useful signal features [24]. The steps of wavelet threshold denoising interference but also retains useful signal features [24]. The steps of wavelet threshold denoising are are as follows: as follows: (1) Wavelet decomposition of signals (1) Wavelet decomposition of signals Using a four-layer decomposition as an example, the wavelet decomposition process is graphically Using a four-layer decomposition as an example, the wavelet decomposition process is represented. As shown in Figure 4, A1–A5 represent the approximation functions of the layers, whereas graphically represented. As shown in Figure 4, A1–A5 represent the approximation functions of the D1–D5 represent the detail functions of the layers. layers, whereas D1–D5 represent the detail functions of the layers.

Figure 4. Four-layer wavelet decomposition process forfor signals. Figure 4. Four-layer wavelet decomposition process signals.

(2)(2) Threshold quantization Threshold quantization After the original signal is is decomposed byby thethe wavelet, anan appropriate threshold is is used to to After the original signal decomposed wavelet, appropriate threshold used quantize thethe high-frequency coefficients (D1–D5) of of each layer to to reduce thethe noise. The MATLAB quantize high-frequency coefficients (D1–D5) each layer reduce noise. The MATLAB software provides three threshold methods, that is, forced denoising, default threshold software provides three threshold methods, that is, forced denoising, default threshold and and given given threshold. threshold. (3)(3) Signal reconstruction Signal reconstruction   Through threshold quantization, the decomposition coefficients c j,n , d j,n of each layer are Through threshold quantization, the decomposition coefficients c j ,n , d j , n of each layer are used to reconstruct the signal using the MATLAB algorithm. used to reconstruct the signal using the MATLAB algorithm. 4.2. EMD 4.2. EMD The EMD theory holds that any signal can be composed of a series of oscillating components with The EMDcalled theoryintrinsic holds that anyfunctions signal can be composed of a series oscillating components different scales, mode (IMFs). IMFs should meet of two basic conditions as with different scales, called intrinsic mode functions (IMFs). IMFs should meet two basic conditions follows [25]: as (1) follows [25]: number of IMF extreme points is equal to the number of zeroes, or at most, there is The total (1) The total number of IMF extreme points is equal to the number of zeroes, or at most, there is a deviation. a deviation. (2) At any point in the IMF, the average value of the maximal and minimal points corresponding (2) At and any lower point in the IMF, the average value of the maximal and minimal points corresponding to the upper envelopes, respectively, is equal to zero. to the upper and lower envelopes, respectively, is equal to zero. Based on these principles, the initial signal can be reconstructed from all decomposed IMF

  

components and the final residual component

rn  t  , as follows:

Sensors 2018, 18, 3628

6 of 17

Based on these principles, the initial signal can be reconstructed from all decomposed IMF components and the final residual component rn (t), as follows:

Sensors 2018, 18, x FOR PEER REVIEW

6 of 17

n

X (t) = ∑n hi (t) + rn (t) X  t  i= 1 hi  t   rn  t 

(8) (8)

i 1

In this study, we initially use the EMD method to decompose the AE signal and obtain the IMFs this study, we initially thewe EMD method to decompose AE signal andfor obtain the IMFs andIn residual components anduse then select a useful signal with the evident features reconstruction, and residual components and then we select a useful signal with evident features for reconstruction, thus achieving an effective denoising. thus achieving an effective denoising. 5. Experiment on Pipeline Leakage Detection Based on SVM 5. Experiment on Pipeline Leakage Detection Based on SVM 5.1. Experimental Scheme 5.1. Experimental Scheme model for the operation of a liquid-filled pipe is constructed, as shown in First, a simplified Figure 5.a simplified model for the operation of a liquid-filled pipe is constructed, as shown in Figure 5. First,

Figure5.5.Schematic Schematicdiagram diagramofofpipeline pipelineleakage leakagetest testplatform. platform. Figure

TheAE AEtesting testingequipment equipmentconsists consistsmainly mainlyofofthe thefollowing followingcomponents: components:eight-channel eight-channelMicro-II Micro-II The DigitalAE AE System equipment produced by PAC for signal acquisition and analysis; SR40M AE Digital System equipment produced by PAC for signal acquisition and analysis; SR40M AE sensor sensor produced by ShengHua Technology Company used for collecting signals; an intelligent AE produced by ShengHua Technology Company used for collecting signals; an intelligent AE preamplifier that provides 20 dB, 40 dB and 60 dB gain; a bandwidth of 10 kHz–1.2 MHz that effectively preamplifier that provides 20 dB, 40 dB and 60 dB gain; a bandwidth of 10 kHz–1.2 MHz that amplifies the originalthe AEoriginal signal; an such asdevice a computer or keyboard. Table effectively amplifies AEattached signal; device an attached such screen as a computer screen or 1 shows the parameter settings of the AE instruments. keyboard. Table 1 shows the parameter settings of the AE instruments. Table1.1.Parameter Parametersettings settingsofofacoustic acousticemission emission(AE) (AE)instruments. instruments. Table Hit Length

Threshold

Sampling Rate

Preamplifier Gain

PDT/µs

HDT/µs

Hit Length

Threshold

Sampling Rate

Preamplifier Gain

PDT/μs

HDT/μs

1024

20/35 dB

1 MHz

40 dB

200

800

1024

20/35 dB

1 MHz

40 dB

200

800

HLT/µs HLT/ 1000 μs

1000

The main pipe used is a galvanized steel pipe with a length of 2 m, outer diameter of 165 mm and main pipe galvanized steelthe pipe withpressure a lengthinofthe 2 m, outer diameter by of adjusting 165 mm wallThe thickness of 4 used mm. is Inathe experiment, water pipe is regulated and wall thickness of 4 mm. In the experiment, the water pressure in the pipe is regulated by adjusting the ball valve at both ends. The laboratory model is shown in Figure 6. The waterproof sensor is the ball valve placed at bothinside ends.the Thepipe laboratory showntoin Figurethe 6. transmission The waterproof sensor is innovatively to detectmodel the AEissignal, explore characteristics innovatively the pipe detect the AE to explore the transmission characteristics of the signalplaced in theinside pipe when thetopipe leaks andsignal, to explore the application of wireless sensors in ofthe thefuture. signal In in addition, the pipe when the pipe leaks and to explore the application of wireless sensors in the we propose the use of mobile robots for installation in practical application. future. In addition, weapropose usemm of is mobile robots forflange installation practical application. Therefore, a hole with diameterthe of 18 reserved on the plate atinboth ends of the pipe for Therefore, a hole with a diameter of 18 mm is reserved on the flange plate at both ends of the pipe installation of the AE sensor. After installation, a leakage test is performed using a mud plug. for installation of the AE sensor. After installation, a leakage test is performed using a mud plug.

Sensors 2018, 18, 3628 Sensors 2018, 18, x FOR PEER REVIEW

7 of 17 7 of 17

Figure Figure 6. 6. Laboratory Laboratory model. model.

According of of a leakage signal in ainpipeline is low; thus,thus, the According to tothe theliterature, literature,the thepeak peakfrequency frequency a leakage signal a pipeline is low; SR40M sensor of theofnarrowband low frequency is wellismatched. The related parameters of the sensor the SR40M sensor the narrowband low frequency well matched. The related parameters of the are listed Tablein2.Table 2. sensor areinlisted Table 2. 2. Technological Technological parameters parameters of of the the SR40M SR40M sensor. sensor. Table Size/mm Size/mm

Temperature/◦ C Temperature/°C

Bandwidth/kHz Bandwidth/kHz

× 36.8  22Φ22 × 36.8

−20–120 −20–120

15–70 15–70

Resonant Peak Resonant Peak Frequency/kHz Sensitivity/dB Frequency/kHz Sensitivity/dB 40 40

>75>75

The processes processes of ofthe theleakage leakagedetection detectionalgorithm algorithm liquid-filled pipes based on SVM are The forfor liquid-filled pipes based on SVM are the the following: following: (1) Collect Collect the the characteristic characteristic parameters parameters of of the the pipeline pipeline under under different different conditions conditions using using an an AE AE (1) instrument, select appropriate parameters and normalize them and set the normalization interval instrument, select appropriate parameters and normalize them and set normalization interval to [0, 1 . ] to 0, 1 . (2) Define labels for different states of the pipeline. For instance, set the pipeline standing state to (2) Defineoperation labels forstate different states the pipeline. 1, the normal to 2 and theof leakage state toFor 3. instance, set the pipeline standing state to 1, the normal operation state to 2 and the leakage state to 3. and test sets and choose a suitable (3) Divide all sample data into two parts, namely training (3) Divide all sample data into two parts, namely training and test sets and choose a suitable kernel function and corresponding parameters according to actual problems. kernel andthe corresponding parameters to actual problems. (4)function Complete testing process, build theaccording classification prediction model using the determined (4) Complete the testing process, build the classification prediction using determined kernel functions and parameters and establish the classification boundarymodel function f k (the x ) to determine the optimal hyperplane. kernel functions and parameters and establish the classification boundary function f k x to (5) Use optimal model constructed in Step (4) to perform diagnostic tests and input the determine thethe optimal hyperplane. characteristic parameters to complete the classification prediction of the test set. tests and input the (5) Use the optimal model constructed in Step (4) to perform diagnostic

 

 

characteristic parameters to complete the classification prediction of the test set. 5.2. Experimental Process and Result Analysis

A total of five conditions are set up in the experiment, as listed in Table 3. 5.2. Experimental Process and Result Analysis A total of five conditions are set up in the experiment, as listed in Table 3.

Sensors 2018, 18, 3628 Sensors 2018, 18, x FOR PEER REVIEW

8 of 17 8 of 17

Table Table 3. Experimental Experimental cases. cases.

Case Case Pipeline StatusStatus Pressure/MPa Pipeline Pressure/MPa Sample SampleNumber Number C1 resting — 600 C1 resting — 600 C2 operational 600 C2 operational — — 600 C3 leakage 0.2 600 C3 leakage 0.2 600 C4 small pressure leakage 0.05 600 C4 small pressure leakage 0.050.3 600 C5 large pressure leakage 600 C5 large pressure leakage 0.3 600

Distance from Leakage/mm Distance from Leakage/mm 400 400 400 400 400 400 400 400 400 400

According to the statistical mean and standard deviation of the collected data, several AE According to the statistical mean and standard deviation of the collected data, several AE parameters with distinct distinct discrimination discriminationand andgood goodstability stabilitywere wereselected, selected,such suchas asenergy, energy,amplitude, amplitude, parameters with ASL, average frequency and RMS. ASL, average frequency and RMS. The experimental experimental conditions conditions in in Table Table 33 were were combined combined to to complete complete the the following following The classification experiments. classification experiments. (1) Two-way Two-way classification: classification: resting resting (C1) (C1) and and leakage leakage (C3) (C3) (1) Pipeline resting and leakage are clearly different. Therefore, thethe linear kernel function can be used Pipeline resting and leakage are clearly different. Therefore, linear kernel function can be to train the SVM model. First, 100 datasets for each case are selected randomly to form the training used to train the SVM model. First, 100 datasets for each case are selected randomly to form the set, which used is toused trainto the classification model.model. Then, Then, the remaining 500 samples for each case training set,iswhich train the classification the remaining 500 samples for each are imported into the classification model for test verification. The classification results are shown in case are imported into the classification model for test verification. The classification results are Figure in 7. Figure 7. shown 3 Actual classification Recognition classification

Lable

2.5 2

1.5 1

0

200

400

Sample

600

800

1000

Figure Figure7. 7.Classification Classificationresults resultsof ofthe thetest testsample sample (Two-way). (Two-way).

As shownininthe the results of Figure 7, there are erroneous seven erroneous judgments for the leakage As shown results of Figure 7, there are seven judgments for the leakage conditions conditions and the recognition accuracy rate is 99.3%.based Therefore, based on statistical training data, and the recognition accuracy rate is 99.3%. Therefore, on statistical training data, the pipeline the pipeline can be accurately detected using the SVM method in the event of a leak. can be accurately detected using the SVM method in the event of a leak. (2) (2) Three-way Three-way classification: classification: resting resting (C1), (C1), normal normal operation (C2) and leakage (C3) First, 200 datasets of each case are randomly First, 200 datasets of each case are randomly selected selected to to form form the the training training set. set. Then, Then, the the simplest simplest linear kernel is used to train the model. The recognition results are not ideal, with a rate linear kernel is used to train the model. The recognition results are not ideal, with a rate of of accurate accurate classification classification of of only only 87.5%. 87.5%. Therefore, the RBF Therefore, the RBF kernel kernel function function is is used used to construct the SVM model. Meanwhile, Meanwhile, the the grid grid g. search method isisused usedtotofind find the corresponding penalty parameter and parameter kernel parameter search method the corresponding penalty parameter c andckernel g. The result of parameter selection is shown in Figure 8. The SVM model is established through the training set. The result of parameter selection is shown in Figure 8. The SVM model is established through the By inputting theinputting test sample model, each condition can be classifiedcan andbedetermined training set. By theinto testthe sample into theoperating model, each operating condition classified and determined then the operating status the pipeline can identification results are shown and and then the of operating status of be thedetected. pipeline The can test be detected. The test identification in Figure results are9.shown in Figure 9.

Sensors 2018, 18, 3628

9 of 17

Sensors Sensors 2018, 2018, 18, 18, xx FOR FOR PEER PEER REVIEW REVIEW Sensors 2018, 18, x FOR PEER REVIEW

99 of of 17 17 9 of 17

Best Best c=0.020617 c=0.020617 g=16 g=16 CVAccuracy=100% CVAccuracy=100% Best c=0.020617 g=16 CVAccuracy=100%

Best Best c=0.020617 c=0.020617 g=16 g=16 CVAccuracy=100% CVAccuracy=100% c=0.020617 g=16 CVAccuracy=100%

88 Best 8 97 97 66 97 6

100 100 100 90 90 90 80 80 80 70 70 70 60 60 60 50 50 50 40 40 40 30 30 30

Accuracy(%) Accuracy(%) Accuracy(%)

44 4

log2g log2g log2g

22 2

97 97 00 97 0 92.5 92.5 92.5 -2 -2 -2

997 9997722.5 92.5 .5

-4 -4 -4

55 5

00

0 log2g log2g log2g

-5 -5 -5

-5 -5 -5

55 5

00

(a) 3D illustration (a) (a) 3D 3D illustration illustration

88 88 88

-6 -6 -6 -8 -8-8 -8-8 -8

0 log2c log2c log2c

-6 -6 -6

-4 -4 -4

-2 -2 -2

997 9972 9722..585 .8588 88 00 22 log2c 0 2 log2c

log2c

44 4

66 6

88 8

(b) Contour map (b) (b) Contour Contour map map

Figure Figure 8. 8. Results Results of of parameter parameter selection. selection. Figure 8. Results of parameter selection. 3 3 3

Case1 Case1 Case1 Case2 Case2 Case2 Case3 Case3 Case3

Lable Lable Lable

2.5 2.5 2.5 2 2 2

1.5 1.5 1.5 1 10 10 0

50 50 50

100 100 100

150 150 150

200 200

250 250

300 300

350 350

200 250 300 350 Sample Sample Sample Figure 9. Test results based on Figure on the the model model with with the the RBF RBF kernel kernel function. function. Figure 9. 9. Test Test results results based based Figure 9. Test results based on on the the model model with with the the RBF RBF kernel kernel function. function.

400 400 400

The test results of the RBF kernel function model for classification are quite accurate and the The The test test results results of of the the RBF RBF kernel kernel function function model model for for classification classification are are quite quite accurate accurate and and the the correct recognition rate for each case exceeds 98%. Unlike the recognition result of the linear kernel correct Unlike the the recognition recognition result result of correct recognition recognition rate rate for for each each case case exceeds exceeds 98%. 98%. Unlike Unlike the recognition result of the the linear linear kernel kernel function, the integrated recognition accuracy rate is improved by 11%. function, function, the the integrated integrated recognition recognition accuracy accuracy rate rate is is improved improved by by 11%. 11%. (3) Four-way classification: resting (C1), normal operation (C2), small pressure leakage (C4) and (3) Four-way classification: resting (C1), normal operation (C2), (3) Four-way classification: resting (C1), normal operation (C2), small small pressure pressure leakage leakage (C4) (C4) and and large pressure leakage (C5) large large pressure pressure leakage leakage (C5) (C5) Generally, the larger SVM model will be. A total of Generally, the training dataset is, the more accurate the SVM model will be. A of larger the thetraining trainingdataset datasetis, is,the themore moreaccurate accuratethe the SVM model will A total Generally, the the larger larger the training dataset is, the more accurate the SVM model will be.be. A total total of 100 and 300 data training models are selected for each case to investigate the effect of the sample 100 and 300 data training models are of and 300 data training models areselected selectedfor foreach eachcase caseto investigatethe theeffect effect of of the 100100 and 300 data training models are selected for each case totoinvestigate investigate the effect of the sample sample number on the classification results. Figures 10 and 11 show the results. The SVM model is also number on the classification results. Figures 10 11 the The SVM is results. Figures 10 and 11 show the results. The SVM is also trained number on onthe theclassification classification results. Figures 10 and and 11 show show the results. results. The model SVM model model is also also trained using the RBF kernel function. trained using the RBF kernel function. using the RBF kernel function. trained using the RBF kernel function. By comparing the above By two results, we can see that with the increase in training data samples, above two two results, results, we we can can see see that that with with the theincrease increase in in training training data data samples, samples, By comparing comparing the the above above two results, we can see that with the increase in training data samples, the classification accuracy rate improves significantly. Moreover, the four cases can be clearly divided the classification accuracy rate improves significantly. Moreover, the be clearly significantly. Moreover, the four cases can the classification accuracy rate improves significantly. Moreover, the four cases can be clearly divided divided and the number of undetermined cases is significantly reduced. and the number of undetermined cases is significantly reduced. and the number of undetermined cases is significantly reduced. 5 5 5

Lable Lable Lable

4 4 4

Case1 Case1 Case1 Case2 Case2 Case2 Case4 Case4 Case4 Case5 Case5 Case5

3 3 3 2 2 2 1 10 10 0

50 50 50

100 100 100

150 150 150

200 200 200

250 250

300 300

250 300 Sample Sample Sample

350 350 350

400 400 400

450 450 450

500 500 500

Figure Figure 10. 10. Classification Classification results results of of the the test test sample sample using using SVM SVM (100 (100 data data for for each each case). case). Figure using SVM SVM (100 (100 data data for for each each case). case). Figure 10. 10. Classification Classification results results of of the the test test sample sample using

Sensors 2018, 18, 3628

10 of 17

Sensors 2018, 18, x FOR PEER REVIEW Sensors 2018, 18, x FOR PEER REVIEW

10 of 17 10 of 17

In addition, addition, the the classification classification results results of of the the proposed proposed SVM SVM method method were were also also compared compared with with the the In In addition, the classification results ofin the proposed SVM method were also compared with the widely-used BP neural network. As shown Figures 11 and 12, the results of proposed SVM method widely-used BP neural network. As shown in Figures 11 and 12, the results of proposed SVM method widely-used BP same neuralaccuracy network.with As shown in Figures 11 and 12, the results of proposed SVM method nearly have the the BP neural nearly have same accuracy with BP neural network. network. nearly have the same accuracy with BP neural network. 5 5 4 4

Lable Lable

Case1 Case1 Case2 Case2 Case4 Case4 Case5 Case5

3 3 2 2 1 10 0

50 50

100 100

150

150 Sample Sample

200 200

250 250

300 300

Figure Figure 11. 11. Classification Classification results results of of the the test test sample sample using using SVM SVM (300 (300 data data for for each each case). case). Figure 11. Classification results of the test sample using SVM (300 data for each case).

Figure 12. Classification results of the test sample using BP neural network. Figure 12. Classification results of the test sample using BP neural network.

6. Decomposition and and EMD EMD 6. Pipeline Pipeline Leakage Leakage Location Location Based Based on on Wavelet Wavelet Decomposition 6. Pipeline Leakage Location Based on Wavelet Decomposition and EMD After detecting detecting a leak in the pipeline, a method should be adopted to further determine the After detecting a leak in the pipeline, a method should be adopted to further determine the source of the damage. This This section section addresses addresses the the problem problem of of locating locating the the source source of of aa pipeline pipeline leak. leak. source of the damage. This section addresses the problem of locating the source of a pipeline leak. The propagation speed of AE waves on liquid-filled pipes was obtained through a lead-cutting The propagation speed of AE waves on liquid-filled pipes was obtained through a lead-cutting experiment. The final calculated average average wave wave speed speed is is 3446 3446 m/s. m/s. In Inpractical practical applications, applications, aa database database experiment. The final calculated average wave speed is 3446 m/s. In practical applications, a database of wave speeds for different materials must be facilitated for for the the application application of of the the method. method. of wave speeds for different materials must be facilitated for the application of the method. The pressure in the pipe is stable at 0.2 MPa and sensors S1 and S2 are placed on both sides of The pressure in the pipe is stable at 0.2 MPa and sensors S1 and S2 are placed on both sides of the leak in the pipe. Liquid leakage leakage detection detection is is performed performed after after the the pressure pressure is stable and the signal the leak in the pipe. Liquid leakage detection is performed after the pressure is stable and the signal is collected using an AE instrument. Moreover, Moreover, five five sensor sensor placement placement conditions conditions are are set set up up for for the the is collected using an AE instrument. Moreover, five sensor placement conditions are set up for the study. The distances (d1–d2) from S1 and S2 to the hole are 400–600, 500–800, 650–800, 700–800 and study. The distances (d1–d2) from S1 and S2 to the hole are 400–600, 500–800, 650–800, 700–800 and 800–950 mm. 800–950 mm. 6.1. Wavelet Decomposition Denoising 6.1. Wavelet Decomposition Denoising According to the wavelet denoising denoising theory, the wavelet denoising analysis analysis of the leaking AE According According to the wavelet denoising theory, the wavelet denoising analysis of the leaking AE signal is performed using a DB6 wavelet, five-level decomposition and default default threshold threshold denoising. denoising. signal is performed using a DB6 wavelet, five-level decomposition and default threshold denoising. The AE signal is decomposed decomposed into five scale wavelets, as shown in Figure 13, where a1–a5 The AE approximation signal is decomposed into five scale as shown in Figure 13, whereina1–a5 called thewavelets, low-frequency coefficients) decomposed each represent the the approximationfunctions functions(also (also called the low-frequency coefficients) decomposed in represent the approximation functions (also called the low-frequency coefficients) decomposed in layer layer and d1–d5 the detail functions (also called the high-frequency coefficients) decomposed each and represent d1–d5 represent the detail functions (also called the high-frequency coefficients) each layer and d1–d5 represent theexist detail functions (also coefficients. called the high-frequency coefficients) in each layer. signals usually in high-frequency As shown As in the figure, d1 decomposed inNoise each layer. Noise signals usually exist in high-frequency coefficients. shown in the decomposed in each layer. Noise signals usually exist in high-frequency coefficients. As shown in the and d2 have noticeable noise signal characteristics. Figure 14 showsFigure the spectrum corresponding to figure, d1 and d2 have noticeable noise signal characteristics. 14 shows the spectrum figure, d1 and d2 coefficient have noticeable noise signal characteristics. Figure 14 shows thethe spectrum the decomposition of each layer after the final wavelet decomposition. Using default corresponding to the decomposition coefficient of each layer after the final wavelet decomposition. corresponding to the decompositioncoefficient coefficientfor of each eachlayer layerisafter the final wavelet decomposition. threshold a high-frequency initially quantized with thequantized threshold Using the method, default threshold method, a high-frequency coefficient for each layer is initially Using the default threshold method, a high-frequency coefficient for each layer is initially quantized with the threshold value, then the wavelet reconstruction is performed with a low-frequency with the threshold value, then the wavelet reconstruction is performed with a low-frequency coefficient a5 of the last layer and the signal after noise removal can be finally obtained. coefficient a5 of the last layer and the signal after noise removal can be finally obtained.

Sensors 2018, 18, 3628

11 of 17

value, then18,the wavelet reconstruction is performed with a low-frequency coefficient a5 of the last Sensors 2018, x FOR PEER REVIEW 11 of 17 layer and the signal after noise removal can be finally obtained. Sensors 2018, 18, x FOR PEER REVIEW 11 of 17

s1

0.1

s1

00.1 -0.1 0 0.05 -0.1

d5

d5

0.05 0

-0.05 0 0.1 0

d4

d4

-0.05 0.1 0

-0.1 0.05-0.1

d3

0.05

d3

0

0

-0.05 -0.05 0.05

d2

0.05

d2

0 0

d1

-0.05 -0.05 0.05 0.05

d1

0 0

-0.05 -0.05 0 0

0.50.5

11

1.5 1.5

2

2.5 2.5

Time/ms Time/ms

33

3.53.5

4 4

4.5 4.5

5

5

Figure13. 13.Details Details of of signal Figure 13. signal wavelet decomposition. Figure Details of signalwavelet waveletdecomposition. decomposition.

d1

d1

0.2 0.1

0.2 0.1 0

d2

d2

0 2 2 1

1

0

10 0 40 0 40 20

d4

d4

d3

d3

0 20 20 10

20 0 10

d5

0 10 5

d5

5 0 2

a5

a5

0 1 2 0

1 0

0

0

50

50

100

100

150

150

200

250

300

Frequency/kHz

200

250

300

350

350

400

400

450

500

450

500

Frequency/kHz Figure 14. Spectrum of each layer after the signal wavelet decomposition. Figure 14. of after the signal wavelet decomposition. Figure 14. Spectrum Spectrum of each each layer layer afterthe thecross signalcorrelation wavelet decomposition. On the basis of the cross-correlation theory, function of the AE signal detected by the two sensors is obtained. The point of greatest cross correlation is the cross coordinate On the basis of the cross-correlation theory, the cross correlation function of the AE signal corresponding to the peak value in the function graph and, thus, the time difference t . Graphs (a) detected by the two sensors is obtained. The The point point of of greatest greatest cross correlation is the cross coordinate and (b) in Figure 15 represent the original signals and the cross correlation function diagram, corresponding to the peak value in the function graph thus, thethe time difference ∆t. Graphs (a) and t . Graphs corresponding thewavelet peak value in theoffunction graphand, and, thus, time difference (a) respectively, to after denoising the two sensors. As shown in the graphs, the timedifference (b) in Figure 15 represent the original signals and the cross correlation function diagram, respectively, andbetween (b) in the Figure 15 represent the original signals and the cross correlation function diagram, original signals reaching the sensor is t = 0.045 ms and the time difference obtained

after wavelet denoising of the two sensors. As shown in As theshown graphs, the time difference between the respectively, wavelet denoising sensors. graphs, the time difference t into after noise after reduction is  . Bytwo substituting D,  andinthe Equation (5), we can t  0.056 of msthe original signals reaching the sensor is ∆t = 0.045 ms and the time difference obtained after noise between the original signals reaching the sensor is t = 0.045 ms and the time difference obtained calculate the distance between S1 and the leak point d1 to locate the leak source. reduction is ∆t = 0.056isms. substituting D, substituting ν and ∆t into Equation (5), tweinto can Equation calculate the  and after noise (5),distance we tBy  0.056 ms . By The reduction experimental results of condition 1 are listed inD, Table 4. As presented in the table, the can between S1 and the leak point d1 to locate the leak source. location accuracy of the leakage source evidently improves after the original signal is denoised by calculate the distance between S1 and the leak point d1 to locate the leak source. The experimental results of condition 1 are listed ininTable 4.5.As presented in the table, the location the wavelet. The results of other conditions are listed Table The positioning results at various The experimental results of condition 1 are listed in Table 4. As presented in the table, the accuracy of the source evidently after the original signal is5%, denoised by the wavelet. distances areleakage relatively accurate and improves the relative errors are thus meeting the by location accuracy of the leakage source evidently improves afterless thethan original signal is denoised Therequirements results of other conditions are listed in Table 5. The positioning results at various distances engineering practice. the wavelet. Theofresults of other conditions are listed in Table 5. The positioning results at various are relatively accurate and the relative errors are less than 5%, thus meeting the requirements of distances are relatively accurate and the relative errors are less than 5%, thus meeting the engineering practice. requirements of engineering practice.

Sensors 2018, 18, 3628

12 of 17

Sensors 2018, 18, x FOR PEER REVIEW

12 of 17

0.3

0.4 X: 0.045 Y: 0.2977

0.3 0.2

0.1

互相关函数

互相关函数

0.2

0 -0.1 -0.2

0 -0.1 -0.2

-0.3 -0.4 -0.5

0.1

X: 0.056 Y: -0.3228

-0.3

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

时间/ms (a) Original signal

0.3

0.4

-0.4 -0.5

0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

(b) Signal时间/ms after denoising

0.4

0.5

Figure Figure 15. 15. Cross-correlation Cross-correlation function function diagram diagram of of signals signals S1 S1 and and S2. S2. Table 4. Case 1: Comparison of the results of locating the leakage source. Table 4. Case 1: Comparison of the results of locating the leakage source.

Result

Result

Original signal Original signal Wavelet denoising Wavelet denoising

d1/mm

d1/mm

400 400 400 400

d2/mm d2/mm

600

600 600 600

Δt/ms

∆t/ms

0.045

0.045 0.056 0.056

Calculated

Calculated Value: Value: d1 d1

422.47

422.47 403.51 403.51

Absolute

Absolute ErrorError

22.47

Relative

Relative Error Error

5.62%

22.47 3.51 3.51

5.62% 0.88%0.88%

Table Table 5. 5. Results Results of of locating locating the the leakage leakage source source using using the the wavelet wavelet denoising. denoising.

Experiment Experiment d1/mm d1/mm d2/mm d2/mm NumberNumber 1 400 400 600 1 600 2 800 2 500 500 800 3 800 3 650 650 800 4 700 800 4 700 800 5 800 950 5 800 950

Δt/ms ∆t/ms 0.056 0.056 0.079 0.079 0.049 0.049 0.017 0.017 0.023 0.023

Calculated Absolute Calculated Absolute RelativeRelative d1 Error Error Error Error Value:Value: d1 403.51 403.51 513.88 513.88 640.57 640.57 720.71 720.71 835.37

835.37

3.51 13.88 9.43 20.71 35.37

3.51 13.88 9.43 20.71 35.37

0.88% 2.78% 1.45% 2.96% 4.42%

0.88% 2.78% 1.45% 2.96% 4.42%

6.2. EMD of AE signal 6.2. EMD of AE signal The original nonstationary AE signal can be decomposed in time and frequency domains using The original nonstationary AE signal can be decomposed in time and frequency domains using the EMD method and then the first several IMFs with relatively high energy and correlation with the the EMD method and then the first several IMFs with relatively high energy and correlation with the original signal are reconstructed. original signal are reconstructed. Using condition 1 (d1-400, d2-600) as an example, the original AE signal of the pipeline leakage Using condition 1 (d1-400, d2-600) as an example, the original AE signal of the pipeline leakage is decomposed using the EMD. Each stage waveform after decomposition and its corresponding is decomposed using the EMD. Each stage waveform after decomposition and its corresponding frequency spectrum are shown in Figure 16. The original signal is decomposed into nine IMFs and one frequency spectrum are shown in Figure 16. The original signal is decomposed into nine IMFs and residual component Rn . In the time domain waveforms, the amplitude of the IMF gradually decreases. one residual component Rn. In the time domain waveforms, the amplitude of the IMF gradually Therefore, IMF components with lower signal amplitudes and lower frequencies are negligible and decreases. Therefore, IMF components with lower signal amplitudes and lower frequencies are these components are possibly caused by a noisy environment. negligible and these components are possibly caused by a noisy environment. The energy of each layer component E and the energy ratio of this layer R MFi can be obtained The energy of each layer componentMFiEMFi and the energy ratio of this Ilayer RIMFi can be to further quantify the IMF using the formula as follows: obtained to further quantify the IMF using the formula as follows: 1 N −N11 EE I MFi 22 E I MFi =  1∑hhi (i m 100% I MFI EIMFi RIMFI= 9 9 IMFi ×  100%  m) , R， NN m=m0 0 ∑ EEI MFi IMFi

(9) (9)

E I MFi E R I MFI = × 100% 9 RIMFI  9 IMFi  100% E ∑ I MFi i = 1 EIMFi

(10) (10)

i =i11

i 1

where hi (m) is the magnitude of the ith and the IMF component and N is the length of the signal. where hi m is the magnitude of the ith and the IMF component and N is the length of the signal.

 

Sensors 2018, 18, x FOR PEER REVIEW

13 of 17

IMF1

Sensors 2018, 18, 3628 0.5

10

Sensors 2018, 18, x FOR PEER REVIEW

5 0 40 10

IMF2 IMF1

0

-0.5 0.2 0.5

5 20 0 405

0 0

20

-0.1-0.2 0.02 0.1 0 0

0 45

Rn IMF9 IMF8 IMF7 IMF6 IMF5 IMF4 IMF3 IMF9 IMF8 IMF7 IMF6 IMF5 IMF4 IMF3 IMF2

0 0

-0.2-0.5 0.1 0.2

2 0 40 2 2 1 0 20 2 1 1 0 20 2 1 1 0 20 41 20 4 0 22

-0.1 -0.020.02 0.02 0 0 -0.02 -0.020.02 0.01 0 0 -0.02 -0.010.01 0.01 0 0 -0.01 -0.010.01 0.01 0

0 -0.01 0.01 -0.01 0.002 0 0 -0.01 0.002 -0.002 0.005 0

01 2 0 14

Rn

-0.002 0 0.005 -0.005 0 0 -0.005

13 of 17

13 of 17

1

0

2 1

3

4

5

02 4 0 20

50

100

150

Frequency/kHz 50 100 Frequency/kHz

Time/ms 0 2 3 4 5 FigureTime/ms 16. EMD results and spectral diagram of the signal from S1. 0

150

Figure Figure16. 16.EMD EMDresults resultsand andspectral spectraldiagram diagramof ofthe thesignal signalfrom fromS1. S1.

The energy ratio of the IMF at various levels is shown in Figure 17. As presented in the figure, The energy of the IMF various levels isisshown 17.17. AsAs presented in the figure, theof energy the IMFat at various levels showninwhich inFigure Figure presented in the figure, the firstThe three IMFsratio contain almost 98% of the entire signal, can reflect the main information first three IMFs contain almost 98% of the entire signal, which can reflect the main information of the the first three IMFs contain almost 98% of the entire signal, which can reflect the main information the original signal completely. Therefore, the three principal components, that is, IMF1, IMF2 of and original completely. Therefore, the three principal components, that is, IMF1, IMF2 IMF2 and IMF3, the original completely. Therefore, the three principal components, that is, lower IMF1, and IMF3, can signal be signal used for signal reconstruction, whereas other components with energy can can be used for signal reconstruction, whereas other components with lower energy can be ignored. IMF3, can be used for signal reconstruction, whereas other components with lower energy can be ignored. be ignored. 60

60 60

60

50

50

50

Energy Energyratio/% ratio/%

Energy ratio/%

Energy ratio/%

50 40

40

30 20 10 0

30 20 10 0

40 40

30 30

20

20

10

10

1

1

2

2

3

3

4

4

5

5

6

6

IMF1-9 IMF1-9

7

7

8

8

9 9

(a)(a) Signal Signalcollected collectedby bysensor sensorS1 S1

0

0

1

1

2

2

3

3

4

4

5

5

6

6

7

7

8

8

9

9

IMF1-9 IMF1-9 (b) Signal Signalcollected collectedby bysensor sensorS2S2 (b)

Figure IMF component. Figure 17. Energy ratio Figure17. 17.Energy Energyratio ratioof ofeach eachIMF IMFcomponent. component.

The correlationfunction functionofof ofthe thereconstructed reconstructed waveform The cross correlation The correlation waveform isis calculated. calculated.The Thecross crosscorrelation correlation The correlation function the reconstructed function diagram under five conditions is shown in Figure 18. The time difference is obtained from the function diagramunder five fiveconditions conditionsisis shown shown in Figure from function diagram Figure 18. 18. The Thetime timedifference differenceisisobtained obtained from peak value of of the figure andand the location on theon pipe calculated throughthrough the leakage location peak value ofthe thefigure figure andleakage theleakage leakage location the isiscalculated the leakage thethe peak value the location theispipe pipe calculated through the leakage formula. The positioning results are listed in Table 6. location formula. The positioning results are listed in Table 6. location formula. The positioning results are listed Table 6. Based on the results inthe thetable, table,the theoriginal original AE AE signal signal is is decomposed decomposed by the EMD, the main Based the results the table, the original AE main Based onon the results inin signal is decomposedby bythe theEMD, EMD,the the main IMF reconstruction relatively high energy and large correlation with the original signal is IMF reconstruction signal with high energy and large correlation with the original signal IMF reconstruction signal with relatively high energy and large correlation with the original signal selected and the main IMFs are selected to reconstruct the signal. The leakage source is determined is selected and the mainIMFs IMFsare areselected selectedto to reconstruct reconstruct the determined is selected and the main the signal. signal. The leakagesource sourceisis determined using a time-delay estimation method based on cross correlation. The accuracy of the using a time-delay estimation method based on cross correlation. The accuracy of thelocation location

Sensors 2018, 18, 3628

14 of 17

Sensors 2018, 18, x FOR PEER REVIEW

14 of 17

using a time-delay estimation method based on cross correlation. The accuracy of the location calculationisis high high and basically controlled within 5%, 5%, thusthus fullyfully meeting the actual calculation and the therelative relativeerror erroris is basically controlled within meeting the requirements of theof project. actual requirements the project. 0.4

0.5

0.3

0.4 0.3 0.2

互相关函数

互相关函数

0.2 0.1 0 -0.1

0.1 0 -0.1 -0.2

-0.2

-0.3

-0.3 -0.4 -0.5

X: 0.082 Y: 0.4919

-0.4

X: 0.051 Y: -0.3865

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

时间/ms (a) Condition 1 (400–600)

0.4

-0.5 -0.5

0.5

0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0.1

0.2

0.3

0.4

0.5

时间/ms (b) Condition 2 (500–800)

0.6

0.4

0.4

0.2

互相关函数

互相关函数

0.3

0.1 0 -0.1 -0.2 -0.3

X: 0.049 Y: -0.4304

-0.4 -0.5 -0.5

-0.4

-0.3

-0.2

-0.1

0.2 0 -0.2 -0.4 -0.6

0

0.1

0.2

0.3

时间/ms (c) Condition 3 (650–800)

0.4

0.5

X: 0.039 Y: -0.5292

-0.8 -0.5

-0.4

-0.3

-0.2

-0.1

0

时间/ms (d) Condition 4 (700–800)

0.5 0.4

X: 0.016 Y: 0.4045

互相关函数

0.3 0.2 0.1 0 -0.1 -0.2 -0.3 -0.4 -0.5 -0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

时间/ms (e) Condition 5 (800–950)

0.3

0.4

0.5

Figure Figure18. 18.Correlation Correlationfunctions functionsofofdifferent differentconditions conditionsand andEMD EMDdenoising. denoising. Table6.6.Locating Locatingresults resultsofofthe theleakage leakagesource sourceusing usingAE AEtechnology. technology. Table Experiment Experiment d1/mm Number d1/mm Number 1 1 400 400 2 2 500 500 3 650 3 650 700 4 4 700 800 5 5 800

Calculated Calculated Absolute Absolute Relative Relative Value: d1 d1 Error ErrorError Value: Error

d2/mm d2/mm

∆t/ms Δt/ms

600 600 800 800 800 800 800 800 950 950

0.051 0.051 0.082 0.082 0.049 0.049 0.039 0.039 0.016

412.13 412.13 508.71 508.71 640.57 640.57 682.80 682.80 847.43

12.1312.13 8.71 8.71 9.43 17.2 9.43 47.4317.2

3.03%3.03% 1.74% 1.74% 1.45% 2.46%1.45% 5.97%2.46%

0.016

847.43

47.43

5.97%

Sensors 2018, 18, 3628

15 of 17

Sensors 2018, 18, x FOR PEER REVIEW

15 of 17

6.3. Wiener Prefilter and GCC Method 6.3. Wiener Prefilter and GCC Method

The Wiener pre-filters and GCC have been widely used for denoising and leakage location [14]. The Wiener the pre-filters and GCC have between been widely used for denoising and [14]. Figure 19 provides correlation functions different conditions and theleakage Wienerlocation pre-filter and Figure 19 provides the correlation functions between different conditions and the Wiener pre-filter GCC method and Table 7 shows the location results of all test conditions. It is clear that the proposed and GCC method and Table 7 showsprocess the location results of allsimilar test conditions. is clear that wavelet denoising and EMD denoising have an accuracy to WienerItfiltering. The the basic proposed wavelet denoising and EMD denoising process have an accuracy similar to Wiener cross correlation (BCC) works as well as the GCC. filtering. The basic cross correlation (BCC) works as well as the GCC.

(a) Condition 1 (400–600)

(b) Condition 1 (600–800)

(c) Condition 1 (650–800)

(d) Condition 1 (700–800)

(e) Condition 1 (800–950) Figure 19. 19. Correlation Correlation functions functions between between different different conditions conditions and and the the Wiener Wiener pre-filter pre-filterand andGCC GCCmethod. method. Figure

Table 7. 7. Results GCC and andWiener Wienerpre-filter. pre-filter. Table Resultsofoflocating locatingthe theleakage leakagesource source using the GCC

Experiment Experiment Number Number 1 1 2 2 3 3 4 4 5 5

d1 d1(mm) (mm)

d2(mm) (mm) d2

Δt/ms ∆t/ms

400 400 500 500 650 650 700 700 800 800

600 600 800 800 800 800 800 800 950

0.058 0.058 0.076 0.076 0.053 0.053 0.017 0.017 0.027

950

0.027

Calculated Absolute Absolute Relative Relative Calculated Value: d1 d1 ErrorError ErrorError Value: 400.06 400.06 519.05 519.05 633.68 633.68 717.26 717.26 828.48 828.48

0.06 0.06 19.0519.05 16.3216.32 17.2617.26 35.37

35.37

0.015% 0.015% 3.81% 3.81% 2.51% 2.51% 2.47% 2.47% 3.36% 3.36%

Sensors 2018, 18, 3628

16 of 17

7. Conclusions In this study, a leakage AE signal is collected via an experiment with sensors inside a liquid-filled pipe. First, detection and recognition of the pipe running status are performed using SVM and AE technology. Then, wavelet decomposition and EMD are used to effectively denoise the original AE signal. Finally, the pipeline leakage source is accurately identified using the cross correlation delay estimation method. This study provides a reference value for a real-time online monitoring of pipeline operational status and has broad application prospects. The main research conclusions are summarized as follows: (1) In a liquid-filled pipe, the characteristics of the AE signal under different operating conditions vary significantly. The signal is continuous with standing wideband and low frequency, which is a characteristic of a typical noise signal. When liquid flows, a small fluctuation exists in the signal and the frequency spectrum indicates an area of energy concentration, whereas the amplitude is small. When leakage occurs, the signal suddenly presents the typical characteristics of an AE signal with narrowband and noticeable peak frequencies: the energy is concentrated and the amplitude is large. (2) The SVM method can effectively classify the conditions of pipeline operation with an accuracy rate of more than 95%. The test results show that for inseparable linear problems, the RBF kernel function is mapped to a high-dimensional space for classification and the recognition accuracy evidently improves. Meanwhile, the larger the training dataset, the more accurate the SVM model and the better the classification effect will be. (3) The propagation speed of the AE wave on the surface of a liquid-filled pipe is 3446 m/s, as determined by the lead-cutting experiment. Both types of noise reduction methods, namely wavelet decomposition and EMD, have ideal denoising ability for original leakage AE signals. The cross correlation time-delay estimation method can be used to accurately locate the pipeline leakage source. The positioning error is basically within 5%, which has certain practical significance in engineering applications. This article mainly used built-in sensors to detect pipeline leakage. The next phase of our research should include: (1) research with mobile robots and the optimization of robot fixed position, (2) the impact of leakage at different locations on the signal, which requires further study. Author Contributions: S.P. and Z.X. designed the experiments, analyzed the data. D.L. (Dang Lu) finished the experiments and processed some experiment results. D.L. (Dongsheng Li) proposed the methodology and writing—review & editing. Funding: This research received the financial support from the National Key Research and Development Program of China (Project No. 2017YFC0703410) and National Natural Science Foundation of China (NSFC) under Grant Nos. 51478079, 51778104, 51678110 and 51678108. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2.

3. 4. 5. 6.

Zhang, W.; Shi, Y.B.; Li, Y.J.; Luo, Q.W. A study of quantifying thickness of ferromagnetic pipes based on remote field eddy current testing. Sensors 2018, 18, 2769. [CrossRef] [PubMed] Lais, H.; Lowe, P.S.; Gan, T.H.; Wrobel, L.C.; Kanfoud, J. Characterization of the use of low frequency ultrasonic guided waves to detect fouling deposition in pipelines. Sensors 2018, 18, 2122. [CrossRef] [PubMed] Choi, J.; Shin, J.; Song, C.; Han, S.; Park, D. Leak detection and location of water pipes using vibration sensors and modified ML prefilter. Sensors 2017, 17, 2104. [CrossRef] [PubMed] Huang, J.; Zhou, Z.; Zhang, L.; Chen, J.; Ji, C.; Pham, D.T. Strain modal analysis of small and light pipes using distributed fibre bragg grating sensors. Sensors 2016, 16, 1583. [CrossRef] [PubMed] Wang, Y.Q.; Wan, S.; Zhou, Y.C.; Yang, H.G. Research on Technology of gas pipeline leakage detection based on infrasonic wave. Appl. Mech. Mater. 2013, 401, 1106–1109. [CrossRef] Jiao, J.; He, C.; Wu, B.; Fei, R. A new technique for modal acoustic emission pipeline leak location with one sensor. J. Mech. Eng. 2004, 40, 77–81. [CrossRef]

Sensors 2018, 18, 3628

7. 8. 9. 10. 11. 12. 13. 14. 15.

16. 17. 18. 19. 20. 21. 22. 23. 24. 25.

17 of 17

Li, Z.L.; Zhang, H.F.; Tan, D.J.; Chen, X.; Lei, H.X. A novel acoustic emission detection module for leakage recognition in a gas pipeline valve. Process SAF Environ. 2017, 105, 32–40. [CrossRef] Guo, C.; Wen, Y.; Li, P.; Wen, J. Pipeline leakage acoustic signal enhancement using EMD. Chin. J. Sci. Instrum. 2015, 36, 1397–1405. Li, H.; Shen, R. Pipeline leak detection method for thermal network based on improved empirical mode decomposition method. Sci. Technol. Eng. 2014, 14, 224–227. Sun, L.; Li, Y.; Qu, Z.; Wang, W.K. Study on acoustic emission pipeline leak detection based on EMD signal analysis method. J. Vib. Shock 2007, 26, 161–164. Li, D.S.; Ou, J.P.; Lan, C.M.; Li, H. Monitoring and Failure Analysis of Corroded Bridge Cables under Fatigue Loading Using Acoustic Emission Sensors. Sensors 2012, 12, 3901–3915. [CrossRef] [PubMed] Wu, Z.; Zhou, S.; Jiang, G. Signal processing of acoustic emission detection of pipeline leakage based on wavelet transform. Manuf. Autom. 2012, 34, 21–23. Liu, G.; Xu, M.; Wang, X.; Jiang, D.L. Research on acoustic emission detection of inner leakage of pipeline valve based on HHT. J. Vib. Shock 2012, 31, 62–66. Gao, Y.; Brennan, M.; Joseph, P.A. Comparison of time delay estimators for the detection of leak noise signals in plastic water distribution pipes. J. Sound Vib. 2006, 292, 552–570. [CrossRef] Shama, A.; El-Shaib, M.; Sharara, A.; Nasser, D.Y. Experimental study for leakage detection in subsea pipeline by applying acoustic emission technique. In Proceedings of the International Congress of the International Maritime Association of the Mediterranean, Lisbon, Portugal, 9–11 October 2017. Xu, C.; Gong, P.; Xie, J.; Shi, H.; Chen, G.; Song, G. An acoustic emission based multi-level approach to buried gas pipeline leakage localization. J. Loss Prev. Process Ind. 2016, 44, 397–404. [CrossRef] Mostafapour, A.; Davoodi, S. Leakage locating in underground high pressure gas pipe by acoustic emission method. J. Nondestruct. Eval. 2013, 32, 113–123. [CrossRef] Mostafapour, A.; Davoudi, S. Analysis of leakage in high pressure pipe using acoustic emission method. Appl. Acoust. 2013, 74, 335–342. [CrossRef] Yu, L.; Li, S. Acoustic emission (AE) based small leak detection of galvanized steel pipe due to loosening of screw thread connection. Appl. Acoust. 2017, 120, 85–89. [CrossRef] Liu, X.; Jia, D.; Li, H.; Jiang, J. Research on Kernel parameter optimization of support vector machine in speaker recognition. Sci. Technol. Eng. 2010, 10, 1669–1673. Khulief, Y.; Khalifa, A.; Mansour, R.B.; Habib, M.A. Acoustic detection of leaks in water pipelines using measurements inside pipe. J. Pipeline Syst. Eng. Pract. 2011, 3, 47–54. [CrossRef] Li, S.; Wen, Y.; Li, P.; Yang, J.; Yang, L. Determination of acoustic speed for improving leak detection and location in gas pipelines. Rev. Sci. Instrum. 2014, 85, 024901. [CrossRef] [PubMed] Prosser, W.H.; Jackson, K.E.; Kellas, S.; Smith, B.T.; McKeon, J.; Friedman, A. Advanced, waveform based acoustic emission detection of matrix cracking in composites. NDT E Int. 1995, 30, 1052–1058. Zhao, Y.; Zhuang, X.; Min, S. A new method of leak location for the natural gas pipeline based on wavelet analysis. Energy 2010, 35, 3814–3820. Liu, B.; Riemenschneider, S.; Xu, Y. Gearbox fault diagnosis using empirical mode decomposition and Hilbert spectrum. Mech. Syst. Signal Process. 2006, 20, 718–734. [CrossRef] © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).