UwDAS gas pipeline leakage identification method based on SE-2DCNN model

Abstract

The uwDAS gas pipeline leakage identification method based on the SE-2DCNN model uses the uwDAS system to collect pipeline leakage signals S(n); constructs a pipeline leakage data set and extracts the MFCC features of the signals; constructs a SE-2DCNN model; optimizes the network parameters of the SE-2DCNN model; and inputs the SE-2DCNN model with the optimized parameters to identify the pipeline leakage data set. The gas pipeline leakage identification method introduces a spatial expansion two-dimensional convolutional neural network (SE-2DCNN), optimizes the feature extraction and classification performance of the processed leakage signals, and the SE-2DCNN structure can more deeply capture and learn the spatial features of the leakage signals, so that the efficient learning mechanism improves the identification accuracy of the pipeline leakage events.

Description

Technical Field

[0001] This invention relates to the field of gas pipeline leak monitoring technology, specifically to a uwDAS gas pipeline leak identification method based on the SE-2DCNN model. Background Technology

[0002] Pipelines are the most common, safest, and lowest-cost mode of gas transportation. However, pipelines are prone to leaks due to factors such as natural aging, corrosion, climate change, and human interference. If leaks are not accurately identified and effectively handled, they can lead to significant property damage, fires, and explosions, causing casualties. Therefore, improving the accuracy of pipeline leak detection is crucial.

[0003] Distributed Acoustic Sensing (DAS) is a technology that achieves distributed detection of vibrations or acoustic waves on optical fibers by demodulating the intensity or phase information of coherent backscattered Rayleigh light in the fiber. Due to its advantages such as high spatial resolution and long sensing distance, it has been widely used in pipeline leak monitoring. However, it is susceptible to false alarms due to the influence of signal-to-noise ratio and external environment. Ultra-weak Fiber Bragg Rating Distributed Acoustic Sensing (uwDAS) technology retains the original advantages of DAS technology while utilizing broadband gratings with a reflectivity of less than 0.1% to enhance the signal, improving the directionality of coherent light and increasing the signal-to-noise ratio by more than 20 dB, which reduces false alarms and missed alarms to some extent. Therefore, uwDAS technology can achieve more efficient and accurate pipeline leak monitoring. However, in practical applications, it has been found that using deep learning methods can further improve the accuracy of pipeline leak monitoring. Therefore, deep learning methods are introduced into pipeline leak monitoring based on uwDAS technology.

[0004] Current deep learning methods applied to pipeline leak detection include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long-term Recurrent Convolutional Networks (LRCNs). In this process, feature selection is crucial for building a more efficient classifier and reducing computational complexity. However, these deep learning methods are limited by their insufficient handling of interactions between network channels. This prevents them from fully utilizing multi-dimensional features and necessitates independent feature extraction for each candidate region, leading to redundant computation. This not only reduces the algorithm's efficiency but also negatively impacts accuracy.

[0005] The paper [Hu, Jie et al. “Squeeze-and-Excitation Networks.” 2018 IEEE / CVFConference on Computer Vision and Pattern Recognition (2017): 7132-7141.] proposed the SE (Squeeze-and-Excitation) module. It can adaptively capture the dependencies between different channels, thereby enhancing the network's sensitivity to key features and suppressing unimportant features. Furthermore, by adaptively adjusting the feature weights on each channel of the feature map, it avoids repetitive feature extraction for each candidate region, effectively improving the algorithm's efficiency and recognition accuracy.

[0006] Defects and shortcomings of existing technology:

[0007] (1) The main problems with traditional DAS systems in practical applications are low sensitivity and signal-to-noise ratio. This is because the Rayleigh scattering intensity in standard single-mode fiber is weak, resulting in a low signal-to-noise ratio of the demodulated signal, which limits the system's ability to monitor minute external disturbances, thus affecting the accuracy and reliability of monitoring. Therefore, a uwDAS system with higher sensitivity and signal-to-noise ratio is needed.

[0008] (2). Since acoustic signals are one-dimensional signals, existing methods mainly focus on extracting their time-domain features, while obtaining frequency-domain features is relatively difficult. Therefore, converting one-dimensional signals into two-dimensional signals allows for the simultaneous acquisition of both time-domain and frequency-domain information, thus enabling more comprehensive extraction of signal features.

[0009] (3) When using deep learning to extract features from pipeline intrusion signals, some unimportant features may reduce the recognition accuracy. Furthermore, most deep learning-based methods have long recognition times and high time complexity. In practical applications, time costs need to be considered. Therefore, an SE module is added to the original CNN model to extract more important and significant features. Summary of the Invention

[0010] To address the shortcomings of existing deep learning methods and the insufficient sensitivity and signal-to-noise ratio of current DAS technology for vehicle detection, this invention introduces uwDAS technology and an improved convolutional neural network into pipeline leak monitoring. It provides a uwDAS gas pipeline leak identification method based on the SE-2DCNN model, which can more accurately identify pipeline leaks.

[0011] The technical solution adopted in this invention is as follows:

[0012] The uwDAS gas pipeline leak identification method based on the SE-2DCNN model includes the following steps:

[0013] Step 1: Use the uwDAS system to collect the pipeline leak signal S(n);

[0014] Step 2: Construct a pipeline leak dataset and extract MFCC features from the signals;

[0015] Step 3: Construct the SE-2DCNN model;

[0016] Step 4: Optimize the network parameters of the SE-2DCNN model;

[0017] Step 5: Input the optimized SE-2DCNN model to identify pipeline leak datasets.

[0018] Step 2 includes the following steps:

[0019] Step 2.1: Construct a pipeline leak dataset, dividing the dataset into a training set, a validation set, and a test set in a 6:2:2 ratio;

[0020] Step 2.2: After pre-emphasing the pipeline leakage signal S(n) in the segmented dataset, we obtain S'(n):

[0021] S'[n]=S[n]-α×S[n-1]

[0022] In the above formula, S'[n] is the sampled value of the pre-emphasized signal at discrete time point n; S[n] is the sampled value of the original audio signal at discrete time point n; S[n-1] is the sampled value of the original audio signal at discrete time point n-1; α is the pre-emphasis coefficient, which is usually taken as approximately 0.95 to 0.97.

[0023] Step 2.3: Divide the pre-emphasized frames into overlapping frames, with a frame length of N and an overlap length of M.

[0024] x frame (m,n)=s′(n+m)

[0025] In the above formula, m represents the relative position within each frame; n is the index of the discrete time point; x frame (m,n) represents the m-th sample value of the n-th frame audio signal; s′(n+m) represents the value of the pre-emphasized signal at time point n+m.

[0026] Step 2.4: Apply windowing functions to each frame of signal:

[0027] x windowed (m,n)=w(m)×x frame (m,n)

[0028] In the above formula, x windowed (m,n) represents the m-th sample value of the n-th frame audio signal after applying the windowing function; w(m) is the window function.

[0029] Step 2.5: Perform an FFT transform on each frame of the signal to obtain the spectrum:

[0030] X(m,k)=FFT(x windowed (m,n))

[0031] In the above formula, X(m,k) is the spectral representation of the m-th frame audio signal; k is the frequency index.

[0032] Step 2.6: Map the spectrum X(m,k) to the Mel frequency to obtain the Mel spectrum S. m (f):

[0033]

[0034] In the above formula, S m (f) represents the frequency response of the m-th Mel filter bank; N is the length of each frame, i.e., the number of samples contained; H m (k) is a weighting function used to emphasize or weaken the influence of certain frequency components.

[0035] Step 2.7: Take the logarithm of the Mel spectrogram to obtain the logarithmic Mel spectrogram:

[0036] S mel-log (m,f)=log(S m (f)

[0037] Step 2.8: Perform DCT transform on the log-Melbourne spectrum to obtain the MFCC coefficients:

[0038]

[0039] In the above formula, C n is the nth MFCC coefficient; M is the size of the Mel spectrum.

[0040] In step 3, based on the characteristics of the leakage signal, an SE-2DCNN model network is constructed, mainly consisting of 3 convolutional blocks, 1 pooling layer, and 2 fully connected layers. The convolutional blocks primarily include convolutional layers, batch normalization layers, LeakyReLU activation functions, SE modules, and shortcut connections. Specifically, the following steps are included:

[0041] Step 3.1: Construct the convolutional layers in the convolutional block, using the following formula:

[0042]

[0043] In the above formula, g(i) represents the feature map generated by the i-th convolutional kernel; m represents the width of the input tensor; n represents the height of the input tensor; and p represents the depth of the input tensor.

[0044] a x,y,z It is the input data; w i,x,y,z and b i These are the weights and biases of the convolution kernel, respectively.

[0045] Step 3.2: Construct the batch normalization layer in the convolutional block, using the following formula:

[0046]

[0047] In the above formula, BN(g(i)) is the output after batch normalization; γ is the scaling parameter; μ is the mean deviation; σ is the standard deviation; ∈ is a small constant to prevent the denominator from being zero; and β is the translation parameter.

[0048] Step 3.3: Construct the LeakyReLU activation function in the convolutional block, as shown in the following formula:

[0049]

[0050] In the above formula, X relu1 This is the output after the activation function.

[0051] Step 3.4: Construct the SE module in the convolutional block. First, construct the average pooling of the SE module, and the formula is as follows:

[0052]

[0053] In the above formula, X sel The output is the result of global average pooling; H is the height of the feature map; W is the width of the feature map.

[0054] Step 3.5: Construct the fully connected layer 1 of the SE module, with the following formula:

[0055] X fc1 =max(0,W se1 X se1 +b se1 )

[0056] In the above formula, X fc1 This is the output of fully connected layer 1; W se1 b represents the weights of fully connected layer 1; se1 This is the bias for fully connected layer 1.

[0057] Step 3.6: Construct the fully connected layer 2 of the SE module, with the following formula:

[0058] X fc2 =σ(Wse2 X fc1 +b se2 )

[0059] In the above formula, X fc2 This is the output of fully connected layer 2; W se2 b represents the weights of the fully connected layer 2. se2 σ is the bias of the fully connected layer 2; σ(·) is the Sigmoid function.

[0060] Step 3.7: Finally, adjust the weights in the SE module using the following formula:

[0061] X se2 =X relu1 ·X fc2

[0062] In the above formula, X relu1 It is the output after the activation function in step 3.3; X se2 It is for X relu1 The result after channel attention weighting (by element-wise multiplication).

[0063] Step 3.8: Construct the residual connections for the convolutional blocks, using the following formula:

[0064]

[0065] In the above formula, X shortcut The output of the residual connection; x is the index of the input tensor c in the width direction; y is the index of the input tensor c in the height direction; z is the index of the input tensor c in the depth direction; c x,y,z The value of the input data at (x, y, z); w shortcut,i,x,y,z b represents the weights of the residual connections. shortcut,i This is the bias for the residual connection.

[0066] Step 3.9: After constructing the first convolutional block, construct the second and third convolutional blocks according to steps 3.1 to 3.8, and construct a global average pooling layer, with the following formula:

[0067]

[0068] In the above formula, X pool The output of global average pooling; H is the height of the feature map; W is the width of the feature map; X is the height of the feature map. block3 This is the output of the third convolutional block; X block3 (i,j) represents the feature map value at position (i,j).

[0069] Step 3.10: Construct fully connected layer 1:

[0070] Xfc1 =max(0,W fc1 X flatten +b fc1 )

[0071] In the above formula, X fc1 This is the output of fully connected layer 1; W fc1 X represents the weights of fully connected layer 1; flatten The output after the flattening operation; b fc1 This is the bias for fully connected layer 1.

[0072] Step 3.11: Construct fully connected layer 2:

[0073] X fc2 =W fc2 X dropout +b fc2

[0074] In the above formula, X fc2 This is the output of fully connected layer 2; W fc2 The weights of fully connected layer 2; X dropout This is the output after Dropout; b fc2 This is the bias for the fully connected layer 2.

[0075] In step 4, the parameters of the constructed SE-2DCNN model are tuned, as follows:

[0076] Three different batch sizes (32, 64, 128), four different optimizers (Adam, Sgd, Rmsprop, Adamamax), and eight different learning rates (1E-1, 1E-2, 1E-3, 1E-4, 1E-5, 1E-6, 1E-7, 1E-8) were considered, and various combinations of these parameters were tested. For each parameter combination, 10 tests were performed to ensure the stability of the results, and the average accuracy and training time were calculated. Finally, the optimal parameter combination was obtained by comparing accuracy and training time, with a batch size of 64, using the Adam optimizer, and a learning rate of 1E-5.

[0077] This invention discloses a gas pipeline leak identification method based on the SE-2DCNN model using uwDAS, with the following technical advantages: 1) The uwDAS device used in this invention has high sensitivity and signal-to-noise ratio, improving the signal-to-noise ratio by more than 20dB compared to traditional DAS. By enhancing signal quality and clarity, the leak signal becomes more prominent in complex noisy environments, thus providing higher-quality data input for subsequent signal analysis and effectively improving the accuracy and reliability of leak identification. 2) This invention optimizes the feature extraction and classification performance of processed leak signals by introducing a spatially dilated two-dimensional convolutional neural network (SE-2DCNN). The SE-2DCNN structure can more deeply capture and learn the spatial features of leak signals. Through its efficient learning mechanism, it improves the accuracy of pipeline leak event identification, especially under various leak types.

[0078] 3) This invention optimizes the system parameters of the improved algorithm to identify leaks, which has a higher accuracy rate than traditional algorithms and can more accurately identify pipeline leaks and different pipeline leak situations. Attached Figure Description

[0079] The present invention will be further described below with reference to the accompanying drawings and examples;

[0080] Figure 1 This is a flowchart of the present invention.

[0081] Figure 2 This is a diagram of the experimental apparatus of the present invention.

[0082] Figure 3(a) shows the time-domain and frequency-domain plots (No Leak) of different levels of leakage collected by the uwDAS system;

[0083] Figure 3(b) shows the time-domain and frequency-domain plots (PL24) of different levels of leakage collected by the uwDAS system;

[0084] Figure 3(c) shows the time-domain and frequency-domain plots (PL26) of different levels of leakage collected by the uwDAS system;

[0085] Figure 3(d) shows the time-domain and frequency-domain plots (PL28) of different levels of leakage collected by the uwDAS system;

[0086] Figure 3(e) shows the time-domain and frequency-domain plots (PL44) of different levels of leakage collected by the uwDAS system;

[0087] Figure 3(f) shows the time-domain and frequency-domain plots (PL46) of different levels of leakage collected by the uwDAS system;

[0088] Figure 3(g) shows the time-domain and frequency-domain plots (PL48) of different levels of leakage collected by the uwDAS system;

[0089] Figure 3(h) shows the time-domain and frequency-domain plots (PL64) of different levels of leakage collected by the uwDAS system;

[0090] Figure 3(i) shows the time-domain and frequency-domain plots (PL66) of different levels of leakage collected by the uwDAS system;

[0091] Figure 3(j) shows the time-domain and frequency-domain plots (PL68) of different levels of leakage collected by the uwDAS system.

[0092] Figure 4 The structure diagram of the improved SE-2DCNN model.

[0093] Figure 5 To improve the model's average accuracy and training time under different parameters. Detailed Implementation

[0094] like Figure 1 As shown, the uwDAS gas pipeline leak identification method based on the SE-2DCNN model includes the following steps:

[0095] Step 1: Use the uwDAS system to collect the pipeline leak signal S(n);

[0096] Step 2: Construct a pipeline leak dataset and extract MFCC features from the signals;

[0097] Step 3: Construct the SE-2DCNN model;

[0098] Step 4: Optimize the network parameters of the SE-2DCNN model;

[0099] Step 5: Input the optimized SE-2DCNN model to identify pipeline leak datasets.

[0100] In step 1, the uwDAS system is used to collect signals S(n) for different pipeline pressures and leakage rates. The monitoring optical cable is a distributed optical cable with an ultra-weak fiber optic grating array with a grating spacing of 5m, and two grating points are suspended at the left and right ends of the pipeline, respectively. When the pipeline pressure and leakage rate are adjusted, the monitoring optical cable can sense a large amount of leakage data on the pipeline. After the optical cable is connected to the uwDAS system, the system can acquire this pipeline leakage data in real time.

[0101] In step 2, a pipeline leak dataset is constructed, and MFCC features of the signal are extracted. The specific steps are as follows:

[0102] 2.1: Construct a pipeline leakage dataset and divide the dataset into a training set, a validation set, and a test set in a 6:2:2 ratio.

[0103] 2.2: After pre-emphasizing the leakage signal S(n) in the divided dataset, we obtain S'(n):

[0104] S'[n]=S[n]-α×S[n-1]

[0105] In the above formula, S'[n] is the pre-emphasized signal, the sampled value at discrete time point n; S[n] is the original audio signal, the sampled value at discrete time point n; S[n-1] is the original audio signal, the sampled value at discrete time point n-1; α is the pre-emphasis coefficient, which is usually taken as approximately 0.95 to 0.97.

[0106] 2.3: Divide the pre-emphasized frames into overlapping frames, with a frame length of N and an overlap length of M:

[0107] x frame (m,n)=s′(n+m)

[0108] In the above formula, m represents the relative position within each frame; n is the index of the discrete time point; x frame (m,n) represents the m-th sample value of the n-th frame audio signal; s′(n+m) represents the value of the pre-emphasized signal at time point n+m.

[0109] 2.4: Windowing is performed on each frame of signal using window functions:

[0110] x windowed (m,n)=w(m)×x frame (m,n)

[0111] In the above formula, x windowed (m,n) represents the m-th sample value of the n-th frame audio signal after applying the windowing function; w(m) is the window function.

[0112] 2.5: Perform an FFT transform on each frame of the signal to obtain the spectrum:

[0113] X(m,k)=FFT(x windowed (m,n))

[0114] In the above formula, X(m,k) is the spectral representation of the m-th frame audio signal; k is the frequency index.

[0115] 2.6: Map the spectrum X(m,k) to the Mel frequency to obtain the Mel spectrum S. m (f):

[0116]

[0117] In the above formula, S m (f) represents the frequency response of the m-th Mel filter bank; N is the length of each frame (i.e., the number of samples contained); Hm (k) is a weighting function used to emphasize or reduce the influence of certain frequency components.

[0118] 2.7: Taking the logarithm of the Mel spectrogram yields the logarithmic Mel spectrogram:

[0119] S mel-log (m,f)=log(S m (f)

[0120] 2.8: Perform DCT transform on the log-Melbourne spectrum to obtain the MFCC coefficients:

[0121]

[0122] In the above formula, C n is the nth MFCC coefficient; M is the size of the Mel spectrum.

[0123] In step 3, based on the characteristics of the leakage signal, an SE-2DCNN network is constructed, mainly consisting of 3 convolutional blocks, 1 pooling layer, and 2 fully connected layers. The convolutional blocks primarily include convolutional layers, batch normalization layers, LeakyReLU activation functions, SE modules, and shortcut connections. The specific steps are as follows:

[0124] 3.1: Construct the convolutional layer in the convolutional block, using the following formula:

[0125]

[0126] In the above formula, g(i) represents the feature map generated by the i-th convolutional kernel; a x,y,z It is the input data; w i,x,y,z and b i These are the weights and biases of the convolution kernel, respectively; m represents the width of the input tensor; n represents the height of the input tensor; and p represents the depth of the input tensor.

[0127] 3.2: Construct the batch normalization layer in the convolutional block, using the following formula:

[0128]

[0129] In the above formula, BN(g(i)) is the output after batch normalization; γ is the scaling parameter; μ is the mean deviation; σ is the standard deviation; ∈ is a small constant to prevent the denominator from being zero; and β is the translation parameter.

[0130] 3.3: Construct the LeakyReLU activation function in the convolutional block, as shown in the following formula:

[0131]

[0132] In the above formula, X relu1This is the output after the activation function.

[0133] 3.4: Constructing the SE module in the convolutional block. First, construct the average pooling of the SE module, the formula of which is as follows:

[0134]

[0135] In the above formula, X sel The output is the result of global average pooling; H is the height of the feature map; W is the width of the feature map.

[0136] 3.5: Construct the fully connected layer 1 of the SE module, using the following formula:

[0137] X fc1 =max(0,W se1 X se1 +b se1 )

[0138] In the above formula, X fc1 This is the output of fully connected layer 1; W se1 b represents the weights of fully connected layer 1; se1 This is the bias for fully connected layer 1.

[0139] 3.6: Construct the fully connected layer 2 of the SE module, with the following formula:

[0140] X fc2 =σ(W se2 X fc1 +b se2 )

[0141] In the above formula, X fc2 This is the output of fully connected layer 2; W se2 b represents the weights of the fully connected layer 2. se2 σ is the bias of the fully connected layer 2, and σ(·) is the Sigmoid function.

[0142] 3.7: Finally, the weights are adjusted in the SE module, using the following formula:

[0143] X se2 =X relu1 ·X fc2

[0144] In the above formula, X relu1 It is the output after the activation function in step 3.3; X se2 It is for X relu1 The result after channel attention weighting (by element-wise multiplication).

[0145] 3.8: Constructing residual connections for convolutional blocks, the formula is as follows:

[0146]

[0147] In the above formula, X shortcut The output of the residual connection; x is the index of the input tensor c in the width direction; y is the index of the input tensor c in the height direction; z is the index of the input tensor c in the depth direction; c x,y,z The value of the input data at (x, y, z); w shortcut,i,x,y,z B represents the weights of the residual connections. shortcut,i This is the bias for the residual connection.

[0148] 3.9: After constructing the first convolutional block, construct the second and third convolutional blocks according to 3.1-3.8, and construct a global average pooling layer, the formula of which is as follows:

[0149]

[0150] In the above formula, X pool The output of global average pooling; H is the height of the feature map; W is the width of the feature map; X is the height of the feature map. block3 This is the output of the third convolutional block; X block3 (i,j) represents the feature map value at position (i,j).

[0151] 3.10: Constructing the fully connected layer 1:

[0152] X fc1 =max(0,W fc1 X flatten +b fc1 )

[0153] In the above formula, X fc1 This is the output of fully connected layer 1; W fc1 X represents the weights of fully connected layer 1; flatten The output after the flattening operation; b fc1 This is the bias for fully connected layer 1.

[0154] 3.11: Constructing the fully connected layer 2:

[0155] X fc2 =W fc2 X dropout +b fc2

[0156] In the above formula, X fc2 The output of fully connected layer 2; W fc2 The weights of fully connected layer 2; X dropout This is the output after Dropout; b fc2 This is the bias for the fully connected layer 2.

[0157] In step 4, the constructed SE-2DCNN model undergoes parameter tuning. Three different batch sizes (32, 64, 128), four different optimizers (Adam, Sgd, Rmsprop, Adamamax), and eight different learning rates (1E-1, 1E-2, 1E-3, 1E-4, 1E-5, 1E-6, 1E-7, 1E-8) are considered, and various combinations of these parameters are tested. For each parameter combination, 10 tests are performed to ensure the stability of the results. The average accuracy and training time are calculated, and the optimal parameter combination is obtained, with a batch size of 64, using the Adam optimizer, and a learning rate of 1E-5.

[0158] In step 5, the optimal parameters obtained in step 4 are set into the network to identify the collected leaked data.

[0159] Based on this, a complete pipeline leak monitoring system was built, and the accuracy of gas pipeline leak identification was improved by using SE-2DCNN with optimized parameters.

[0160] Example:

[0161] The data used in the experiments of this invention are leakage signals collected by the uwDAS system on the pipeline. A diagram of the specific experimental setup is shown below. Figure 2 As shown, the following parameters should be set when monitoring and identifying these leakage signals.

[0162] uwDAS system data acquisition parameter settings:

[0163] The system has a sampling frequency of 10kHz, a sampling length of 200m, a grating spacing of 5m, a pulse width of 24ns, and a pulse period of 100μs.

[0164] Creating a dataset:

[0165] Data were collected for 10 different leakage conditions based on different leakage amounts and pipeline pressures, as shown in Table 1. Figures 3(a) to 3(j) The following are the time-domain and frequency-domain plots of these ten data types. Among them, Figure 3(a) is the time-domain and frequency-domain plot under the condition of no leakage, where the phase amplitude remains relatively stable and the frequency response exhibits low amplitude characteristics. Figures 3(b) to 3(j) The time-domain and frequency-domain plots show the results under different leakage rates and pipeline pressures. As pipeline pressure increases, the phase amplitude gradually increases, indicating that higher pressure has a greater impact on the acoustic signal. Furthermore, as the leakage rate increases, the phase amplitude also shows an increasing trend, and the frequency domain amplitude becomes larger. Close observation reveals that the frequency domain response of all leakage data is concentrated below 500Hz, indicating that the acoustic signal generated by leakage tends to exhibit significant characteristics in the lower frequency range.

[0166] Table 1 Data for different leakage conditions

[0167]

[0168] Mel frequency cepstral coefficient calculation parameter settings:

[0169] The filter bank is set to 64, and the cepstral coefficients are set to 20.

[0170] SE-2DCNN network parameter settings:

[0171] Figure 4 The SE-2DCNN network structure mainly consists of three convolutional blocks, fully average pooling layers, and fully connected layers. The SE module is contained within the convolutional blocks. Its parameter settings are shown in Table 2. A moderate 3×3 square convolutional kernel with a stride of 2 is used in the convolutional layers; in the SE module, the kernel size of the two convolutional modules is 1×1 with a stride of 1.

[0172] Table 2 Parameter Settings

[0173]

[0174]

[0175] SE-2DCNN network hyperparameter settings:

[0176] Figure 5 This study investigates the impact of different parameter combinations on average accuracy and training time when using the SE-2DCNN model for gas pipeline leak detection. The bar chart represents the average accuracy for each parameter combination, while the line chart reflects the training time. Ten tests were conducted for each parameter combination to ensure the stability of the results, and the average accuracy was calculated. A batch size of 64, the Adam optimizer, and a learning rate of 1E-5 were ultimately selected.

[0177] The following three methods were used to compare the identification effects of pipeline leakage signals collected by uwDAS:

[0178] Method 1: Reference [Han, Shunjie et al. "Parameter selection in SVM with RBFkernel function."]

[0179] The SVM algorithm in World Automation Congress 2012 (2012): 1-4.

[0180] Method 2: The basic CNN algorithm without the SE module, whose network parameters are shown in Table 3.

[0181] Table 3. Network parameters of the basic CNN algorithm

[0182]

[0183] Method 3: The SE-2DCNN recognition method proposed in this invention.

[0184] Table 4 Performance metrics of SE-2DCNN, baseline CNN model, and SVM model

[0185]

[0186] Table 4 shows the performance metrics of SE-2DCNN, baseline CNN, and SVM models in terms of precision, recall, accuracy, and F1 score, all averaged results from 10 runs. SE-2DCNN achieves a precision of 0.9317, superior to the baseline CNN's 0.8463 and SVM's 0.8811. In terms of recall, SE-2DCNN demonstrates stronger ability to identify positive samples. The SE-2DCNN model also outperforms the baseline CNN and SVM models in accuracy. Furthermore, the difference in accuracy and F1 score between the SE-2DCNN model and the baseline CNN model is 2.23%, while the difference is 5.27% for the baseline CNN model and 3.68% for the SVM model. This indicates that the SE-2DCNN model exhibits less bias in classification compared to the baseline CNN and SVM models. Therefore, the SE-2DCNN model demonstrates superior classification performance in leak detection and prediction compared to the other two methods.

[0187] This invention discloses a gas pipeline leak identification method based on ultra-weak fiber Bragg grating distributed acoustic wave sensing, and introduces an improved convolutional neural network to enhance the accuracy of gas pipeline leak identification. For 10 types of pipeline leak signals acquired by uwDAS, this invention first extracts the MFCC features of the signals, and then uses an optimized hyperparameter SE-2DCNN network for identification. Compared with traditional identification methods, this invention's pipeline leak identification method extracts two-dimensional features from the pipeline leak signals acquired by uwDAS, increasing information density and better capturing spatial correlation. Subsequently, through an attention mechanism, important features are preserved to the maximum extent while unimportant features are suppressed, resulting in a significant improvement in the accuracy of pipeline leak signal identification.

Claims

1. A method for identifying gas pipeline leaks in uwDAS based on the SE-2DCNN model, characterized in that... Includes the following steps: Step 1: Use the uwDAS system to collect pipeline leak signals S ( n ); Step 2: Construct a pipeline leak dataset and extract MFCC features from the signals; Step 3: Construct the SE-2DCNN model; Step 4: Optimize the network parameters of the SE-2DCNN model; Step 5: Input the optimized SE-2DCNN model to identify pipeline leaks in the dataset; In step 3, based on the characteristics of the leakage signal, an SE-2DCNN model network is constructed, including 3 convolutional blocks, 1 pooling layer, and 2 fully connected layers. Each convolutional block includes a convolutional layer, a batch normalization layer, a LeakyReLU activation function, an SE module, and shortcut connections. Specifically, the steps include: Step 3.1: Construct the convolutional layers in the convolutional block, using the following formula: In the above formula, Indicates the first Feature mapping maps generated by each convolutional kernel; Indicates the width of the input tensor; Indicates the height of the input tensor; Indicates the depth of the input tensor; It is the input data; and These are the weights and biases of the convolution kernel, respectively. Step 3.2: Construct the batch normalization layer in the convolutional block, using the following formula: In the above formula, This is the output after batch normalization; For scaling parameters; The mean difference; Standard deviation; To prevent a small constant with a denominator of zero; These are translation parameters; Step 3.3: Construct the LeakyReLU activation function in the convolutional block, as shown in the following formula: In the above formula, This is the output after the activation function; Step 3.4: Construct the SE module in the convolutional block. First, construct the average pooling of the SE module, and the formula is as follows: In the above formula, This is the output after global average pooling; The height of the feature map; The width of the feature map; Step 3.5: Construct the fully connected layer 1 of the SE module, with the following formula: In the above formula, This is the output of fully connected layer 1; These are the weights of the fully connected layer 1; This is the bias for fully connected layer 1; Step 3.6: Construct the fully connected layer 2 of the SE module, with the following formula: In the above formula, This is the output of fully connected layer 2; These are the weights of the fully connected layer 2; For the bias of fully connected layer 2; For the Sigmoid function; Step 3.7: Finally, adjust the weights in the SE module using the following formula: In the above formula, It is the output after the activation function in step 3.3; Yes The result after channel attention weighting; Step 3.8: Construct the residual connections for the convolutional blocks, using the following formula: In the above formula, The output of the residual connection; For input tensors Index in the width direction; For input tensors Index in the height direction; For input tensors Index in the depth direction; For input data in The value at; The weights for the residual connections; The bias is for the residual connection; Step 3.9: After constructing the first convolutional block, construct the second and third convolutional blocks according to steps 3.1 to 3.8, and construct a global average pooling layer, with the following formula: In the above formula, This is the output of global average pooling; The height of the feature map; The width of the feature map; This is the output of the third convolutional block; For in position Feature map values ​​at; Step 3.10: Construct fully connected layer 1: In the above formula, This is the output of fully connected layer 1; These are the weights of the fully connected layer 1; This is the output after the flattening operation; This is the bias for fully connected layer 1; Step 3.11: Construct fully connected layer 2: In the above formula, This is the output of fully connected layer 2; These are the weights of the fully connected layer 2; This is the output after Dropout; This is the bias for the fully connected layer 2.

2. The uwDAS gas pipeline leak identification method based on the SE-2DCNN model according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Construct a pipeline leak dataset, dividing the dataset into a training set, a validation set, and a test set in a 6:2:2 ratio; Step 2.2: Divide the pipeline leak signals in the dataset into segments. S ( n After pre-weighting treatment, the result is... S' ( n ): In the above formula, For the pre-emphasized signal at discrete time points The sampled values; The original audio signal at discrete time points The sampled value at that location; The original audio signal at discrete time points The sampled value at that location; This is the pre-weighting factor, with a value ranging from 0.95 to 0.97; Step 2.3: Divide the pre-emphasized data into overlapping frames, and set the frame length as . The overlap length is : In the above formula, To indicate the relative position within each frame; Indexes for discrete time points; For the first The first frame of the audio signal Each sample value; For the pre-emphasized signal at time point The value at; Step 2.4: Apply windowing functions to each frame of signal: In the above formula, The first one after applying the windowing function The first frame of the audio signal Each sample value; For window functions; Step 2.5: Perform an FFT transform on each frame of the signal to obtain the spectrum: In the above formula, For the first Spectral representation of frame audio signals; It is a frequency index; Step 2.6: Spectrum Mapping to Mel frequencies yields the Mel spectrum. : In the above formula, For the first Frequency response of a Mel filter bank; It is the length of each frame, i.e., the number of samples it contains; It is a weighting function used to emphasize or weaken the influence of certain frequency components; Step 2.7: Take the logarithm of the Mel spectrogram to obtain the logarithmic Mel spectrogram: Step 2.8: Perform DCT transform on the log-Melbourne spectrum to obtain the MFCC coefficients: In the above formula, For the first One MFCC coefficient; yes Size of the spectrogram.

Images