A gas pipeline leakage diagnosis method based on unbalanced data

Abstract

The present application relates to the technical field of leakage diagnosis, and particularly relates to a gas pipeline leakage diagnosis method based on unbalanced data, which comprises collecting acoustic emission signals under different leakage states, and constructing training and test data sets; performing principal component analysis on time-frequency domain signals of the acoustic emission signals to obtain relevant signals; reconstructing the relevant signals into gray scale images; constructing a five-layer multi-layer feature extraction module of a four-layer structure convolution branch and a side convolution branch to perform feature extraction on the gray scale images; constructing a triple weighting module based on channel weighting, position weighting and identity mapping, and inputting output feature maps of the multi-layer feature extraction module into the triple weighting module; inputting output feature maps of the triple weighting module into a working condition discrimination module, and outputting pipeline leakage types. The present application solves the problem that the prediction accuracy of unbalanced data needs to be improved because normal working conditions are much more than abnormal working conditions in gas pipeline monitoring data.

Description

Technical Field

[0001] The present invention relates to the technical field of leakage diagnosis, and in particular to a gas pipeline leakage diagnosis method based on imbalance data. Background Art

[0002] Urban gas pipeline systems are generally built directly outdoors. Non-buried gas pipelines are exposed and easily affected by various external factors, causing corrosion, damage, perforation, cracks and other leakages in the gas pipelines, which seriously threaten the safety of urban residents and bring huge economic losses.

[0003] In addition, since most gas pipelines are installed near residential areas, other ground public facilities are often interspersed around the pipelines. The complex usage environment further increases the possibility of pipelines being hit, squeezed and twisted by humans, resulting in an increased risk of leakage. However, as a special equipment, gas pipelines have far more normal working conditions than leakage or abnormal conditions, and leakage usually only occurs in very rare cases. This results in the normal operating data in the gas pipeline monitoring data being far greater than the leakage data, forming a typical unbalanced data structure. This long-tail data characteristic makes it difficult for conventional fault diagnosis methods to accurately detect leakage signals in a small amount of abnormal data, and it is easy to cause the model to overfit to the normal state, weakening the ability to identify abnormal states.

[0004] Existing gas pipeline detection methods such as magnetic particle method, magnetic flux leakage method, ultrasonic detection method, etc. perform poorly in the case of unbalanced data. They have disadvantages such as limited detection objects, complex operation procedures, and high false detection and missed detection rates. They are difficult to effectively deal with a small number of abnormal signals. In order to ensure the safe and stable operation of gas pipelines and reduce the economic losses and social impact caused by leakage, it is particularly necessary to design an intelligent gas pipeline detection method that can efficiently diagnose leaks in the case of unbalanced data. Summary of the Invention

[0005] In view of the shortcomings of the existing methods, the present invention solves the problem that the accuracy of unbalanced data prediction under normal working conditions far outnumbers abnormal working conditions in gas pipeline monitoring data needs to be improved.

[0006] The technical solution adopted by the present invention is: a gas pipeline leakage diagnosis method based on imbalance data includes the following steps:

[0007] Step 1: Collect acoustic emission signals under different leakage conditions and construct training and test data sets;

[0008] As a preferred embodiment of the present invention, the leakage status includes: healthy, light leakage, moderate leakage and heavy leakage.

[0009] As a preferred embodiment of the present invention, the leakage state is achieved by adjusting the opening of the pipeline valve.

[0010] Step 2: Perform principal component analysis on the time-frequency domain signals of the acoustic emission signal to obtain the relevant signal;

[0011] As a preferred embodiment of the present invention, the acoustic emission signal includes: mean, root mean square, kurtosis, energy, amplitude, rise time, skewness, correlation coefficient, corner rate, frequency, spectral peak, peak slope, and coherence.

[0012] As a preferred embodiment of the present invention, the correlation signal includes: energy, mean value, root mean square, kurtosis, rise time and count.

[0013] Step 3: reconstruct the relevant signal into a grayscale image;

[0014] Step 4: Construct a five-layer multi-layer feature extraction module consisting of a four-layer convolution branch and a side convolution branch to extract features from the grayscale image;

[0015] As a preferred embodiment of the present invention, the first layer of the multi-layer feature extraction module is: 8×8 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the second layer is 7×7 convolution, hole convolution, 3×3 convolution and 1x1 convolution; the third layer is 6×6 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the fourth layer is 5×5 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the fifth layer is 3×3 convolution and 1×1 convolution; the five-layer branch outputs the feature map through softmax.

[0016] Step 5: Construct a triple weighted module based on channel weighting, position weighting and identity mapping, and input the output feature map of the multi-level feature extraction module into the triple weighted module;

[0017] As a preferred embodiment of the present invention, the triple weighted module includes: channel weighting adjusts the input C×H×W of the spatial dimension of the input feature map to the specified output C×1×1 through adaptive average pooling, and through the lifting and lowering channel operation, the input C×1×1 is changed to 2C×1×1 and then to C×1×1; position weighting is the input feature map after global average pooling, edge filling, element averaging and bilinear interpolation; the channel-weighted and position-weighted feature maps are multiplied with the input feature map to obtain the weighted feature map.

[0018] Step 6: Input the output feature map of the triple weighted module into the working condition discrimination module to output the pipeline leakage type;

[0019] As a preferred embodiment of the present invention, the operating condition discrimination module is a fully connected neural network.

[0020] As a preferred embodiment of the present invention, a gas pipeline leakage diagnosis system based on imbalance data includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a gas pipeline leakage diagnosis method based on imbalance data.

[0021] As a preferred embodiment of the present invention, a computer readable medium stores computer program code, and when the computer program code is executed by a processor, a gas pipeline leakage diagnosis method based on imbalance data is implemented.

[0022] Beneficial effects of the present invention:

[0023] 1. This invention introduces acoustic emission technology and neural network models in machine learning into gas pipeline leak diagnosis. It constructs a multi-level feature extraction module that extracts different features from the leak image through five branches. The feature extraction based on the gas pipeline acoustic emission signal can quickly obtain pipeline characterization parameters. Combined with the neural network, it can accurately diagnose pipeline leaks.

[0024] 2. The present invention constructs a triple weighted module to reduce the risk of overfitting and achieve double weighting of important features in key channels and key positions;

[0025] 3. The test cost of the present invention is low. It only needs to collect the acoustic emission signals of the pipeline operation through the acoustic emission system, and then a simple neural network can be established through computer processing to obtain the test results. The characteristic parameter processing and calculation of the acoustic signal are flexible, and multiple different input neurons can be set according to the diagnostic requirements. This method can predict and diagnose leaks in gas pipelines.

[0026] 4. The training data set of the neural network of the present invention is gradually updated and enriched as the number of tests increases, which can significantly improve the efficiency and accuracy of subsequent tests;

[0027] 5. The present invention can be effectively applied in gas pipeline leakage diagnosis and obtain excellent diagnostic results, and is suitable for fault diagnosis of engineering equipment and instruments such as gas pipelines. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the gas pipeline leakage diagnosis method based on imbalance data of the present invention;

[0029] Figure 2 is the principal component analysis result of acoustic emission signal;

[0030] Figure 3 This is a simulation diagram of a special leakage valve for gas pipeline leakage test;

[0031] Figure 4 Schematic diagram of the model structure of the present invention;

[0032] Figure 5 is the confusion matrix result;

[0033] Figure 6 These are the accuracy results when the ratio of normal data to abnormal data is 1:4, 1:10, 1:30, and 1:50. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore only shows the components related to the present invention.

[0035] like Figure 1 As shown, a gas pipeline leakage diagnosis method based on imbalance data includes the following steps:

[0036] Step 1: Combine the gas pipeline test platform and gas pipeline leakage simulation to collect acoustic emission signals under different leakage conditions and construct training and test data sets;

[0037] The gas pipeline test platform is a three-layer coiled structure of the gas pipeline with a total length of 55 meters, a total width of 2 meters, a total height of 1.05 meters, a total wall thickness of 5 mm, and a pipeline diameter of 25 mm.

[0038] Step 2: By controlling the opening, simulate the leakage condition and collect the signals collected at different openings; for the acoustic emission signal;

[0039] Gas pipeline leakage simulation refers to controlling the leakage degree by adjusting the opening of the dedicated leakage valve of the gas pipeline;

[0040] like Figure 3 When the gas pipeline valve opening is 0°, the gas pipeline is in a healthy state and can operate normally; when the gas pipeline valve opening is 15-30°, the gas pipeline is in a mild leakage state, and the gas pipeline emits a slight leakage sound; when the gas pipeline valve opening is 30-45°, the gas pipeline is in a moderate leakage state, and the gas pipeline emits an obvious leakage sound; when the gas pipeline valve opening is 45-60°, the gas pipeline is in a severe leakage state, and the gas pipeline emits a loud leakage sound.

[0041] Feature extraction of gas pipeline leakage acoustic emission signals, including time domain and frequency domain signals. Time domain features include mean, root mean square, kurtosis, energy, amplitude, rise time, skewness, correlation coefficient, corner rate, and count; frequency domain features include signal frequency, spectral peak, peak slope, and coherence.

[0042] like Figure 2Using principal component analysis, the six most relevant signals are analyzed, including acoustic emission energy (AEEnergy), root mean square value (RMS), acoustic emission mean value (AE MeanValue), kurtosis, rise time (AE Rise Time), and ring count.

[0043] Counting refers to the number of times a signal exceeds a set threshold. After the acoustic emission signal is collected, the signal is compared with the preset threshold based on the voltage changes recorded continuously over time, thereby calculating the number of times the threshold is exceeded.

[0044] like Figure 4 ,Step 3, process the collected one-dimensional acoustic signal and then reconstruct it into a two-dimensional reconstructed grayscale image;

[0045] Two-dimensional reconstructed grayscale image is to reconstruct one-dimensional signal data into a two-dimensional matrix, and then convert it into a grayscale image for visualization or feature extraction;

[0046] Step 4: Construct a multi-level feature extraction module to extract features from the two-dimensional reconstructed grayscale image;

[0047] The multi-layer feature extraction module consists of four layers of similar convolutional branches and one side convolutional branch. The first to fourth layers of branches have convolutional layers of different sizes (8×8, 7×7, 6×6, 5×5), void convolution, 3×3 convolution kernel, and 1×1 convolution kernel; the fifth layer of side branch mechanism consists of 3×3 convolution kernel and 1×1 convolution kernel.

[0048] The convolution layer gradually becomes smaller from 8×8, 7×7, 6×6, and 5×5, and its receptive field gradually becomes smaller, which can help the model gradually go from coarse-grained global features to finer-grained local features, making the information at different levels richer; the module can extract features layer by layer at different scales, which helps to capture the continuity of global patterns and local details in the signal; the hollow convolution enhances the receptive field without increasing the amount of calculation, thereby extracting multi-scale features; the combination of 3×3 convolution kernel and 1×1 convolution kernel is designed to complement each other in the asymmetric multi-scale feature extraction module, so that the model can not only extract multi-scale features but also control the complexity of the model.

[0049] The side branches only have 3×3 and 1×1 convolutions to prevent the model from being designed too deep. The gradient may decay during propagation, resulting in information loss or training difficulties.

[0050] The five-layer branches of the multi-layer feature extraction module are summarized and output through the softmax function.

[0051] Step 5: Construct a triple weighted module and use the output features of the multi-level feature extraction module as the input of the triple weighted module;

[0052] The triple weighted module consists of three parts: channel weighting part, position weighting part, and identity mapping part;

[0053] In the channel weighting part, adaptive average pooling is first used to adjust the spatial dimension of the input feature map C×H×W to the specified output C×1×1;

[0054] Channel weighting can reduce the data size and subsequent calculation amount, thereby reducing the risk of overfitting; pooling window height and width k h and k w Calculated by the following formula:

[0055]

[0056] Among them, k h is the pooling window height, k w is the pooling window width, H×W is the original output size, and H′×W′ is the specified output size.

[0057] Channel weighting is achieved through channel changes. Through the up-channel and down-channel operations, the input C×1×1 is transformed into a 2C×1×1 and then into a C×1×1 structure as the output; the up-channel enhances the feature expression, and the down-channel generates weights and restores the dimension, ensuring the effective flow and weighting of information.

[0058] The z of the channel weighted part is the eigenvector before the channel-raising operation, z' is the eigenvector after the channel-raising operation (before the channel-dropping operation), z" is the eigenvector before the channel-dropping operation, W1 is the convolution matrix for the channel-raising operation, and W2 is the convolution matrix for the channel-dropping operation. The formula is:

[0059] z / =W1z(2)

[0060] z / / =W2z / (3)

[0061] The channel weighting part performs sigmoid activation on the output features after channel reduction, so that the weight of each channel is mapped to the range of [0,1]. The weight value represents the importance of each channel. The closer the value is to 1, the more important the feature of the channel is; the closer the value is to 0, the less important the feature of the channel is.

[0062] The position weighted part includes: global average pooling, edge padding, element averaging, and bilinear interpolation.

[0063] The position weighted part is compressed to C×1×1 through global average pooling of the spatial dimensions C×H×W;

[0064]

[0065] Among them, Z c represents the global average value obtained by global average pooling of the cth channel, x c (i, j) represents the eigenvalue of the cth channel at position (i, j); H and W are the height and width of the feature map respectively.

[0066] The position weighted part is expanded to C×(H+pad)×(W+pad) by edge padding operation C×1×1 spatial dimension;

[0067] The position-weighted part is averaged over all elements, and the formula is:

[0068]

[0069] The position weighting part uses bilinear interpolation to adjust the size of the feature map. It is mainly used to scale the two-dimensional data from C×(H+pad)×(W+pad) to C×H×W. It calculates the pixel value of the target position by taking the weighted average of the values ​​of the four surrounding known pixels and their relative distances to the target position. The formula is:

[0070] x'(i'j')=w 00 x(i0j0)+w 01 x(i0j1)+w 10 x(i1j0)+w 11 x(i1j1) (6)

[0071] Among them, x'(i'j') is the target position (i ' ,j ' ) The four nearest known locations around 00 ,w 01 ,w 10 ,w 11 It is the position weight calculated based on the distance between the target position and the surrounding pixels, and the weight satisfies w 00 +w 01 +w 10 +w 11 =1.

[0072] The identity mapping part effectively solves problems such as gradient vanishing, information loss, and training difficulties by simply passing the input directly to the output without any transformation or processing;

[0073] x3=x3 (7)

[0074] The output y of the triple weighted module consists of three parts: the channel weighted part x1, the position weighted part x2, and the identity map x3. The weighted feature map is obtained by multiplying the channel and position weights and the calculated local mean weights as shown in formula (8). This step realizes the double weighting of important features in key channels and key positions. The formula is:

[0075] y=x1x2x3 (8)

[0076] Step 6: Build a working condition discrimination module and use the working condition discrimination model to identify the pipeline leakage type;

[0077] The working condition discrimination module is composed of a fully connected neural network, which includes an input layer, multiple hidden layers and an output layer; the multiple hidden layers include the convolution part and the fully connected part for feature extraction; among them, the convolution part includes: convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, convolution layer 3, and global pooling layer in sequence; the fully connected part includes fully connected layer 1 and fully connected layer 2; fully connected layer 1 consists of 128 neurons, the activation function is ReLU, and the Dropout is set to 0.5; fully connected layer 2 consists of 64 neurons, the activation function is ReLU, and the Dropout is set to 0.2; the output layer uses a softmax activation function containing 4 neurons to convert the output of the model into the probability of four categories; the model is trained with the cross entropy loss function, which is often used for multi-classification problems.

[0078] In this embodiment, the principal component analysis of the gas pipeline acoustic emission signal is performed, and the average acoustic emission energy (AE Energy), root mean square value (RMS), acoustic emission mean (AE Mean Value), kurtosis (Kurtosis), rise time (AE Rise Time), and count (Ring Count) are selected as network input neurons in the time domain signal. After analysis, the required neurons can better characterize the pipeline working conditions.

[0079] By controlling the opening of the special leakage simulation valve of the gas pipeline, as shown in the attached Figure 3 The leakage degree of the gas pipeline is defined as four states: healthy pipeline, slightly leaking pipeline, moderately leaking pipeline, and severely leaking pipeline. When the gas pipeline valve opening is 0°, the gas pipeline is in a healthy state and can operate normally; when the gas pipeline valve opening is 15-30°, the gas pipeline is in a slightly leaking state, and the gas pipeline emits a slight leakage sound; when the gas pipeline valve opening is 30-45°, the gas pipeline is in a moderate leakage state, and the gas pipeline emits an obvious leakage sound; when the gas pipeline valve opening is 45-60°, the gas pipeline is in a severely leaking state, and the gas pipeline emits a loud leakage sound.

[0080] In this embodiment, the data format of the neural network data set is:

[0081] [Input layer: [acoustic emission energy x1, root mean square value x2, acoustic emission mean x3, kurtosis x4, rise time x5, count x6];

[0082] [Output layer: healthy pipeline y1, slightly leaking pipeline y2, moderately leaking pipeline y3, severely leaking pipeline y4];

[0083] In this embodiment, 3000 sets of acoustic emission data were collected through the gas pipeline test platform to construct a data set, of which 2400 sets were normal operation acoustic signals and 600 sets were leakage signals of various degrees; 1800 sets were used for neural network learning and 1200 sets were used for neural network testing. The specific collection process of gas pipeline acoustic emission signals is as follows:

[0084] Before the formal collection, the basic collection settings of the acoustic emission collection system were performed, the sampling frequency of the acoustic emission signal collector was set to 2500kHz, the gain of the preamplifier was set to 40dB, the lock time of the signal collection was set to 2000us, and the peak interval of the signal collection was set to 1000us. Before the formal collection of data, the environmental noise was measured, and then 5-10dB was added to the environmental noise level as the threshold voltage value of the instrument. After the measurement, the threshold value of the acoustic emission instrument was set to 40dB.

[0085] Connect the acoustic emission sensor to the input port and the output port to the output port through a professional transmission cable. Before connecting, it is necessary to visually inspect the acoustic emission sensor for visible damage and cracks. If there are any, the sensor should be replaced in time. When using multiple acoustic emission sensors at the same time, the transmission cables need to be sorted and numbered to ensure that the processor interface, transmission cable and sensor interface correspond to each other. The cables need to be sorted to avoid the cables becoming tangled and difficult to distinguish during use.

[0086] Use pipe sandpaper to remove the yellow paint on the pipe surface to expose the actual pipe body. Apply an even amount of vaseline to the bottom of the sensor for coupling. Use the sensor magnetic bracket to fix the sensor to the polished part of the gas pipe.

[0087] The Hsu-Nielsen signal source was used to perform the lead-breaking operation to rule out the possibility that micro cracks or damage to the acoustic emission sensor would affect the experiment. The sensor calibration results were normal and the sensor could be used normally.

[0088] Control pipeline leakage by simulating valves to collect acoustic emission signals.

[0089] In this embodiment, it is assumed that x=[x1, x2, ..., x d ] T is the input of the network, y=[y1,y2,…,y d ] T is the output of the network. The number of neurons in the input and output layers depends on the number of feature categories and prediction categories of the dataset, respectively.

[0090] The input one-dimensional sound signal features are processed into a two-dimensional image. After passing through a multi-level feature extraction module, various information and their links in the image are extracted to obtain more complete information; then through a multiple weighting module; finally, through a convolutional neural network, the processed signal is identified and classified.

[0091] The confusion matrix is ​​a table layout that supports the visualization of the classification performance of the algorithm model and can intuitively express the target value and the predicted value. This embodiment evaluates and analyzes the gas pipeline diagnosis results. The confusion matrix is ​​shown in the attached figure. Figure 5 Given 960 pipeline signals (actual) from the healthy state, the algorithm model accurately classified 960 signals (predicted) as the healthy working state of the pipeline; for mild leakage, among 80 leakage signals (actual), the algorithm model accurately classified 77 signals (predicted) as mild leakage; for moderate leakage, among 80 leakage signals (actual), the model accurately classified 75 signals (predicted) as moderate signals, and incorrectly classified 5 signals as severe leakage signals; for severe leakage, among 80 leakage signals (actual), the model accurately classified 78 signals (predicted) as severe signals, and incorrectly classified 2 signals as moderate leakage signals.

[0092] The F1 score is the classification performance value of the algorithm model based on precision and recall, so it can be used to evaluate precision and recall.

[0093] This example evaluates and analyzes the gas pipeline diagnosis results. The F1 score and the overall accuracy are as follows: Figure 6 The F1 scores of the method of the present invention for the healthy state, mild leakage, moderate leakage and severe leakage of the gas pipeline are 99.7%, 100%, 95.6% and 95.7% respectively, and the overall accuracy is 97.8%.

[0094] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A gas pipeline leakage diagnosis method based on imbalance data, characterized in that: The following steps are involved: Step 1: Collect acoustic emission signals under different leakage conditions and construct training and test data sets; Step 2: Perform principal component analysis on the time-frequency domain signals of the acoustic emission signal to obtain the relevant signal; Step 3: reconstruct the relevant signal into a grayscale image; Step 4: Construct a five-layer multi-layer feature extraction module consisting of a four-layer convolution branch and a side convolution branch to extract features from the grayscale image; The first layer of the multi-layer feature extraction module is: 8×8 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the second layer is 7×7 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the third layer is 6×6 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the fourth layer is 5×5 convolution, hole convolution, 3×3 convolution and 1×1 convolution; the fifth layer is 3×3 convolution and 1×1 convolution; the five-layer branch outputs the feature map through softmax; Step 5: Construct a triple weighted module based on channel weighting, position weighting and identity mapping, and input the output feature map of the multi-level feature extraction module into the triple weighted module; The triple weighted module includes: channel weighting adjusts the spatial dimension of the input feature map C×H×W to C×1×1 through adaptive average pooling; then through the lifting and lowering channel operation, C×1×1 is converted to 2C×1×1 and then to C×1×1; position weighting performs global average pooling, edge padding, element averaging and bilinear interpolation on the input feature map; the output feature map of the channel weighting and position weighting is multiplied with the input feature map to obtain the weighted feature map; Step 6: Input the output feature map of the triple weighted module into the working condition discrimination module to output the pipeline leakage type; The working condition discrimination module is a fully connected neural network.

2. The gas pipeline leakage diagnosis method based on imbalance data according to claim 1 is characterized in that: Leak status includes: Healthy, Minor Leak, Moderate Leak, and Major Leak.

3. The gas pipeline leakage diagnosis method based on imbalance data according to claim 2 is characterized in that: The leakage status is determined by adjusting the opening of the pipeline valve.

4. The gas pipeline leakage diagnosis method based on imbalance data according to claim 1, characterized in that: Acoustic emission signals include: mean, root mean square, kurtosis, energy, amplitude, rise time, skewness, correlation coefficient, corner rate, count, frequency, spectral peak, peak slope, and coherence.

5. The gas pipeline leakage diagnosis method based on imbalance data according to claim 4 is characterized in that: Related signals include: energy, mean, RMS, kurtosis, rise time, and counts.

6. Gas pipeline leakage diagnosis system based on imbalance data, characterized by: include: a memory for storing instructions executable by the processor; A processor, configured to execute instructions to implement the gas pipeline leakage diagnosis method based on imbalance data as described in any one of claims 1 to 5.

7. A computer-readable medium storing computer program code, characterized in that When the computer program code is executed by a processor, the method for diagnosing gas pipeline leakage based on imbalance data according to any one of claims 1 to 5 is implemented.

Images