Pipeline buried microcrack positioning method based on fusion feature matrix and deep learning

Abstract

The invention discloses a pipeline burial microcrack positioning method based on a fusion feature matrix and deep learning, and belongs to the technical field of material detection. The positioning method comprises the following steps: firstly, selecting a proper mode as an ultrasonic guided wave excitation signal, arranging a group of circular sensor arrays at a specified position on the surface of a metal pipeline, and then acquiring zero-frequency component characteristics of a to-be-detected area through a receiving sensor; point excitation is adopted to emit ultrasonic guided wave excitation signals, damage index features of a to-be-detected area are calculated by receiving a sensor frequency spectrum, and finally zero-frequency component features are spliced with the damage index features to construct a fusion feature matrix to construct a data set for training a deep learning model to carry out pipeline burial microcrack positioning. Experimental data show that the buried microcrack positioning can be realized on the premise of lower error by applying the deep learning model constructed by the method disclosed by the invention, the buried microcrack outside the positioning area can be reliably positioned, and the method has good generalization in the aspect of buried microcrack positioning.

Description

Technical Field

[0001] The present invention belongs to the technical field of material detection, and specifically discloses a method for locating buried microcracks in a pipeline based on a fusion feature matrix and deep learning. Background Art

[0002] As one of the five major means of transportation, pipelines play a vital role in the construction of the national defense industry and the national economic construction. Since pipelines are generally installed in harsh environments such as underground and underwater, with the accumulation of service time, they often develop cracks, holes and other damages due to external forces, bacterial erosion and other reasons, which in turn lead to pipeline leakage or even explosion, which will not only cause property losses, but also threaten life safety, cause environmental pollution and social chaos.

[0003] In the existing technology, the research on pipeline microcrack positioning often only focuses on through-going microcracks. However, in real life, buried microcracks are more difficult to find and more dangerous than through-going microcracks, but they are often ignored in research. At the same time, since the nonlinear effect produced by the interaction between ultrasonic guided waves and buried microcracks is too weak, it is difficult to use nonlinear ultrasonic guided waves to locate buried microcracks. Summary of the invention

[0004] Technical problem to be solved by the invention

[0005] In view of the difficulty in detecting buried microcracks in the prior art, the present invention designs a pipeline buried microcrack positioning method based on a fusion feature matrix and a deep learning model based on the characteristics of buried microcracks.

[0006] Technical Solution

[0007] To achieve the above purpose, the technical solution provided by the present invention is:

[0008] The pipeline buried microcrack location method based on fusion feature matrix and deep learning includes the following steps:

[0009] Step 1: Obtain the dispersion curve of ultrasonic guided waves of metal pipes and select a suitable mode as the ultrasonic guided wave excitation signal.

[0010] Step 2, setting the metal pipe area to be inspected, setting an excitation sensor at the end face of the pipe, and setting a group of circular sensor arrays on the surface of the pipe in the area to be inspected as receiving sensors to receive the echo signal of the area to be inspected;

[0011] Step 3, symmetrically transmitting ultrasonic guided wave excitation signals on the upper and lower surfaces of the metal pipe at the excitation sensor, and obtaining the zero-frequency component characteristics of the area to be detected through the receiving sensor;

[0012] Step 4: For the damaged pipeline and the non-destructive pipeline, a receiving sensor in the area to be detected is used as an excitation sensor, and other receiving sensors are used as receiving sensors. The ultrasonic guided wave excitation signal is emitted by point excitation, and the damage index characteristics of the area to be detected are calculated through the receiving sensor spectrum;

[0013] Step 5, concatenating the zero-frequency component feature obtained in step 3 with the damage index feature in step 4 to construct a fusion feature matrix, and multiple fusion feature matrices constitute the original data set;

[0014] Step 6, divide the original data set obtained in step 5 into a training set, a validation set, and a test set, use the training set to train the deep learning model parameters, use the validation set to verify the process model, and select the deep learning model with the best performance;

[0015] Step 7: Use the deep learning model trained in step 6 to locate buried microcracks on the test set.

[0016] Furthermore, in step 3, the method for extracting the zero-frequency component feature of the area to be detected is: the excitation sensor sends an ultrasonic guided wave, the received time domain signal is extracted at the receiving sensor and converted into the frequency domain, and the amplitude of the frequency domain signal at zero frequency is extracted as a feature.

[0017] Furthermore, the method for extracting the damage index feature of the area to be detected in step 4 is: select a certain receiving sensor as the excitation sensor in the area to be detected, use other receiving sensors as receiving sensors, send excitation signals to obtain the receiving sensor spectrum, and then calculate the damage index DI by the following formula:

[0018]

[0019] where f k is the frequency of the spectrum, FI and FD are the amplitudes of the corresponding frequencies of the spectrum of the lossless pipeline and the lossy pipeline, respectively, s is the sampling rate.

[0020] Furthermore, the method for selecting the excitation signal in step 1 is: according to the dispersion curve of the metal pipeline, a low-dispersion, fast-speed L(0,2) modal signal is selected as the excitation signal.

[0021] Furthermore, the deep learning model trained in step 6 is a convolutional neural network model.

[0022] Beneficial Effects

[0023] The method of the present invention fuses the zero-frequency component and the damage index features, and the constructed feature matrix contains rich buried microcrack location information. The trained deep learning network model can predict the location of buried microcracks with a low error.

[0024] The method proposed in the present invention combines the fusion feature matrix and the deep learning model, which breaks through the limitations of traditional methods. The results obtained in locating buried microcracks outside the positioning area also have good reliability and good generalization in locating buried microcracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention;

[0026] Figure 2 This is a schematic diagram of the experimental pipeline model of the method of the present invention;

[0027] Figure 3 A schematic diagram of a two-dimensional positioning area of ​​the method of the present invention;

[0028] Figure 4 (a) is the experimental result of the dispersion curve group velocity of ultrasonic guided waves in the pipeline according to the method of the present invention;

[0029] Figure 4 (b) is the experimental result of the dispersion curve group phase velocity of ultrasonic guided waves in the pipeline of the method of the present invention;

[0030] Figure 5 Schematic diagram of the steps of training the convolutional neural network model of the present invention. DETAILED DESCRIPTION

[0031] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and specific implementation methods.

[0032] In practical applications, both the zero-frequency component and the damage index contain the location information of buried microcracks, but both characteristic factors have their own shortcomings. Although the zero-frequency component has stronger nonlinear characteristics than the double frequency, since the detected object is buried microcrack damage, the amplitude of the zero-frequency component changes only when the damage position changes. Considering that there are 1353 different situations for the location of buried microcracks, it is not enough to study the location information of buried microcracks by relying solely on the change of this characteristic factor; the damage index is the calculated value of the degree of damage on each path. In order to achieve the most accurate positioning, an array composed of as many excitation point-receiving point sensors as possible is required, which puts higher requirements on the configuration of the experimental equipment. Reducing the number of arrays requires adding other characteristic factors to improve the positioning accuracy. Based on the above analysis, the two characteristic factors of zero-frequency component and damage index are combined in the method of the present invention to construct a fusion feature matrix containing zero-frequency component and damage index. The matrix has more damage location information and can achieve accurate positioning of buried microcracks in pipelines.

[0033] like Figure 1As shown, the pipeline buried microcrack location method based on the fusion feature matrix and deep learning of the present invention includes four stages: signal acquisition, data processing, model training, and prediction results, and mainly includes the following steps:

[0034] Step 1: Obtain the dispersion curve of ultrasonic guided waves of metal pipes and select a suitable mode as the ultrasonic guided wave excitation signal. In this embodiment, the L(0,2) mode signal with low dispersion and high speed is selected as the excitation signal;

[0035] Step 2: Set the target area of ​​the metal pipe to be detected, set an excitation sensor outside the target area, and set a group of circular sensor arrays at the specified position on the surface of the metal pipe as receiving sensors to receive the echo signal of the area to be detected.

[0036] Step 3: symmetrically transmit ultrasonic guided wave excitation signals on the upper and lower surfaces of the metal pipe at the excitation sensor, and obtain the zero-frequency component characteristics of the area to be detected through the receiving sensor.

[0037] Step 4: Use a sensor in the circular sensor array of the area to be detected as an excitation sensor, and other sensors as receiving sensors. Use point excitation to emit ultrasonic guided wave excitation signals at the excitation sensor, and calculate the damage index characteristics of the area to be detected through the frequency spectrum of the receiving sensor.

[0038] Step 5, fusing the zero-frequency component obtained in step 3 with the damage index feature obtained in step 4 to construct a fusion feature matrix, and constructing an original data set consisting of multiple fusion feature matrices;

[0039] Step 6: Divide the original data set into a training set, a validation set, and a test set, use the training set to train the deep learning model parameters, use the validation set to verify the process model, and select the deep learning model with the best performance;

[0040] Step 7: Use the deep learning model trained in step 6 to locate buried microcracks on the test set.

[0041] The present invention uses ABAQUS software to establish a three-dimensional pipeline simulation model, and uses the Abaqus / Explicit dynamic analysis method to simulate the propagation of nonlinear ultrasonic guided waves under different microcrack characteristics. The simulation model is as follows Figure 2 As shown in Figure 1, the overall model is a three-dimensional pipe with an axial length of 1000 mm, an outer diameter and an inner diameter of 70 mm and 64 mm respectively. The material is stainless steel, and the specific parameters are shown in Table 1.

[0042] Table 1 Experimental pipeline material parameters

[0043]

[0044]

[0045] like Figure 2 As shown, the microcrack location detection area is located at the center of the upper half of the pipeline model, with an axial length of 220 mm and a circumferential angle of 180 degrees. The center plane of the location detection area on the XY plane is plane 1, and the center plane on the XZ plane is plane 2. A sensor array containing eight sensors (A to H) is set on the upper surface of the circular tube around the location area to capture the response signal from the damaged area, wherein the axial position of sensors A, B, E, and F is 40 mm away from the boundary of the location detection area, and the circumferential position sensors A, B, E, and F are located on both sides of plane 2, respectively, and are spaced 40 degrees from plane 2; the axial position of sensors C, D, G, and H is 50 mm away from plane 1, and the circumferential position is 30 degrees away from the boundary of the location detection area.

[0046] In the microcrack model, since the microcrack type is buried microcrack, component 1 and component 2 are constructed separately for damage modeling, where component 1 is the damage external model and component 2 is the damage regional model. The two are bound to form a unified whole with unified properties. The buried microcrack is located in the center of the microcrack model, with a height of 4 mm, a microcrack length of 20 mm, and an angle of about 37 degrees with the Z axis. For the convenience of modeling, Figure 2 As shown, a cohesive unit with the same value as the buried microcrack is embedded in the middle of component 2 to simulate microcrack damage, and the interaction characteristics of tangential frictionlessness and normal hard contact are adopted for the surface generating microcracks.

[0047] Since the depth of the microcracks is fixed, the three-dimensional pipeline is simplified into a two-dimensional plane for simplicity of operation. Figure 3 The following is a schematic diagram of the two-dimensional positioning detection area. A new XZ coordinate system is established with the center point of the positioning detection area as the origin (0, 0). The X-axis direction is Figure 2 The circumferential angle in the Z-axis direction is Figure 2 The axial length in the three-dimensional coordinate system (consistent with the Z direction of the aforementioned three-dimensional coordinate system). The coordinates of the center point of the buried microcrack represent its position. The distribution in the X-axis direction changes from -80 degrees to 80 degrees, with an interval of 5 degrees. The Z-axis direction changes from -100mm to 100mm, with an interval of 5mm. The intersection of the horizontal and vertical directions is the coordinates of the center point of the crack. Taking the buried microcrack in the lower left corner as an example, the coordinates (-80, -100) can be used to represent the position information. The entire modeling process simulated 1,353 buried microcrack damages in different positions. When calculating the positioning error later, it is only necessary to convert the circumferential distance of the X-axis into the corresponding arc length and calculate the distance difference between the predicted coordinates and the actual coordinates.

[0048] according to Figure 4The dispersion curve of ultrasonic guided waves in the pipeline is shown in Figure 1. The simulation excitation signal selects a single-frequency signal, which consists of a 10-cycle Hanning window plus a windowed sine signal with a center frequency of 50 kHz. During the simulation analysis, although the excitation signal is the same, the selection of the excitation point will change according to the change of the characteristic factors used for subsequent positioning research.

[0049] The zero-frequency component is essentially related to the nonlinear effect caused by the interaction between ultrasonic guided waves and microcracks, and has a single mode. Therefore, it can be more conveniently extracted to analyze the time domain signal. In this embodiment, the excitation point position is set using end face excitation, that is, the same excitation is applied at the center line node of the end face along the axial direction of the pipeline at the same time to obtain the required L(0,2) mode ultrasonic guided wave.

[0050] The calculation of the damage index can include signals of different modes, and it is necessary to calculate the signals of different excitation point-receiving point paths. Therefore, point excitation is adopted as the excitation method. Sensors AH are selected as the excitation points respectively, and the excitation signal is applied along the Z-axis direction (axial direction). When one sensor excites the signal, the other seven sensors receive the corresponding damage signal.

[0051] In this embodiment, it should be specifically noted that in order to ensure stability and convergence in the solution process and at the same time ensure the calculation efficiency and accuracy of the model, the mesh unit division size and analysis time step in the finite element analysis need to meet the following requirements:

[0052]

[0053]

[0054] Among them, L max and L min is the maximum and minimum value of the unit length, λ min represents the shortest wavelength of the excitation signal, Δt is the time step, c g is the propagation velocity of ultrasonic guided waves.

[0055] Since the nonlinearity generated by buried microcracks is weaker than that of penetrating microcracks, it is necessary to ensure higher finite element analysis calculation accuracy. In this embodiment, the mesh unit size of the pipeline model is set to 2mm and the time step is 30ns. In addition, the positioning area and buried microcrack modules require finer meshes, so the local unit sizes of these two parts are set to 1mm and 0.5mm respectively.

[0056] Table 2 Schematic diagram of feature matrix construction

[0057]

[0058] Table 2 is a schematic table of the buried microcrack fusion feature matrix with the position coordinate (0, 0) as an example. First, the zero-frequency component characteristic factor is calculated, and the L(0, 2) mode ultrasonic guided wave is excited on the end face of the pipeline. The time domain signals under the damage state are received by sensors AH respectively and processed in the frequency domain. The amplitude of the frequency domain signal at zero frequency is extracted as the feature, and a total of 8 eigenvalues ​​are obtained, as shown in the last column of Table 2.

[0059] The damage index is calculated for the damaged and non-damaged pipelines respectively. The 8 sensors included in the sensor array can all be used as excitation point sensors. Taking sensor A as the excitation point as an example, the excitation signal is excited along the Z axis at sensor A, and the remaining 7 sensors BH receive the response signals respectively. The damage index DI is calculated by the following formula:

[0060]

[0061] where f k is the frequency of the spectrum, FI and FD are the signal spectrum amplitudes of the lossless pipeline and lossy pipeline structures respectively, and f s is the sampling rate.

[0062] Calculate the damage index under the corresponding paths AB, AC, ... AH. When the buried microcrack position coordinates are (0, 0), the simulation model can obtain a total of 56 paths, and 56 damage index features can be obtained after calculation.

[0063] The above 8 zero-frequency component features and 56 damage index features are arranged as shown in Table 2 to obtain a fusion feature matrix of an actual size of 8 × 8. Through the above operation, the response signal received by the sensor at a specific damage position can be converted into a fusion feature matrix, and a total of 1353 matrices with different data contents can be obtained by changing the position of the buried microcracks.

[0064] Although the above matrix contains damage location information, it is difficult to obtain the coordinate position of buried microcracks by analyzing the matrix information due to the large amount of data. Therefore, all the matrix data obtained need to be used as the input of the deep learning model, and a network model can be generated after multiple iterative updates. When the sensor receives a new signal, it only needs to obtain the fusion feature matrix through the above processing, and input the matrix data into the trained model to obtain the coordinate prediction of the buried microcracks, thereby achieving the purpose of accurately locating the buried microcracks.

[0065] The entire buried microcrack coordinate information dataset contains a total of 1353 sets of data, which are randomly divided into training set, validation set and test set in a ratio of 8:1:1. The dimension of each set of data is 8×8, and the dimension expansion operation is used to generate inputs with channels W, H and C of 8, 8 and 1 respectively.

[0066] In the model training phase, this embodiment adopts Figure 5 The convolutional neural network shown in Figure 1 is used to generate the optimal model for predicting the location coordinates of buried microcracks. Convolutional Neural Network (CNN) is a feedforward neural network composed of basic structures such as convolutional layer, pooling layer and fully connected layer. It filters the input information through the convolutional layer and is a typical algorithm in the field of deep learning.

[0067] For a buried microcrack at a certain position, a characteristic vector x∈ can be obtained by processing the response signal 8×8 , named as a sample, and the corresponding coordinate information is y∈ 1×2 , named as a set of labels. The CNN model can be used as a parametric regression model to implement the mapping function F: F(x)→y. The parameter θ is used to represent the process of fitting the sample x and its corresponding coordinate information y. The process can be formulated as an optimization problem, the purpose of which is to reduce the loss function L(θ) representing the mean squared error (MSE) between the model output and the label:

[0068] L(θ)=||F(x)-y|| 2

[0069] like Figure 5 As shown, first, the 8×8 input matrix shown in Table 2 is passed through the first basic block BasicBlock(8-1-1), which contains a convolution layer with 8 input channels, a convolution kernel size of 1×1, and a stride of 1. The process includes three steps:

[0070] First, the input matrix is ​​convolved through a convolutional layer (Conv2d) as follows:

[0071]

[0072] Among them, y1(i,j) represents the input matrix, W1 represents the convolution kernel, the size is 1×1, y1_conv[i,j] represents the value of the convolution output at position (i,j), and Respectively represent the sum of the rows and columns of the convolution kernel. This step keeps the size of the matrix unchanged through the 1×1 convolution operation, and adjusts the number of channels of the matrix to 8.

[0073] Then the output of the previous step is normalized by batch normalization (BatchNorm2d) as follows:

[0074]

[0075] Where y1_conv represents the output of the convolution operation, μ1 represents the mean of the current batch of data, and σ1 2 represents the variance of the current batch of data, ε is a small positive number to avoid the denominator being zero, γ1 and β1 are learnable scaling and offset parameters, and y1_bn represents the output of batch normalization.

[0076] Finally, the input matrix x is transformed nonlinearly through the ReLU activation function as follows:

[0077] y1_relu=max(0,x),

[0078] Among them, x represents the input matrix, and y1_relu represents the output of the ReLU activation function, that is, all negative values ​​in the input matrix are set to 0.

[0079] Next, the matrix passes through Basic Block (8-3-2), Basic Block (16-3-2) and BasicBlock (32-2-2) in turn. Each basic block contains a convolution layer with the following parameters:

[0080] Basic Block (8-3-2): The number of output channels is 8, the convolution kernel size is 3×3, and the stride is 2;

[0081] Basic Block (16-3-2): The number of output channels is 16, the convolution kernel size is 3×3, and the stride is 2;

[0082] Basic Block (32-2-2): The number of output channels is 32, the convolution kernel size is 2×2, and the stride is 2.

[0083] These basic blocks gradually reduce the size of the matrix while increasing the number of channels to extract higher-level features. After passing through all the basic blocks, the resulting matrix size is 1×1 and the number of channels is 32. This 1×1×32 matrix is ​​then flattened into a 32-dimensional vector through a Squeeze operation and passed through a Dropout layer to reduce overfitting, as follows:

[0084] y_dropout = m*y,

[0085] Where y_dropout represents the output of the neuron during the training phase after applying Dropout, y represents the original output of the neuron, and m is a Bernoulli random variable with probability p being 1 and probability 1-p being 0. The Dropout layer randomly sets the output of some neurons to 0 with a certain probability to prevent over-reliance on certain features.

[0086] In the next step, the 32-dimensional vector is processed by multiple fully connected layers (FC). The processing process in this embodiment is: first, a fully connected layer that reduces the dimension from 32 to 1, then a Dropout layer, followed by a fully connected layer that maintains the dimension as 1, and finally an output layer that maps the dimension from 1 to 2. Finally, a Squeeze operation is performed to remove the dimension of dimension 1, and finally a 2-dimensional output vector is obtained, which indicates the probability or confidence that the sample belongs to 2 categories.

[0087] During the training process, the convolutional neural network calculates the output of a batch of input data based on the current parameters, and compares the output with the corresponding label to calculate the loss function (such as cross entropy loss, mean square error MSE, etc.). Then the gradient of the loss function to each parameter is calculated through the back propagation algorithm, and the parameters of the network are updated using the optimization algorithm (such as stochastic gradient descent SGD, Adam, etc.) to minimize the loss function. The above process is repeated for multiple rounds, and the network obtained at the end of each round is evaluated on the validation set. When the performance on the validation set no longer improves, the training stops, and the final output result is the microcrack coordinates corresponding to the matrix.

[0088] This embodiment uses the Adam optimization algorithm as the optimizer, the mean square error MSE as the loss function, the batch size is 32, the learning rate is 0.001, and the optimizer is defined using fixed step decay. The learning rate adjustment parameter step_size is 2000, and the parameter gamma indicating the degree of decay is 0.5.

[0089] When calculating the mean square error (MSE), since the corresponding label information is the length in the axial direction and the position in the angular direction, the formula The coordinate quantity representing the circumferential angle is converted into the coordinate quantity corresponding to the arc length, where b is the coordinate quantity representing the circumferential angle and a is the length value representing the corresponding arc length.

[0090] During the training process, in addition to using the MSE loss function, the mean normalized distance (MND) is also used to measure the error. The index is defined as:

[0091]

[0092] MND measures the average positioning error of all samples, where N is the number of samples, {y1 (n) ,y2 (n)}and are the actual position coordinates and predicted position coordinates of the nth sample, respectively. In this embodiment, after 1000 training cycles, the MND of the trained CNN model on the validation set is reduced to 5.28 mm.

[0093] After the model training is completed, the test set data is used to verify the model. The average normalized distance finally obtained by the 135 groups of test set data in this embodiment is 4.71 mm, which proves that the trained convolutional neural network model can better complete the task of locating buried microcracks in pipelines.

[0094] The localization effect of microcrack damage outside the experimental scene of the trained convolutional neural network model needs to be further verified. Three sets of buried microcrack models were re-established, and their center coordinates were (-90, 0), (-100, 0) and (-110, 0), respectively. The fusion feature matrix was constructed, and the optimal network model was used to predict their positions. The positioning prediction results are shown in Table 3. It can be seen that the method of the present invention has good generalization ability for the precise positioning of buried microcracks in pipeline structures.

[0095] Table 3 Microcrack location test results

[0096]

[0097] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A pipeline buried microcrack location method based on fusion feature matrix and deep learning, characterized in that: The following steps are involved: Step S1, obtaining the dispersion curve of ultrasonic guided waves of the metal pipeline and selecting a suitable mode as an ultrasonic guided wave excitation signal; Step S2, setting the metal pipe area to be inspected, setting an excitation sensor at the end face of the pipe, and setting a group of circular sensor arrays on the surface of the pipe in the area to be inspected as receiving sensors to receive the echo signal of the area to be inspected; Step S3, symmetrically transmitting ultrasonic guided wave excitation signals on the upper and lower surfaces of the metal pipe at the excitation sensor, and obtaining zero-frequency component characteristics of the area to be detected through the receiving sensor; Step S4, for the damaged pipeline and the non-damaged pipeline, a certain receiving sensor in the area to be detected is used as an excitation sensor, and other receiving sensors are used as receiving sensors, and ultrasonic guided wave excitation signals are emitted by point excitation, and damage index characteristics of the area to be detected are calculated through the receiving sensor spectrum; Step S5, concatenating the zero-frequency component feature obtained in step S3 with the damage index feature obtained in step S4 to construct a fused feature matrix, and multiple fused feature matrices constitute the original data set; Step S6, dividing the original data set obtained in step S5 into a training set, a validation set and a test set, using the training set to train the deep learning model parameters, using the validation set to validate the process model, and selecting the deep learning model with the best performance; Step S7, using the deep learning model trained in step S6 to locate buried microcracks on the test set.

2. The pipeline buried microcrack location method based on fusion feature matrix and deep learning as claimed in claim 1 is characterized in that: The method for extracting the zero-frequency component feature of the area to be detected in step S3 is: the excitation sensor sends ultrasonic guided waves, the receiving sensor extracts the received time domain signal and performs frequency domain conversion, and extracts the amplitude of the frequency domain signal at zero frequency as a feature.

3. The pipeline buried microcrack location method based on fusion feature matrix and deep learning as claimed in claim 1 is characterized in that: The method for extracting the damage index feature of the area to be detected in step S4 is: select a certain receiving sensor as the excitation sensor in the area to be detected, use other receiving sensors as receiving sensors, send excitation signals to obtain the receiving sensor spectrum, and calculate the damage index DI by the following formula: in fk is the frequency of the spectrum, FI and FD are the amplitudes of the corresponding frequencies of the spectrum of the lossless pipeline and the lossy pipeline, respectively. fs is the sampling rate.

4. The pipeline buried microcrack location method based on fusion feature matrix and deep learning as claimed in claim 1 is characterized in that: The method for selecting the excitation signal in step S1 is: according to the dispersion curve of the metal pipeline, a low-dispersion, fast-speed L(0,2) modal signal is selected as the excitation signal.

5. The method for locating buried microcracks in pipelines based on fusion feature matrix and deep learning as claimed in claim 1, characterized in that: The deep learning model trained in step S6 is a convolutional neural network model.

Images