A pipeline fault diagnosis method combining defect detection and dimension estimation

Abstract

The application belongs to the technical field of pipeline fault diagnosis, and discloses a pipeline fault diagnosis method combining defect detection and size estimation, which comprises collecting original magnetic flux leakage signals for measuring the health state of a pipeline and preprocessing to obtain complete magnetic flux leakage signals; complete magnetic flux leakage signals are labeled and segmented to obtain a magnetic flux leakage data set; the magnetic flux leakage data set is input into a shared feature extraction network to extract shared features of defects; the shared features are input into a feature decomposition network to generate private features for specific tasks; the private features are input into a region candidate network to generate defect candidate regions and physical information; the candidate regions and the private features are jointly input into a region pooling layer network to obtain standard features; the standard features are input into a parallel decision sub-network guided by the physical information; the difference between the physical information, network prediction and true value is minimized to update the parameters of the fault diagnosis network, and a fault diagnosis model for detecting defect positions and quantifying defect sizes is obtained. The method can simultaneously realize high-precision defect detection and size estimation.

Description

Technical Field

[0001] This invention relates to the field of pipeline fault diagnosis technology, and more specifically to a pipeline fault diagnosis method that combines defect detection and size estimation. Background Technology

[0002] Pipeline transportation is one of the main methods for transporting oil and gas resources over long distances. However, due to the influence of the transported medium and the external environment, pipelines inevitably develop defects that compromise their integrity. If these defects are not detected in time, they will become increasingly corroded over time, potentially leading to perforation, energy leaks, and explosions. This seriously threatens the stable operation of the pipeline and the safe transportation of oil and gas resources.

[0003] Magnetic flux leakage (MF) testing is widely used in pipeline safety inspection, primarily involving two core tasks: defect detection and defect size estimation. Defect detection aims to identify and locate defects in the pipeline, while size estimation methods are used to estimate the length, width, and depth of identified defects. These analytical results provide important guidance for evaluating pipeline health and are crucial for safe pipeline operation. However, due to the complexity of these two tasks, existing literature often treats them as independent tasks, lacking coordination and interaction, resulting in long analysis times and poor consistency of results.

[0004] Therefore, there is an urgent need for a pipeline fault diagnosis method that combines defect detection and size estimation to ensure mutual reinforcement and coordination between the two, thereby improving the accuracy of defect detection and size estimation. Summary of the Invention

[0005] The purpose of this invention is to address the problem of the inability to jointly optimize pipeline defect detection and defect size estimation in the prior art, as well as the problem of poor consistency of results, and to provide a pipeline fault diagnosis method that combines defect detection and size estimation.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A pipeline fault diagnosis method combining defect detection and defect size estimation includes:

[0008] S1. Obtain the raw magnetic flux leakage signal to measure the health status of the pipeline; preprocess the raw magnetic flux leakage signal to obtain a complete magnetic flux leakage signal;

[0009] The complete magnetic flux leakage signal is segmented and labeled to serve as the magnetic flux leakage dataset required for training the fault diagnosis network. The fault diagnosis network includes a shared feature extractor, a private feature decomposition network, a region candidate network, a region pooling layer network, and a parallel decision sub-network.

[0010] S2. Input the magnetic flux leakage dataset into the shared feature extractor to extract the deep shared features of the defects;

[0011] S3. Input the obtained deep shared features into the private feature decomposition network to generate features for specific tasks, including specific detection features and specific quantization features;

[0012] S4. Input the features of a specific task into the region candidate network to generate defect candidate regions and physical information;

[0013] S5. Input the defect candidate region and the features of the specific task together into the region pooling layer network composed of pooling layer networks to obtain standard features;

[0014] S6. Construct a parallel decision sub-network guided by physical information based on convolutional layers, feed standard features into the parallel decision sub-network, and obtain the prediction output of the parallel decision sub-network.

[0015] S7. By minimizing the difference between physical information and the true value, and the difference between the predicted output of the parallel decision sub-network and the true value, the parameters of the fault diagnosis network are updated to obtain the final fault diagnosis model. Based on the final fault diagnosis model, pipeline defect detection and size estimation are performed to obtain the location of the defect and the length, width and depth of the defect.

[0016] The preprocessing specifically includes:

[0017] Step 1.1: Perform Gaussian filtering on the original leakage magnetic field signal M to obtain the filtered signal B;

[0018] Step 1.2: Based on the differences in sensor baseline values, use formula (1) to unify the dimensions of different sensors:

[0019]

[0020] Among them, X i,j It is the signal of the j-th channel and the i-th sampling point after correction, where q and e represent the number of sampling points and the number of sensors, respectively; m and n are intermediate variables, where m is a number between 1 and q, and n is a number between 1 and e.

[0021] The process of obtaining the magnetic flux leakage dataset is as follows:

[0022] Complete the leakage magnetic signal by channel, divide the sample according to the r×w sliding window, remove invalid samples without defects, and convert it into a pseudo-color image;

[0023] Labels for defect locations and sizes in the pseudo-color image are artificially constructed based on expert experience. The location information of each defect is denoted as Z = [x min ,y min ,x max,y max ], where x min y min Let x be the true coordinate of the top-left corner of the target in the pseudo-color image. max y max The coordinates of the lower right corner of the target in the pseudo-color image are true values; the size information of each defect is denoted as q = [l, w, d], where l, w, and d are the true values ​​of the length, width, and depth of the defect, respectively.

[0024] The labeled and complete magnetic flux leakage signal data are divided into training set S and test set T in an 8:2 ratio.

[0025] The deep shared features of the extraction defects specifically include:

[0026] Step S2.1: Construct a shared feature extractor Γ() based on the concatenation of a pre-trained residual 50 network and a feature pyramid network;

[0027] Step S2.2: Transfer the training set samples s i ∈S is fed into the shared feature extractor, and the deep shared feature f of the defect is obtained using formula (2):

[0028] f=Γ(s,μ) (2)

[0029] Where μ is a parameter of the shared feature extractor.

[0030] The features for generating specific tasks include specific detection features and specific quantization features, specifically including:

[0031] Step S3.1: Construct a private feature decomposition network based on a convolutional neural network and a pooling neural network, denoted as the detection private network T(·) and the quantization private network Y(·), respectively; the detection private network and the quantization private network have the same structure but different parameters, and are responsible for outputting specific detection features and specific quantization features, respectively; the private feature decomposition network includes a horizontal pooling network, a vertical pooling network, a convolutional network, a normalization layer, and a nonlinear layer;

[0032] Step S3.2: Input the deep shared features f of the defects into the horizontal pooling network and the vertical pooling network respectively to generate pooling features A. x and A y ;

[0033] Step S3.3: Pool feature A x and A y The intermediate layer features are obtained by sequentially inputting the convolutional network, normalization layer, and nonlinear layer. and

[0034] Step S3.4: Based on intermediate layer features The deep shared features f of the defects are used to obtain the channel private features f using formulas (3) and (4). ch and spatial private features f sp :

[0035]

[0036]

[0037] in, Represents multiplication, ε(·) represents the sigmoid nonlinear function, W0, W1, b0, b1 represent learnable parameters, and Dconv 3×3 Conv represents the parameters of a deformable convolution with a kernel of 3. 1×1 The convolution parameter represents a kernel of 1, and Norm(·) represents a normalized layer;

[0038] Step S3.5: Based on channel private features f ch and spatial private features f sp The adaptive fusion of the two is achieved through formula (5) to obtain specific detection features f. d :

[0039]

[0040] Step S3.6: Feed the deep shared feature f of the defect into the quantization private network Υ(·) and repeat steps S3.1 to S3.5 above to obtain the specific quantization feature f. q .

[0041] The generation of candidate defect regions and physical information specifically includes:

[0042] Step S4.1: The region candidate network is composed of two layers of shared convolutional neural network and two parallel fully connected networks connected in series, wherein the number of output nodes of the two parallel fully connected networks are 2 and 4 respectively;

[0043] Step S4.2: Apply specific detection features f d and specific quantitative features f q The data is fed into the regional candidate network, and the defect candidate region Θ and physical information P are obtained using formula (6):

[0044]

[0045] Where W4,b4 represents learnable parameters, and Φ(·) represents the physical information calculation method.

[0046] The acquisition of standard features specifically includes:

[0047] Step S5.1: The region pooling layer network RD (·) and the quantized region pooling layer network R Q (·) consists of pooling layer networks of different lengths, with network output sizes of 7×7×256 and 14×14×256, respectively.

[0048] Step S5.2: Apply specific detection features f d Specific quantitative features f q The candidate defect regions are respectively input into the detection region pooling network R. D (·) and the quantized region pooling layer network R Q (·), the detection standard feature f is obtained using formula (7). dn and quantification standard features f qn :

[0049]

[0050] The acquisition of the network's prediction output specifically includes:

[0051] Step S6.1: The parallel decision subnetwork includes a detection subnetwork and a quantization subnetwork. The detection subnetwork is constructed from two parallel fully connected layer networks, with 2 and 4 output nodes in each fully connected layer, respectively. The quantization subnetwork is constructed from three parallel fully connected layer networks, each with 1 output node.

[0052] Step S6.2: Detect the standard feature f dn Quantitative standard features f qn The physical information P is fed into the detection subnetwork and the quantization subnetwork respectively, and the network's predicted output is obtained using formula (8):

[0053]

[0054] Where c,t represents the location and category predicted by the detection subnetwork, p l ,p w ,p d These represent the length, width, and depth of the defect predicted by the quantized subnetwork, respectively. θ represents the parameters of the detection subnetwork and the quantization subnetwork, respectively.

[0055] The parameter update of the fault diagnosis network specifically includes:

[0056] Step S7.1: Based on the physical information P and the predicted output of the detection sub-network, calculate the detection loss function Loss using formula (9). d :

[0057]

[0058] Where, N cls Nreg This represents the number of classification anchor boxes and regression anchor boxes used during the training of the candidate network for the region; Represents the classification loss function; ρ represents the normalized weights; This represents the location regression loss function; The label represents the true value, and its value is 0 or 1. It is 1 for a positive sample and 0 for a negative sample; p i t represents the probability that the anchor box is predicted to be the target; i Parametric coordinates representing the predicted location; Parametric coordinates representing the actual labeled location; Let p represent the physical constraint loss function, where p i This represents the physical information of the predicted defect candidate region; Labels indicating physical information about defects;

[0059] Step S7.2: Based on the predicted output of the quantization sub-network, calculate the quantization loss function Loss using formula (10). q :

[0060]

[0061] Step S7.3: Add the detection loss function and the quantization loss function, and update the parameters of the fault diagnosis network by minimizing formula (11) to obtain the final fault diagnosis model:

[0062] Loss = αLoss d +(1-α)Loss q (11)

[0063] Here, α is an adjustable factor.

[0064] The beneficial effects of this invention are as follows: According to the above technical solution, the pipeline fault diagnosis method based on joint defect detection and size estimation proceeds as follows: S1. Obtain the original magnetic flux leakage signal to measure the health status of the pipeline; S2. Preprocess the original magnetic flux leakage signal to obtain a complete magnetic flux leakage signal; S3. Segment and label the complete magnetic flux leakage signal to obtain the magnetic flux leakage dataset required for training the fault diagnosis network; S4. Construct a shared feature extractor based on the pre-trained residual 50 network and feature pyramid network, input the magnetic flux leakage dataset into the shared feature extractor, and extract the deep shared features of the defects; S5. Construct a private feature segmentation... The fault diagnosis network is divided into several steps: S6, S7, S8, S9, S10, S20, S30, S40, S50, S60, S70, S80, S90, S110, S20, S120, S20, S130, S20, S140, S20, S150, S20, S160, S20, S170, S20, S20, S180, S20, S190, S20, S20, S100, S20, S110, S20, S20, S110, S20, S20, S110, S20, S20, S110, S20, S20, S110, S20, S20, S110, S20, S20, S20, S110, S20, S20, S20, S20, S110, S20, S20, S20, S20, S20, S20, S20, S20, S3 ... Attached Figure Description

[0065] Figure 1 This is a flowchart of a pipeline fault diagnosis method that combines defect detection and size estimation. Detailed Implementation

[0066] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0067] The main concept of this invention is as follows: First, complete training and testing samples are constructed through Gaussian filtering, baseline correction, data segmentation, and data labeling. Then, to obtain features specific to a particular task, a private feature decomposition network is designed to highlight key information for different tasks, thereby enhancing the representation of the specific task and providing a complete feature library foundation for subsequent outputs. Second, addressing the bottleneck of achieving accurate prediction across multiple tasks, a parallel decision sub-network based on physical information is proposed. This parallel decision sub-network embeds knowledge of magnetic flux leakage detection mechanisms to supervise the optimization of the fault diagnosis network. Simultaneously, the constructed parallel decision sub-network prevents negative transfer problems during the optimization process of multi-task models, ensuring the accuracy of defect detection and defect size estimation. This invention's detection method, for the first time, unifies defect detection and defect size estimation into an end-to-end framework, achieving joint optimization of both and improving the accuracy of defect detection and size estimation.

[0068] Based on the above concept, this invention provides a pipeline fault diagnosis method that combines defect detection and size estimation, such as... Figure 1 As shown, this pipeline fault diagnosis method combining defect detection and size estimation includes:

[0069] S1. Obtain the raw magnetic flux leakage signal to measure the health status of the pipeline; preprocess the raw magnetic flux leakage signal to obtain a complete magnetic flux leakage signal;

[0070] The complete magnetic flux leakage signal is segmented and labeled to serve as the magnetic flux leakage dataset required for training the fault diagnosis network. The fault diagnosis network includes a shared feature extractor, a private feature decomposition network, a region candidate network, a region pooling layer network, and a parallel decision sub-network.

[0071] S2. Input the magnetic flux leakage dataset into the shared feature extractor to extract the deep shared features of the defects;

[0072] S3. Input the obtained deep shared features into the private feature decomposition network to generate features for specific tasks, including specific detection features and specific quantization features;

[0073] S4. Input the features of a specific task into the region candidate network to generate defect candidate regions and physical information;

[0074] S5. Input the defect candidate region and the features of the specific task together into the region pooling layer network composed of pooling layer networks to obtain standard features;

[0075] S6. Construct a parallel decision sub-network guided by physical information based on convolutional layers, feed standard features into the parallel decision sub-network, and obtain the prediction output of the parallel decision sub-network.

[0076] S7. By minimizing the difference between physical information and the true value, and the difference between the predicted output of the parallel decision sub-network and the true value, the parameters of the fault diagnosis network are updated to obtain the final fault diagnosis model. Based on the final fault diagnosis model, pipeline defect detection and size estimation are performed to obtain the location of the defect and the length, width and depth of the defect.

[0077] According to the above technical solution, the pipeline fault diagnosis method based on the combined defect detection and size estimation can simultaneously achieve high-precision detection of defect location and quantification of defect size during application, effectively realizing the monitoring of pipeline health status.

[0078] In the pipeline fault diagnosis method with combined defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, in step S1, acquiring the original magnetic flux leakage signal to measure the pipeline condition specifically includes:

[0079] An internal inspection robot based on magnetic flux leakage pipelines acquires raw magnetic flux leakage signals to measure the pipeline's condition.

[0080] In the pipeline fault diagnosis method with combined defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, the original magnetic flux leakage signal is preprocessed, specifically including:

[0081] Step 1.1: Perform Gaussian filtering on the original leakage magnetic field signal M collected from a 100-meter pipeline to obtain the filtered signal B;

[0082] Step 1.2: Based on the differences in sensor baseline values, use formula (1) to unify the dimensions of different sensors:

[0083]

[0084] Among them, X i,j It is the signal of the j-th channel and the i-th sampling point after correction. q and e represent the number of sampling points and the number of Hall sensors, respectively, which are 50000 and 144.

[0085] In the pipeline fault diagnosis method with joint defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, the complete magnetic flux leakage signal is segmented and labeled to obtain the magnetic flux leakage dataset required for training the fault diagnosis network, specifically including:

[0086] Complete the magnetic flux leakage signal by channel, divide it into 500×180 sliding windows, remove invalid samples without defects, and convert it into a pseudo-color image.

[0087] Labels for defect locations and sizes in the pseudo-color image are artificially constructed based on expert experience. The location information of each defect is denoted as Z = [x min ,y min ,x max ,y max ], where x min y min Let x be the coordinate of the top-left corner of the target in the image. max y max Let q be the coordinates of the lower right corner of the target in the image; the size information of each defect is denoted as q = [l, w, d], where l, w, and d are the length, width, and depth of the defect, respectively.

[0088] Following an 8:2 ratio of training to testing, the 496 labeled complete magnetic flux leakage data samples were divided into a training set S and a testing set T.

[0089] In the pipeline fault diagnosis method with combined defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, in step S2, the extraction of deep shared features of defects specifically includes:

[0090] Step S2.1: Construct a shared feature extractor Γ(·) based on the pre-trained residual 50 network and feature pyramid network;

[0091] Step S2.2: Transfer the training samples s i ∈S is fed into the shared feature extractor, and the deep shared feature F of the defect is obtained using formula (2):

[0092] F=Γ(s,μ) (2)

[0093] Where μ is a parameter of the shared feature extractor.

[0094] In the pipeline fault diagnosis method with joint defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, in step S3, the generation of features specific to the task includes:

[0095] Step S3.1: Construct a private feature decomposition network based on a convolutional neural network and a pooling neural network, denoted as the detection private network T(·) and the quantization private network Y(·), respectively. The private feature decomposition network includes a horizontal pooling network, a vertical pooling network, a convolutional network, a normalization layer, and a nonlinear layer. The detection private network and the quantization private network have the same structure but different parameters, and are responsible for outputting specific detection features and specific quantization features, respectively.

[0096] Step S3.2: Input the deep shared features f of the defects into the horizontal pooling network and the vertical pooling network respectively to generate pooling features A. x and A y ;

[0097] Step S3.3: Pool feature A x and A y The intermediate layer features are obtained by sequentially inputting the convolutional network, normalization layer, and nonlinear layer. and

[0098] Step S3.4: Based on intermediate layer features The deep shared features f of the defects are used to obtain the channel private features f using formulas (3) and (4). ch and spatial private features f sp :

[0099]

[0100]

[0101] in, Represents multiplication, ε(·) represents the sigmoid nonlinear function, W0, W1, b0, b1 represent learnable parameters, and Dconv 3×3 Conv represents the parameters of a deformable convolution with a kernel of 3. 1×1 The convolution parameter represents a kernel of 1, and Norm(·) represents a normalized layer;

[0102] Step S3.5: Based on channel private features f ch and spatial private features f sp The adaptive fusion of the two is achieved through formula (5) to obtain specific detection features f. d :

[0103]

[0104] Step S3.6: Feed the deep shared feature f of the defect into the quantization private network Υ(·) and repeat steps S3.1 to S3.5 above to obtain the specific quantization feature f. q .

[0105] In the pipeline fault diagnosis method with joint defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, in step S4, generating defect candidate regions and physical information specifically includes:

[0106] Step S4.1: The region candidate network is composed of two layers of shared convolutional neural network and two parallel fully connected networks connected in series, wherein the number of output nodes of the two parallel fully connected networks are 2 and 4 respectively;

[0107] Step S4.2: Apply specific detection features f d and specific quantitative features f q The data is fed into the regional candidate network, and the defect candidate region Θ and physical information P are obtained using formula (6):

[0108]

[0109] Where W4,b4 represents learnable parameters, and Φ(·) represents the physical information calculation method.

[0110] The acquisition of standard features specifically includes:

[0111] Step S5.1: The region pooling layer network R D (·) and the quantized region pooling layer network R Q (·) consists of pooling layer networks of different lengths, with network output sizes of 7×7×256 and 14×14×256, respectively.

[0112] Step S5.2: Apply specific detection features f d Specific quantitative features f q The candidate defect regions are respectively input into the detection region pooling network R. D (·) and the quantized region pooling layer network R Q (·), the detection standard feature f is obtained using formula (7). dn and quantification standard features f qn :

[0113]

[0114] The acquisition of the network's prediction output specifically includes:

[0115] Step S6.1: The parallel decision subnetwork includes a detection subnetwork and a quantization subnetwork. The detection subnetwork is constructed from two parallel fully connected layer networks, with 2 and 4 output nodes in each fully connected layer, respectively. The quantization subnetwork is constructed from three parallel fully connected layer networks, each with 1 output node.

[0116] Step S6.2: Detect the standard feature f dn Quantitative standard features f qn The physical information P is fed into the detection subnetwork and the quantization subnetwork respectively, and the network's predicted output is obtained using formula (8):

[0117]

[0118] Where c,t represents the location and category predicted by the detection subnetwork, p l ,p w ,p d These represent the length, width, and depth of the defect predicted by the quantized subnetwork, respectively. θ represents the parameters of the detection subnetwork and the quantization subnetwork, respectively.

[0119] In the pipeline fault diagnosis method with joint defect detection and size estimation described in this invention, in order to improve the detection accuracy of pipeline defects and the quantification accuracy of defect size, preferably, in step S7, the parameter update of the fault diagnosis network specifically includes:

[0120] Step S7.1: Based on the physical information and the prediction output of the parallel decision sub-network, calculate the detection loss using formula (9). d :

[0121]

[0122] Where, N cls This indicates the number of anchors used during the training of the candidate network for the region; Represents the classification loss function; ρ represents the normalized weights; N reg Indicates the size of the feature map; This represents the location regression loss function; The label represents the true value, and its value is 0 or 1. It is 1 for a positive sample and 0 for a negative sample; p i t represents the probability that the anchor is predicted to be the target; i Parametric coordinates representing the predicted location; Parametric coordinates representing the actual labeled location; Let p represent the physical constraint loss function, where p i This represents the physical information of the predicted defect candidate region; Labels that indicate physical information about defects.

[0123] Step S7.2: Based on the predicted output of the quantization network, calculate the quantization loss function Loss using formula (10). q :

[0124]

[0125] Step S7.3: Add the detection loss function and the quantization loss function, and update the parameters of the fault diagnosis network by minimizing formula (11) to obtain the fault diagnosis model:

[0126] Loss = αLoss d+(1-α)Loss q (11)

[0127] Here, α is an adjustable factor. Based on the above fault diagnosis model, the test set samples were analyzed. The analysis results show that the defect detection accuracy can reach over 95%, and the average quantization error of length, width, and depth is reduced by 10% compared with existing methods, demonstrating good prospects for practical application.

[0128] The pipeline fault diagnosis method with combined defect detection and size estimation provided by this invention can simultaneously achieve high-precision detection of defect location and quantification of defect size during application, effectively realizing the monitoring of pipeline health status.

[0129] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings; however, the present invention is not limited thereto. Within the scope of the inventive concept, various simple modifications can be made to the technical solutions of the present invention. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately. However, these simple modifications and combinations should also be considered as the content disclosed in the present invention and are all within the protection scope of the present invention.

Claims

1. A pipeline fault diagnosis method combining defect detection and defect size estimation, characterized in that, The pipeline fault diagnosis method includes: S1. Obtain the raw magnetic flux leakage signal to measure the health status of the pipeline; preprocess the raw magnetic flux leakage signal to obtain a complete magnetic flux leakage signal; The complete magnetic flux leakage signal is segmented and labeled to serve as the magnetic flux leakage dataset required for training the fault diagnosis network. The fault diagnosis network includes a shared feature extractor, a private feature decomposition network, a region candidate network, a region pooling layer network, and a parallel decision sub-network. S2. Input the magnetic flux leakage dataset into the shared feature extractor to extract the deep shared features of the defects; S3. Input the obtained deep shared features into the private feature decomposition network to generate features for specific tasks, including specific detection features and specific quantization features; S4. Input the features of a specific task into the region candidate network to generate defect candidate regions and physical information; S5. Input the defect candidate region and the features of the specific task together into the region pooling layer network composed of pooling layer networks to obtain standard features; S6. Construct a parallel decision sub-network guided by physical information based on convolutional layers, feed standard features into the parallel decision sub-network, and obtain the prediction output of the parallel decision sub-network. S7. By minimizing the difference between physical information and the true value, and the difference between the predicted output of the parallel decision sub-network and the true value, the parameters of the fault diagnosis network are updated to obtain the final fault diagnosis model. Based on the final fault diagnosis model, pipeline defect detection and size estimation are performed to obtain the location of the defect and the length, width and depth of the defect. The features for generating specific tasks include specific detection features and specific quantization features, specifically including: Step S3.1: Construct a private feature decomposition network based on convolutional neural networks and pooling neural networks, denoted as the detection private network respectively. And Quantified Private Network The detection private network and the quantization private network have the same structure but different parameters, and are responsible for outputting specific detection features and specific quantization features, respectively; the private feature decomposition network includes a horizontal pooling network, a vertical pooling network, a convolutional network, a normalization layer, and a nonlinear layer; Step S3.2: Input the deep shared features f of the defects into the horizontal pooling network and the vertical pooling network respectively to generate pooling features A. x and A y ; Step S3.3: Pool feature A x and A y The intermediate layer features are obtained by sequentially inputting the convolutional network, normalization layer, and nonlinear layer. and ; Step S3.4: Based on intermediate layer features , The deep shared features f of the defects are used to obtain the channel private features using formulas (3) and (4). and spatial private characteristics : （3） （4） in, Represents multiplication. Represents the sigmoid nonlinear function. Represents learnable parameters. This represents the parameters of a deformable convolution with a kernel of 3. This represents the convolution parameters with a kernel of 1. Represents the normalization layer; Step S3.5: Based on channel private features and spatial private characteristics The adaptive fusion of the two is achieved through formula (5) to obtain specific detection features. : （5） Step S3.6: Feed the deep shared features f of the defects into the quantization private network. Repeat steps S3.1 to S3.5 above to obtain specific quantization features. .

2. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 1, characterized in that, The preprocessing specifically includes: Step 1.1: Perform Gaussian filtering on the original leakage magnetic field signal M to obtain the filtered signal B; Step 1.2: Based on the differences in sensor baseline values, use formula (1) to unify the dimensions of different sensors: （1） in, It is the corrected number The first channel The signal at each sampling point and These represent the number of sampling points and the number of sensors, respectively; m and n are intermediate variables, where m is a number between 1 and q, and n is a number between 1 and e.

3. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 1 or 2, characterized in that, The process of obtaining the magnetic flux leakage dataset is as follows: Complete the leakage magnetic field signal by channel, and then sort the completed samples according to... The sliding window segmentation removes invalid samples without defects and converts them into pseudo-color images; Labels for defect locations and sizes are artificially constructed in the pseudo-color image; the location information of each defect is recorded as follows: ,in This represents the true coordinates of the top-left corner of the target in the pseudo-color image. The coordinates of the lower right corner of the target in the pseudo-color image are the true values; the size information of each defect is denoted as... ,in , , These are the true values ​​for the length, width, and depth of the defect, respectively. The labeled and complete magnetic flux leakage signal data were divided into a training set at an 8:2 ratio. and test set .

4. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 3, characterized in that, The deep shared features of the extraction defects specifically include: Step S2.1: Construct a shared feature extractor based on the concatenation of a pre-trained residual 50 network and a feature pyramid network. ; Step S2.2: Transfer the training set samples The data is fed into the shared feature extractor, and the deep shared feature f of the defect is obtained using formula (2): （2） in, These are the parameters of the shared feature extractor.

5. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 4, characterized in that, The generation of candidate defect regions and physical information specifically includes: Step S4.1: The region candidate network is composed of two layers of shared convolutional neural network and two parallel fully connected networks connected in series, wherein the number of output nodes of the two parallel fully connected networks are 2 and 4 respectively; Step S4.2: Specific detection features and specific quantitative features The data is fed into the regional candidate network, and defect candidate regions are obtained using formula (6). and physical information : （6） in, Represents learnable parameters. This represents the method of calculating physical information.

6. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 5, characterized in that, The acquisition of standard features specifically includes: Step S5.1: The region pooling layer network And quantized region pooling layer network It consists of pooling layer networks of different lengths, and their network output sizes are respectively and ; Step S5.2: Specific detection features Specific quantitative characteristics The candidate defect regions are respectively input into the detection region pooling network. and quantized region pooling layer network The detection standard features are obtained using formula (7). and quantitative standard features : （7）。 7. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 6, characterized in that, The acquisition of the network's prediction output specifically includes: Step S6.1: The parallel decision subnetwork includes a detection subnetwork and a quantization subnetwork. The detection subnetwork is constructed from two parallel fully connected layer networks, with 2 and 4 output nodes in each fully connected layer, respectively. The quantization subnetwork is constructed from three parallel fully connected layer networks, with 1 output node in each of them. Step S6.2: Detect standard features Quantitative Standard Characteristics The physical information P is fed into the detection subnetwork and the quantization subnetwork respectively, and the network's predicted output is obtained using formula (8): （8） in, This represents the location and category predicted by the detection subnetwork. These represent the length, width, and depth of the defect predicted by the quantized subnetwork, respectively. These represent the parameters of the detection subnetwork and the quantization subnetwork, respectively.

8. The pipeline fault diagnosis method based on combined defect detection and defect size estimation according to claim 7, characterized in that, The parameter update of the fault diagnosis network specifically includes: Step S7.1: Based on physical information The prediction output of the parallel decision subnetwork is calculated using formula (9) to determine the detection loss function. : （9） in, , This represents the number of classification anchor boxes and regression anchor boxes used during the training of the candidate network for the region; Represents the classification loss function; Indicates the normalized weights; This represents the location regression loss function; The true label is represented by a value of 0 or 1. It is 1 when the label is a positive sample and 0 when the label is a negative sample. This indicates the probability that the anchor box is predicted to be the target. Parametric coordinates representing the predicted location; Parametric coordinates representing the actual labeled location; Represents the physical constraint loss function, where This represents the physical information of the predicted defect candidate region; Labels indicating physical information about defects; Step S7.2: Based on the predicted output of the quantization subnetwork, calculate the quantization loss function using formula (10). : （10） Step S7.3: Add the detection loss function and the quantization loss function, and update the parameters of the fault diagnosis network by minimizing formula (11) to obtain the final fault diagnosis model: （11） in, It is an adjustable factor.

Images