Intelligent self-evolution pipeline defect detection method with incremental learning ability

Abstract

The invention relates to the technical field of defect detection, in particular to an intelligent self-evolution pipeline defect detection method with incremental learning ability, and the method comprises the steps: obtaining an original magnetic flux leakage signal for measuring the state of a pipeline, and processing the original magnetic flux leakage signal to obtain a clean magnetic flux leakage signal; segmenting and complementing the clean magnetic flux leakage signal to obtain a magnetic flux leakage data set required by network training; constructing a feature extractor based on the network model, and inputting the magnetic flux leakage data set into the feature extractor to extract deep features of the defect; constructing a knowledge focusing module, and sending the obtained deep features into the knowledge focusing module to generate focusing features and joint optimization features; inputting the focusing features into the region candidate network to generate a target candidate region, and inputting the focusing features and the target candidate region into the region pooling layer network to obtain candidate features; and based on the candidate features and the joint optimization features, training a detection model by adopting an alternate optimization strategy, and obtaining a detection model for detecting defect positions. By applying the method, high-precision incremental defect detection can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of pipeline defect detection, and in particular to an intelligent self-evolving pipeline defect detection method with incremental learning capability. Background Art

[0002] As a mainstream transportation medium, pipelines dominate the transportation of oil on land and at sea. However, as time goes by, pipelines will continue to develop defects due to the continuous corrosion of the transported materials and the influence of the external environment. If the defects are not discovered in time, as the degree of damage of the defects continues to increase, the pipeline will leak or even explode, which seriously threatens the stable operation of the pipeline and the safe transportation of oil and gas resources.

[0003] The most effective means of pipeline safety inspection currently recognized at home and abroad is magnetic flux leakage detection. Its main principle is that the Hall element installed on the detector inside the pipeline records the changes in the surrounding magnetic field and obtains the magnetic flux leakage signal, and determines the state of the pipe wall by analyzing the difference in the magnetic flux leakage signal. Among them, pipeline defect detection data analysis is a particularly important part of pipeline magnetic flux leakage detection data analysis, but the existing defect detection methods lack the ability to self-evolve. Specifically, as time goes by, new types of defects will continue to appear in the pipeline. When the existing defect detection method is adjusted to adapt to the detection of new types of defects, its detection performance for old types of defects will drop sharply, and the current detection method also has a low detection rate.

[0004] Therefore, there is an urgent need for an intelligent self-evolving pipeline defect detection method with incremental learning capability to adapt to different categories of pipeline defect detection and improve the accuracy of pipeline defect detection. Summary of the invention

[0005] The purpose of the present invention is to solve the problems that the pipeline defect detection method in the prior art cannot adaptively detect the need for new types of defects and has low detection accuracy, and to provide an intelligent self-evolving pipeline defect detection method with incremental learning capability.

[0006] In order to achieve the above object, the present invention provides an intelligent self-evolving pipeline defect detection method with incremental learning capability, the detection method comprising: S1, obtaining the original magnetic leakage signal for measuring the pipeline status; S2, preprocessing the original magnetic flux leakage signal to obtain a clean magnetic flux leakage signal; S3, segmenting and completing the clean magnetic flux leakage signal to obtain a complete magnetic flux leakage data set required for incremental detection network training; S4. Build the basic feature extractor at time t-1 and time t based on the pre-trained ResNet-50 network model, input the complete magnetic flux leakage data set into the basic feature extractor, and extract the deep features of the defects; S5. Construct the knowledge focusing module at time t-1 and time t, send the obtained deep features to the knowledge focusing module, and generate focusing features and joint optimization features; S6, inputting the focused features into the region candidate network to generate a target candidate region, and then inputting the focused features and the target candidate region into the region pooling layer network to obtain candidate features of uniform size; S7, based on the candidate features of uniform size and the joint optimization features in step S5, an alternating optimization strategy is adopted to complete the training of the detection model at time t, and obtain the detection model at time t; S8. Based on the detection model at time t, the new defects and old defects of the pipeline at time t are detected simultaneously, and the location of the defects is obtained; S9. Repeat the above steps S1-S8 to evolve the detection model at time t until the detection model at time t+N is obtained.

[0007] Preferably, in step S2, the original magnetic flux leakage signal is preprocessed, specifically including: Step 2.1: Perform Gaussian filtering on the original magnetic flux leakage signal D to obtain the filtered signal G; Step 2.2: Based on the differences in the base values ​​of different sensors, use formula (1) to unify the dimensions of different sensors: (1)

[0008] in, is the corrected Channel No. The signal of the sampling point, and They represent the number of sampling points and the number of Hall sensors respectively.

[0009] Preferably, in step S3, the clean magnetic flux leakage signal is segmented and completed to obtain a complete magnetic flux leakage data set required for incremental detection network training, specifically including: Step 3.1: Segment the original magnetic flux leakage signal samples in step S2, and divide the latter into The sensors of the channels are added to the first sensor of the raw data to form The signal segment is converted into a pseudo-color image;

[0010] Step 3.2: Use labelme software to mark the target categories and locations in the pseudo-color image, where the targets include defects and components, and the location information of each target is recorded as ,in is the coordinate of the upper left corner of the target in the image, is the coordinate of the lower right corner of the target in the image;

[0011] Step 3.3: Obtain a complete training set of MFL data with a training to test ratio of 8:2 and test set .

[0012] Preferably, in step S4, extracting the deep features of the defects specifically includes: Step 4.1: Construct two pre-trained ResNet-50 network models to represent the detection model at time t-1 and the detection model at time t, respectively. The detection model at time t is trained, while the detection model at time t-1 remains unchanged. Step 4.2: Complete training samples The image is sent to the ResNet-50 network model pre-trained on ImageNet, and formula (2) is used to obtain the deep features of the defects at time t-1 and time t. , :

[0013] (2)

[0014] in, are the feature extraction parameters at time t and time t+1 respectively.

[0015] Preferably, the front of the feature extractor at time t The parameters of the stage are frozen.

[0016] Preferably, in step S5, generating the focus feature and the joint optimization feature specifically includes: Step 5.1: Convert the deep features from step 4.2 and The data are sent to the knowledge focus module at time t-1 and time t respectively. and ;

[0017] Step 5.2: Based on the knowledge focusing module in step 5.1, obtain the focus feature at time t , using formula (3) to obtain :

[0018] (3)

[0019] in, represents matrix multiplication, represents the learnable parameters, Normalization of representative layer;

[0020] Step 5.3: Based on the deep features at time t-1 and the knowledge focus module at time t-1, the core feature measurement parameter matrix is ​​obtained using formula (4) and the focusing characteristics at time t-1 :

[0021] (4)

[0022] in, represents learnable parameters;

[0023] Step 5.4: Based on the core feature measurement matrix in step 5.3, the core feature sequence of the old class is obtained by formula (5): :

[0024] (5)

[0025] in, is the importance measurement parameter, C is a constant;

[0026] Step 5.5: Based on the features in step 5.2 and step 5.3 and the sequence in step 5.4 , the joint optimization feature is obtained through formula (6) :

[0027] (6).

[0028] Preferably, in step S7, the alternate optimization strategy is used to complete the training of the detection model at time t to obtain the detection model at time t, which specifically includes: Step 7.1: Construct a gradient distortion layer based on the fully connected layer and define the parameters of the gradient distortion layer at time t as , and the parameters of the non-gradient distortion layer are ;

[0029] Step 7.2: Construct a fixed-length feature sequence Used to store candidate features obtained at time t ;

[0030] Step 7.3: Train the non-gradient distortion layer parameters from the training set by minimizing formula (7); (7)

[0031] in, represents the detection loss function of the incremental detection network; Indicates the number of anchors used in the process of training the region candidate network; represents the classification loss function; represents the normalized weight; Indicates the size of the feature map; represents the position regression loss function; Represents the true label, which takes values ​​of 0 and 1. It is 1 when the label is a positive sample and 0 when the label is a negative sample. Indicates the probability that the anchor is predicted as a target; represents the parameterized coordinates of the predicted location; The parameterized coordinates representing the actual annotation location;

[0032] Step 7.4: Non-gradient distortion layer parameter optimization After that, fix the parameters of the non-gradient distortion layer and get the feature sequence The features used to train the gradient distortion layer parameters are obtained, and the gradient distortion layer parameters are updated through the first minimization formula (8):

[0033] (8)

[0034] Step 7.5: Based on joint optimization features and the focal characteristics at time t , the second minimization formula (9) is used to train the detection model at time t, and the detection model at time t is obtained. M t ():

[0035] (9)

[0036] in, Represents an adjustable parameter.

[0037] Preferably, in step 7.3, the parameters of the gradient distortion layer remain unchanged.

[0038] Preferably, in step S9, repeating the above steps S1-S8 to evolve the detection model at time t until the detection model at time t+N is obtained by evolution specifically includes: To obtain the detection model at time t M t () is used as the basis, and the above steps S1-S8 are repeated to obtain the evolved detection model at time t+1 M t+1 (), and then execute steps S1-S8 again until the evolved detection model at time t+N is obtained. M t+N ().

[0039] Preferably, in step S1, obtaining the original magnetic leakage signal for measuring the pipeline state specifically includes: Based on the internal inspection robot of magnetic flux leakage pipeline, the original magnetic flux leakage signal is obtained to measure the pipeline status.

[0040] According to the above technical scheme, based on the intelligent self-evolution pipeline defect detection method with incremental learning capability, through S1, the original magnetic leakage signal for measuring the pipeline state is obtained; S2, the original magnetic leakage signal is preprocessed to obtain a clean magnetic leakage signal; S3, the clean magnetic leakage signal is segmented and completed to obtain a complete magnetic leakage data set required for incremental detection network training; S4, based on the pre-trained ResNet-50 network model, a basic feature extractor at time t-1 and time t is constructed, and the complete magnetic leakage data set is input into the basic feature extractor to extract the deep features of the defect; S5, a knowledge focusing module at time t-1 and time t is constructed, and the obtained deep features are sent to the knowledge focusing module. Focus module, generate focus features and joint optimization features; S6, input the focus features into the regional candidate network to generate the target candidate region, and then input the focus features and the target candidate region into the regional pooling layer network to obtain the candidate features of uniform size; S7, based on the candidate features of uniform size and the joint optimization features in step S5, adopt the alternating optimization strategy to complete the training of the detection model at time t, and obtain the detection model at time t; S8, based on the detection model at time t, realize the simultaneous detection of new and old defects in the pipeline at time t, and obtain the location of the defects; S9, repeat the above steps S1-S8 to evolve the detection model at time t until the detection model at time t+N is evolved. In the application process, the detection of new and old defects in the pipeline can be realized based on the incremental learning ability as time goes by, and the detection accuracy of new and old defects in the pipeline can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of an intelligent self-evolving pipeline defect detection method with incremental learning capability; Figure 2 It is a network structure diagram of the knowledge focus module in the present invention. DETAILED DESCRIPTION

[0042] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0043] The main concept of the present invention is: first, through Gaussian filtering, base value correction, data segmentation and data annotation, complete training samples and test samples are constructed. Then, in view of the problem of excessive compromise of the model with old knowledge in the process of evolution of the detection source model, a knowledge focusing module is designed, and an adaptive knowledge distillation method is proposed based on the knowledge focusing module, which not only realizes the focus of deep features, but also replaces the global feature transfer of the existing method with the transfer of core features, solving the problem of excessive compromise of the evolution model with old knowledge; secondly, in view of the problem that new types of knowledge in the process of evolution of the detection source model are difficult to learn quickly, an alternating optimization strategy based on the gradient distortion layer is proposed, which balances the gradient optimization direction of the evolution model in the new and old classes at the decision layer, and ensures the rapid learning of the new class. The detection method of the present invention can realize high-precision detection of new and old defects at the same time without the need for old class samples. At the same time, the detection method of the present invention has the ability of lifelong learning, and the model can continue to evolve over time.

[0044] Based on the above concept, the present invention provides an intelligent self-evolving pipeline defect detection method with incremental learning capability, such as Figure 1-2 As shown, the intelligent self-evolving pipeline defect detection method with incremental learning capability includes: S1, obtaining the original magnetic leakage signal for measuring the pipeline status; S2, preprocessing the original magnetic flux leakage signal to obtain a clean magnetic flux leakage signal; S3, segmenting and completing the clean magnetic flux leakage signal to obtain a complete magnetic flux leakage data set required for incremental detection network training; S4. Build the basic feature extractor at time t-1 and time t based on the pre-trained ResNet-50 network model, input the complete magnetic flux leakage data set into the basic feature extractor, and extract the deep features of the defects; S5. Construct the knowledge focusing module at time t-1 and time t, send the obtained deep features to the knowledge focusing module, and generate focusing features and joint optimization features; S6, inputting the focused features into the region candidate network to generate a target candidate region, and then inputting the focused features and the target candidate region into the region pooling layer network to obtain candidate features of uniform size; S7, based on the candidate features of uniform size and the joint optimization features in step S5, an alternating optimization strategy is adopted to complete the training of the detection model at time t, and obtain the detection model at time t; S8. Based on the detection model at time t, the new defects and old defects of the pipeline at time t are detected simultaneously, and the location of the defects is obtained; S9. Repeat the above steps S1-S8 to evolve the detection model at time t until the detection model at time t+N is obtained.

[0045] According to the above technical solution, based on the intelligent self-evolutionary pipeline defect detection method with incremental learning capability, during the application process, as time goes by, the detection of new and old pipeline defects can be realized based on the incremental learning capability, and the detection accuracy of new and old pipeline defects can be effectively improved.

[0046] In the intelligent self-evolution pipeline defect detection method with incremental learning capability described in the present invention, in order to improve the detection accuracy of pipeline defects, preferably, in step S1, the obtaining of the original magnetic leakage signal for measuring the pipeline state specifically includes: Based on the internal inspection robot of magnetic flux leakage pipeline, the original magnetic flux leakage signal is obtained to measure the pipeline status.

[0047] In the intelligent self-evolutionary pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S2, the original magnetic leakage signal is preprocessed, specifically including: Step 2.1: Perform Gaussian filtering on the original magnetic flux leakage signal D to obtain the filtered signal G; Step 2.2: Based on the differences in the base values ​​of different sensors, use formula (1) to unify the dimensions of different sensors: (1)

[0048] in, is the corrected Channel No. The signal of the sampling point, and They represent the number of sampling points and the number of Hall sensors respectively.

[0049] In the intelligent self-evolving pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S3, the clean magnetic leakage signal is segmented and completed to obtain a complete magnetic leakage data set required for incremental detection network training, which specifically includes: Step 3.1: Segment the original magnetic flux leakage signal samples in step S2, and divide the latter into The sensors of the channels are added to the first sensor of the raw data to form The signal segment is converted into a pseudo-color image;

[0050] Step 3.2: Use labelme software to mark the target categories and locations in the pseudo-color image, where the targets include defects and components, and the location information of each target is recorded as ,in is the coordinate of the upper left corner of the target in the image, is the coordinate of the lower right corner of the target in the image;

[0051] Step 3.3: Obtain a complete training set of MFL data with a training to test ratio of 8:2 and test set .

[0052] In the intelligent self-evolutionary pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S4, the deep features of the defects are extracted, specifically including: Step 4.1: Construct two pre-trained ResNet-50 network models to represent the detection model at time t-1 and the detection model at time t, respectively. The detection model at time t is trained, while the detection model at time t-1 remains unchanged. Step 4.2: Complete training samples The image is sent to the ResNet-50 network model pre-trained on ImageNet, and formula (2) is used to obtain the deep features of the defects at time t-1 and time t. , :

[0053] (2)

[0054] in, are the feature extraction parameters at time t and time t+1 respectively.

[0055] In order to prevent overfitting of the network, in a further preferred embodiment, the front of the feature extractor at time t The parameters of the stage are frozen.

[0056] In the intelligent self-evolutionary pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S5, the generating of focus features and joint optimization features specifically includes: Step 5.1: Convert the deep features from step 4.2 and The data are sent to the knowledge focus module at time t-1 and time t respectively. and ;

[0057] Step 5.2: Based on the knowledge focusing module in step 5.1, obtain the focus feature at time t , using formula (3) to obtain :

[0058] (3)

[0059] in, represents matrix multiplication, represents the learnable parameters, Normalization of representative layer;

[0060] Step 5.3: Based on the deep features at time t-1 and the knowledge focus module at time t-1, the core feature measurement parameter matrix is ​​obtained using formula (4) and the focusing characteristics at time t-1 :

[0061] (4)

[0062] in, represents learnable parameters;

[0063] Step 5.4: Based on the core feature measurement matrix in step 5.3, the core feature sequence of the old class is obtained by formula (5): :

[0064] (5)

[0065] in, is the importance measurement parameter, C is a constant;

[0066] Step 5.5: Based on the features in step 5.2 and step 5.3 and the sequence in step 5.4 , the joint optimization feature is obtained through formula (6) :

[0067] (6).

[0068] In the intelligent self-evolutionary pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S7, the alternate optimization strategy is used to complete the training of the detection model at time t to obtain the detection model at time t, which specifically includes: Step 7.1: Construct a gradient distortion layer based on the fully connected layer and define the parameters of the gradient distortion layer at time t as , and the parameters of the non-gradient distortion layer are ;

[0069] Step 7.2: Construct a fixed-length feature sequence Used to store candidate features obtained at time t ;

[0070] Step 7.3: Train the non-gradient distortion layer parameters from the training set by minimizing formula (7); (7)

[0071] in, represents the detection loss function of the incremental detection network; Indicates the number of anchors used in the process of training the region candidate network; represents the classification loss function; represents the normalized weight; Indicates the size of the feature map; represents the position regression loss function; Represents the true label, which takes values ​​of 0 and 1. It is 1 when the label is a positive sample and 0 when the label is a negative sample. Indicates the probability that the anchor is predicted as a target; represents the parameterized coordinates of the predicted location; The parameterized coordinates representing the actual annotation location;

[0072] Step 7.4: Non-gradient distortion layer parameter optimization After that, fix the parameters of the non-gradient distortion layer and get the feature sequence The features used to train the gradient distortion layer parameters are obtained, and the gradient distortion layer parameters are updated through the first minimization formula (8):

[0073] (8)

[0074] Step 7.5: Based on joint optimization features and the focal characteristics at time t , the second minimization formula (9) is used to train the detection model at time t, and the detection model at time t is obtained. M t ():

[0075] (9)

[0076] in, Represents an adjustable parameter.

[0077] In a further preferred embodiment, in step 7.3, the parameters of the gradient distortion layer remain unchanged.

[0078] In the intelligent self-evolution pipeline defect detection method with incremental learning capability described in the present invention, in order to further improve the detection accuracy of pipeline defects, preferably, in step S9, the above steps S1-S8 are repeated to evolve the detection model at time t until the detection model at time t+N is obtained, which specifically includes: To obtain the detection model at time t M t() is used as the basis, and the above steps S1-S8 are repeated to obtain the evolved detection model at time t+1 M t+1 (), and then execute steps S1-S8 again until the evolved detection model at time t+N is obtained. M t+N ().

[0079] The intelligent self-evolving pipeline defect detection method with incremental learning capability provided by the present invention can, during application, detect new and old pipeline defects based on the incremental learning capability over time, and effectively improve the detection accuracy of new and old pipeline defects.

[0080] The preferred embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations. However, these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention.

Claims

1. An intelligent self-evolving pipeline defect detection method with incremental learning capability, characterized in that: The test method includes: S1, obtaining the original magnetic leakage signal for measuring the pipeline status; S2, preprocessing the original magnetic flux leakage signal to obtain a clean magnetic flux leakage signal; S3, segmenting and completing the clean magnetic flux leakage signal to obtain a complete magnetic flux leakage data set required for incremental detection network training; S4. Build the basic feature extractor at time t-1 and time t based on the pre-trained ResNet-50 network model, input the complete magnetic flux leakage data set into the basic feature extractor, and extract the deep features of the defects; S5. Construct the knowledge focusing module at time t-1 and time t, send the obtained deep features to the knowledge focusing module, and generate focusing features and joint optimization features; S6, inputting the focused features into the region candidate network to generate a target candidate region, and then inputting the focused features and the target candidate region into the region pooling layer network to obtain candidate features of uniform size; S7, based on the candidate features of uniform size and the joint optimization features in step S5, an alternating optimization strategy is adopted to complete the training of the detection model at time t, and obtain the detection model at time t; S8. Based on the detection model at time t, the new defects and old defects of the pipeline at time t are detected simultaneously, and the location of the defects is obtained; S9. Repeat the above steps S1-S8 to evolve the detection model at time t until the detection model at time t+N is obtained.

2. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 1 is characterized in that: In step S2, the original magnetic flux leakage signal is preprocessed, specifically including: Step 2.1: Perform Gaussian filtering on the original magnetic flux leakage signal D to obtain the filtered signal G; Step 2.2: Based on the differences in the base values ​​of different sensors, use formula (1) to unify the dimensions of different sensors: （1） in, is the corrected Channel No. The signal of the sampling point, and They represent the number of sampling points and the number of Hall sensors respectively.

3. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 1 or 2, characterized in that: In step S3, the clean magnetic flux leakage signal is segmented and completed to obtain a complete magnetic flux leakage data set required for incremental detection network training, which specifically includes: Step 3.1: Segment the original magnetic flux leakage signal samples in step S2, and divide the latter into The sensors of the channels are added to the first sensor of the raw data to form The signal segment is converted into a pseudo-color image; Step 3.2: Use labelme software to mark the target categories and locations in the pseudo-color image, where the targets include defects and components, and the location information of each target is recorded as ,in is the coordinate of the upper left corner of the target in the image, is the coordinate of the lower right corner of the target in the image; Step 3.3: Obtain a complete training set of MFL data with a training to test ratio of 8:2 and test set .

4. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 3 is characterized in that: In step S4, the deep features of the defects are extracted, specifically including: Step 4.1: Construct two pre-trained ResNet-50 network models to represent the detection model at time t-1 and the detection model at time t, respectively. The detection model at time t is trained, while the detection model at time t-1 remains unchanged. Step 4.2: Complete training samples The image is sent to the ResNet-50 network model pre-trained on ImageNet, and formula (2) is used to obtain the deep features of the defects at time t-1 and time t. , : （2） in, are the feature extraction parameters at time t and time t+1 respectively.

5. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 4 is characterized in that: The front of the feature extractor at time t The parameters of the stage are frozen.

6. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 4 or 5, characterized in that: In step S5, the generating of focusing features and joint optimization features specifically includes: Step 5.1: Convert the deep features from step 4.2 and The data are sent to the knowledge focus module at time t-1 and time t respectively. and ; Step 5.2: Based on the knowledge focusing module in step 5.1, obtain the focus feature at time t , using formula (3) to obtain : （3） in, represents matrix multiplication, represents the learnable parameters, Normalization of representative layer; Step 5.3: Based on the deep features at time t-1 and the knowledge focus module at time t-1, the core feature measurement parameter matrix is ​​obtained using formula (4) and the focusing characteristics at time t-1 : （4） in, represents learnable parameters; Step 5.4: Based on the core feature measurement matrix in step 5.3, the core feature sequence of the old class is obtained by formula (5): : （5） in, is the importance measurement parameter, C is a constant; Step 5.5: Based on the features in step 5.2 and step 5.3 and the sequence in step 5.4 , the joint optimization feature is obtained through formula (6) : （6）。 7. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 1 or 6, characterized in that: In step S7, the alternating optimization strategy is used to complete the training of the detection model at time t to obtain the detection model at time t, which specifically includes: Step 7.1: Construct a gradient distortion layer based on the fully connected layer and define the parameters of the gradient distortion layer at time t as , and the parameters of the non-gradient distortion layer are ; Step 7.2: Construct a fixed-length feature sequence Used to store candidate features obtained at time t ; Step 7.3: Train the non-gradient distortion layer parameters from the training set by minimizing formula (7); （7） in, represents the detection loss function of the incremental detection network; Indicates the number of anchors used in the process of training the region candidate network; represents the classification loss function; represents the normalized weight; Indicates the size of the feature map; represents the position regression loss function; Represents the true label, which takes values ​​of 0 and 1. It is 1 when the label is a positive sample and 0 when the label is a negative sample. Indicates the probability that the anchor is predicted as a target; represents the parameterized coordinates of the predicted location; The parameterized coordinates representing the actual annotation location; Step 7.4: Non-gradient distortion layer parameter optimization After that, fix the parameters of the non-gradient distortion layer and get the feature sequence The features used to train the gradient distortion layer parameters are obtained, and the gradient distortion layer parameters are updated through the first minimization formula (8): （8） Step 7.5: Based on joint optimization features and the focusing characteristics at time t , the second minimization formula (9) is used to train the detection model at time t, and the detection model at time t is obtained. M t (): （9） in, Represents an adjustable parameter.

8. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 7 is characterized in that: In step 7.3, the parameters of the gradient distortion layer remain unchanged.

9. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 7 is characterized in that: In step S9, the above steps S1-S8 are repeated to evolve the detection model at time t until the detection model at time t+N is obtained, which specifically includes: To obtain the detection model at time t M t () is used as the basis, and the above steps S1-S8 are repeated to obtain the evolved detection model at time t+1 M t+1 (), and then execute steps S1-S8 again until the evolved detection model at time t+N is obtained. M t+N ().

10. The intelligent self-evolving pipeline defect detection method with incremental learning capability according to claim 1, characterized in that: In step S1, the acquisition of the original magnetic leakage signal for measuring the pipeline status specifically includes: Based on the internal inspection robot of magnetic flux leakage pipeline, the original magnetic flux leakage signal is obtained to measure the pipeline status.

Images