A quantitative assessment method for axial and bending stress in pipelines based on alternating electromagnetic fields.

Abstract

This invention belongs to the field of nondestructive testing technology for pipeline stress, and particularly relates to a quantitative assessment method for axial and bending stress in pipelines based on alternating current electromagnetic fields. This quantitative assessment method improves the detection sensitivity of axial and bending stresses in pipelines, effectively reduces the risk of missed detections due to the inability to separate axial and bending stresses from mixed stress signals, and achieves simultaneous identification of both types of stress. The quantitative assessment method includes: arranging several alternating current electromagnetic field array detection probes in a double-row, equally spaced configuration; constructing a fitting relationship between the detection signals of the alternating current electromagnetic field array detection probes and the stress; obtaining two sets of magnetic field signal matrices under excitation in different directions; filtering and denoising the magnetic field signal matrices; obtaining the actual pipeline axial stress matrix and the actual pipeline bending stress matrix; establishing a bi-branch stress analysis network; and completing the localization and quantitative identification of the axial and bending stresses in the pipeline to be quantitatively assessed.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for pipeline stress, and particularly relates to a quantitative assessment method for axial and bending stress of pipelines based on alternating electromagnetic fields. Background Technology

[0002] For oil and gas pipelines, during their service life, they are simultaneously subjected to the coupled effects of axial stress (such as the axial component caused by internal pressure and thermal expansion and contraction) and bending stress (such as topographic settlement and external mechanical loads). When these two types of stress are superimposed, fatigue cracks or even failures are easily triggered at the geometric discontinuities of the pipeline or in the soil slip zone. Therefore, technicians need to accurately identify the axial stress and bending stress of the pipeline simultaneously.

[0003] Alternating current electromagnetic field (ACSM) measurement technology, with its advantages of non-contact detection, high quantification accuracy, and strong resistance to lift-off interference, is widely used in the detection of stress in various structures. However, research has revealed that the probe arrangement of existing ACSM detection technologies exhibits a significant direction dependence: for axial stress, high detection sensitivity is achieved only when the excitation magnetic field direction is perpendicular to the stress direction; while for bending stress, an effective response is only obtained when the excitation direction is parallel to the stress direction. Therefore, using excitation in only a single direction will lead to a significant weakening of the response signal for other types of stress, and may even result in missed detections.

[0004] Furthermore, it is worth noting that the stress quantification models of existing ACSM measurement technologies mostly rely on single pipe materials and ideal working conditions under laboratory conditions, making it impossible to directly apply their results to actual construction sites; and the lack of support from real pipeline field test data makes it difficult to achieve effective transformation for engineering applications.

[0005] Therefore, in order to overcome the shortcomings of the existing technology, it is urgent for those skilled in the art to provide a new quantitative assessment method for pipeline axial stress and bending stress, which can retain the original advantages of ACSM detection technology while taking into account the need for high-sensitivity synchronous detection of pipeline axial stress and bending stress, and help to achieve accurate assessment of the safety status of pipeline structure. Summary of the Invention

[0006] This invention provides a quantitative assessment method for axial and bending stress in pipelines based on alternating electromagnetic fields. By employing a composite excitation design for the arrangement of AC electromagnetic field array detection probes, the detection sensitivity for axial and bending stress in pipelines is improved, effectively reducing the risk of missed detection due to the inability to separate axial and bending stresses from mixed stress signals, and achieving simultaneous identification of the two types of stresses. Furthermore, through force calibration and data verification, the detection reliability of this quantitative assessment method is ensured.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A quantitative assessment method for axial and bending stress in pipelines based on alternating electromagnetic fields includes the following steps: Step S1: Define the axial direction of the pipeline to be quantitatively evaluated as the X direction, the circumferential direction of the pipeline to be quantitatively evaluated as the Y direction, and the radial direction of the pipeline to be quantitatively evaluated as the Z direction. A double-row arrangement is used, with several AC electromagnetic field array detection probes evenly spaced along the Y direction at the location on the pipe surface to be quantitatively evaluated. The first row of AC electromagnetic field array detection probes is used to detect the axial stress of the pipe, with the internal excitation coil and the magnetic sensor both positioned in the Y direction. The second row of AC electromagnetic field array detection probes is used to detect the bending stress of the pipe, with the internal excitation coil and the magnetic sensor both positioned in the X direction. Step S2: Perform stress calibration on the AC electromagnetic field array detection probe to establish a fitting relationship between the detection signal of the AC electromagnetic field array detection probe and the stress. Step S3: Using an AC electromagnetic field array detection probe, scan the inside of the pipe to be quantitatively evaluated along the X-axis to obtain two sets of magnetic field signal matrices under excitation in different directions, as follows: , ; Step S4: For B X1 B X2 Each data point in the magnetic field signal matrix is ​​filtered and denoised to obtain the filtered and denoised output matrix. , ; Step S5: Calculate the sensitivity S of each value in each column of the signal in the output matrices B1 and B2 respectively, to obtain the axial stress sensitivity matrix. and bending stress sensitivity matrix Among them, the sensitivity S satisfies, ; Based on the fitting relationship between the AC electromagnetic field array detection probe detection signal and stress obtained in step S2, the axial stress sensitivity matrix S X1 and bending stress sensitivity matrix S X2 Convert to the corresponding stress matrix; After eliminating the stress components caused by the internal pressure of the pipe in the stress matrix, the actual axial stress matrix of the pipe is obtained. and the actual pipe bending stress matrix ; Step S6: Establish a dual-branch stress analysis network based on the fusion of CNN-Transformer feature extraction and Stacking; Specifically, the upper branch of the dual-branch stress analysis network is a bending stress analysis module based on CNN-Transformer feature extraction, which consists of a multi-scale convolutional feature extraction module, a multi-expansion rate residual convolution module, and a Transformer encoder module; its input is the actual pipe bending stress matrix. The output includes the distribution and location of bending stress features, along with mileage information. The lower branch of the bi-branch stress analysis network is an axial stress analysis module based on Stacking integration of XGBoost and Random Forest, which consists of the base model XGBoost and the base model Random Forest. Its input is the bending feature location and various feature signals of the pipeline to be quantitatively evaluated output by the lower branch of the bi-branch stress analysis network, and the output is the axial stress regression prediction value. Step S7: Based on the bending stress feature distribution, bending feature location, and axial stress regression prediction value output by the bi-branch stress analysis network trained in Step S6, complete the location and quantitative identification of the axial stress and bending stress of the pipeline to be quantitatively evaluated.

[0008] Preferably, step S2 specifically includes the following steps: Step S21: Perform tensile stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for tensile stress calibration; place the AC electromagnetic field array detection probe in the middle of the standard test block for tensile stress calibration, and make the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor perpendicular to the tensile direction. Tensile stress is gradually applied starting from 0 MPa with a fixed stress increment; and the stress is paused at a fixed time interval after each tensile stress increment stabilizes. The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the tensile stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the tensile stress are calibrated at fixed intervals after each tensile stress increase stabilizes. Step S22: Perform bending stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for bending stress calibration; place the AC electromagnetic field array detection probe on the upper or lower surface of the standard test block between the three bending pressure application point and the support point, so that the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor are parallel to the surface stress direction of the standard test block during the bending process. The bending stress is gradually applied starting from 0 MPa with a fixed stress increase; and a fixed time interval is set after each bending stress increase stabilizes. The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the bending stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the bending stress are calibrated at fixed intervals after each bending stress increase stabilizes.

[0009] Preferably, in step S4, B... X1 B X2 The process of filtering and denoising each data point in the magnetic field signal matrix is ​​specifically described as follows: Set the width of the movable window to 2a+1; where a is a positive integer; In the moving window, for B X1 B X2 For each data point in the magnetic field signal matrix, select 2m+1 data points within a window centered on it to form a subsequence {xi−m,...,xi,...,xi+m}; where m is a positive integer and satisfies m>a; Perform k-th order polynomial least squares fitting on the subsequence to obtain the fitting polynomial. ;in, ; These are the relative position coordinates within the window; Calculate polynomials The function value y at the center of the window is used as the filtered output value of the data point; Slide the window to the next data point and repeat the above process until all data has been traversed, generating the filtered and denoised output matrix. , .

[0010] A preferred approach is to independently train the branches of the bi-branch stress analytical network in step S6, specifically as follows: The actual pipe bending stress matrix is ​​extracted by a multi-scale convolution feature extraction module and a multi-dilation rate residual convolution module. Local features related to mileage information; Long-range dependencies are captured by the Transformer encoder module; The bending stress feature distribution and bending feature location are output through spatial attention pooling, and then optimized using a composite loss function that combines classification and regression.

[0011] A preferred approach is to describe the independent training process of the branches in the bi-branch stress analysis network in step S6 as follows: After fusing the bending feature positions of the branch outputs on the bi-branch stress analysis network with the feature signals of the pipeline to be quantitatively evaluated, the results are input in parallel into two base learners: XGBoost and Random Forest. XGBoost performs fine feature selection and regularized training through gradient boosting trees, and random forest obtains robust predictions through bagging ensemble. The outputs of the two base learners, XGBoost and Random Forest, are fused using a linear regression meta-model through stacking to obtain the axial stress regression prediction.

[0012] Preferably, before performing step S7, the following steps are also included: Step S60: Through bi-branch loss weighting and multi-task learning, the feature representations of the bending stress analysis module and the axial stress analysis module are mutually coordinated and enhanced; The hyperparameters were optimized using early stopping and cross-validation.

[0013] This invention provides a method for quantitatively assessing the axial and bending stresses of a pipeline based on an alternating current electromagnetic field. The method includes at least the following steps: Step S1: Arranging several AC electromagnetic field array detection probes at equal intervals along the Y-direction at locations on the surface of the pipeline to be quantitatively assessed using a double-row arrangement; Step S2: Performing stress calibration on the AC electromagnetic field array detection probes to establish a fitting relationship between the detection signals and stresses; Step S3: Using the AC electromagnetic field array detection probes, scanning the interior of the pipeline to be quantitatively assessed along the X-direction to obtain two sets of magnetic field signal matrices under excitation in different directions; Step S4: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] X1 B X2 Each data point in the magnetic field signal matrix undergoes filtering and noise reduction processing; Step S5: Calculate the sensitivity S of each value in each column of the output matrices B1 and B2 respectively to obtain the axial stress sensitivity matrix and the bending stress sensitivity matrix; After eliminating the stress components caused by the internal pressure of the pipe in the stress matrix, obtain the actual pipe axial stress matrix and the actual pipe bending stress matrix. Step S6: Establish a dual-branch stress analysis network based on CNN-Transformer feature extraction and Stacking fusion; Step S7: Based on the bending stress feature distribution, bending feature location and axial stress regression prediction value output by the dual-branch stress analysis network trained in Step S6, complete the localization and quantitative identification of the axial stress and bending stress of the pipeline to be quantitatively evaluated.

[0014] The quantitative assessment method for axial and bending stress of pipelines based on alternating electromagnetic fields, which has the above-described steps, has at least the following technical advantages compared to existing technologies: (1) The present invention provides a quantitative assessment method for axial stress and bending stress of pipeline based on AC electromagnetic field. Through the layout design of AC electromagnetic field array detection probe with double row orthogonal excitation, the independent synchronous detection of axial stress and bending stress of pipeline is realized. Furthermore, by constructing a training dual-branch stress analysis network, the complementary enhancement of bending stress analysis and axial stress analysis is realized. (2) The present invention provides a quantitative assessment method for axial stress and bending stress of pipelines based on AC electromagnetic field. By calibrating the fitting relationship between the detection signal and stress of the AC electromagnetic field array detection probe, and then combining the physical decoupling and fusion analysis of stress-related data, it effectively solves a series of technical problems that are difficult to quantify and analyze in the existing single excitation detection, such as inaccurate bending stress identification, large direction interpretation error and large amount of feature data. Finally, it provides highly reliable data support and intelligent analysis means for pipeline integrity management and safety assessment. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the following drawings: Figure 1 A schematic flowchart illustrating the quantitative evaluation method for axial and bending stress of pipelines based on alternating electromagnetic fields provided by this invention. Figure 2a Results of axial stress response under different AC electromagnetic field array probe angles; Figure 2b Results of bending stress response under different AC electromagnetic field array probe angles; Figure 3 A schematic diagram showing the arrangement of the AC electromagnetic field array detection probes; Figure 4a The graph shows the fitting relationship between the detection signal from the AC electromagnetic field array detection probe and the axial stress. Figure 4b The graph shows the fitting relationship between the detection signal from the AC electromagnetic field array detection probe and the bending stress. Figure 5a This is a schematic diagram before the axial stress signal is filtered. Figure 5b This is a schematic diagram of the filtered axial stress signal. Figure 6a This is a schematic diagram of the bending stress signal before filtering. Figure 6b This is a schematic diagram of the filtered bending stress signal. Figure 7a This is a schematic diagram of the network training results without subtracting axial stress. Figure 7bA schematic diagram of the network training results after subtracting axial stress; Figure 8 This is a schematic diagram of the algorithm framework for a bi-branch stress analysis network. Figure 9 This is a schematic diagram of the sample feature vectors used to train a bi-branch stress analytical network. Figure 10 This is a schematic diagram illustrating the antisymmetric variation characteristics along longitude and latitude. Figure 11 An example of axial stress quantification features; Figure 12 A schematic diagram illustrating the importance of axial stress quantification characteristics; Figure 13 This is a diagram of the characteristic signals of bending stress. Figure 14 A visualization of the characteristic segment of bending stress; Figure 15 This is a summary chart of the histogram and quantification information of the characteristic distribution of bending stress. Detailed Implementation

[0016] This invention provides a quantitative assessment method for axial and bending stress in pipelines based on alternating electromagnetic fields. By employing a composite excitation design for the arrangement of AC electromagnetic field array detection probes, the detection sensitivity for axial and bending stress in pipelines is improved, effectively reducing the risk of missed detection due to the inability to separate axial and bending stresses from mixed stress signals, and achieving simultaneous identification of the two types of stresses. Furthermore, through force calibration and data verification, the detection reliability of this quantitative assessment method is ensured.

[0017] like Figure 1 As shown, the present invention provides a quantitative evaluation method for axial and bending stress of a pipeline based on an alternating electromagnetic field, comprising the following steps: Step S1: Define the axial direction of the pipeline to be quantitatively evaluated as the X direction, the circumferential direction of the pipeline to be quantitatively evaluated as the Y direction, and the radial direction of the pipeline to be quantitatively evaluated as the Z direction.

[0018] A dual-row arrangement is used, with several AC electromagnetic field array probes evenly spaced along the Y-axis at the locations on the pipe surface to be quantitatively evaluated. The first row of probes is used to detect the axial stress of the pipe, with its internal excitation coils and magnetic sensors both oriented in the Y-axis. The second row of probes is used to detect the bending stress of the pipe, with its internal excitation coils and magnetic sensors both oriented in the X-axis.

[0019] For example, assuming eight AC electromagnetic field array detection probes are uniformly arranged along the Y direction, the arrangement of these probes can be referenced as follows: Figure 3 As shown in the figure. In the AC electromagnetic field array detection probe, the orientation of the internal excitation coils in the first row and the sensing direction of the magnetic sensor are both in the Y direction; in the AC electromagnetic field array detection probe, the orientation of the internal excitation coils in the second row and the sensing direction of the magnetic sensor are both in the X direction.

[0020] Further as Figure 2a , Figure 2b As shown (where, Figure 2a The results show the response to axial stress under different AC electromagnetic field array probe angles. Figure 2b (The results show the response to bending stress under different AC electromagnetic field array detector probe angles.) Based on the calibration results of axial stress and bending stress, it can be found that for axial stress in the pipeline, the signal sensitivity is high when the excitation coil core is perpendicular to the stress; while for bending stress, the signal sensitivity is even higher when the excitation coil core is parallel to the stress. Therefore, the above-mentioned dual-row orthogonal excitation AC electromagnetic field array detector probe layout design can achieve optimal sensitivity response for different stress types (i.e., the first row of AC electromagnetic field array detector probes adopts an excitation layout perpendicular to the pipeline axis, specifically for high-sensitivity capture of axial stress; the second row of AC electromagnetic field array detector probes adopts an excitation layout parallel to the pipeline axis to achieve optimal detection of bending stress), thus providing a key foundation for subsequent steps to effectively separate and independently extract the axial stress and bending stress components in the electromagnetic response characteristics.

[0021] Step S2: Perform stress calibration on the AC electromagnetic field array detection probe to establish a fitting relationship between the detection signal of the AC electromagnetic field array detection probe and the stress.

[0022] Based on completing step S1, proceed to step S2.

[0023] Specifically, step S2 is used (preferably under laboratory conditions) to perform stress calibration on the AC electromagnetic field array detection probe, thereby establishing a fitting relationship between the detection signal of the AC electromagnetic field array detection probe and the stress.

[0024] In a preferred embodiment of the present invention, step S2 specifically includes the following steps: Step S21: Perform tensile stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for tensile stress calibration; place the AC electromagnetic field array detection probe in the middle of the standard test block for tensile stress calibration, and ensure that the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor are both perpendicular to the tensile direction.

[0025] Tensile stress is gradually applied starting from 0 MPa with a fixed stress increment; and a fixed time interval is set after each tensile stress increment stabilizes.

[0026] The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the tensile stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the tensile stress are calibrated at fixed intervals after each tensile stress increase stabilizes.

[0027] Step S22: Perform bending stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for bending stress calibration; place the AC electromagnetic field array detection probe on the upper or lower surface of the standard test block between the three bending pressure application point and the support point, so that the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor are parallel to the surface stress direction of the standard test block during the bending process.

[0028] The bending stress is gradually applied starting from 0 MPa with a fixed stress increase; and a fixed time interval is set after each bending stress increase stabilizes.

[0029] The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the bending stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the bending stress are calibrated at fixed intervals after each bending stress increase stabilizes.

[0030] To facilitate understanding by those skilled in the art, the following specific implementation method is provided, specifically using a universal testing machine to complete the above-mentioned tensile stress calibration experiment and three-point bending stress calibration experiment. The tensile stress loading process is performed at 50MPa intervals, gradually increasing from 0MPa to 300MPa; with a 1-minute pause at each interval. Simultaneously, the magnetic signals detected by the AC electromagnetic field array detection probes are acquired in real time during the tensile stress loading process. The detection signals acquired at each interval are then calibrated against the tensile stress. Similarly, the bending stress loading process is performed at 50MPa intervals, gradually increasing from 0MPa to 300MPa; with a 1-minute pause at each interval. Simultaneously, the magnetic signals detected by two AC electromagnetic field array detection probes located on the upper or lower surface of the three-point bending pressure application point and support point of the bending stress calibration standard test block are acquired during the bending stress loading process. The rate of change of the detection signals from the two AC electromagnetic field array detection probes at each interval is then calibrated against the stress. The fitting relationship can be referenced as follows: Figure 4a , Figure 4b As shown (where, Figure 4a The graph shows the fitting relationship between the detection signal from the AC electromagnetic field array detection probe and the axial stress. Figure 4b(Figure showing the fitting relationship between the detection signal from the AC electromagnetic field array detection probe and the bending stress).

[0031] After further fitting and solving, the fitting relationship between the AC electromagnetic field array detection probe's detected signal and tensile stress can be described as follows: Regarding the fitting relationship between the signal detected by the AC electromagnetic field array detection probe and the bending stress, it can be described as follows: Therefore, step S2, through a controlled experiment with a load-mode distribution, establishes a precise and traceable quantitative benchmark for the fitting relationship between the AC electromagnetic field array detection probe signal and stress for this quantitative evaluation method.

[0032] Step S3: Using an AC electromagnetic field array detection probe, scan the inside of the pipe to be quantitatively evaluated along the X-axis to obtain two sets of magnetic field signal matrices under excitation in different directions, as follows: , .

[0033] Step S4: For B X1 B X2 Each data point in the magnetic field signal matrix is ​​filtered and denoised to obtain the filtered and denoised output matrix. , .

[0034] Based on step S2, steps S3 and S4 are further implemented. It is worth noting that steps S3 and S4 are specifically used to obtain two sets of magnetic field signal matrices under excitation in different directions, and to perform filtering and noise reduction processing on the magnetic field signal matrices. In a preferred embodiment of the present invention, step S4 involves B... X1 B X2 The process of filtering and denoising each data point in the magnetic field signal matrix is ​​specifically described as follows: Set the width of the movable window to 2a+1, where a is a positive integer.

[0035] In the moving window, for B X1 B X2 For each data point in the magnetic field signal matrix, select 2m+1 data points within a window centered on it to form a subsequence {xi−m,...,xi,...,xi+m}. Here, m is a positive integer and satisfies m>a.

[0036] Perform k-th order polynomial least squares fitting on the subsequence to obtain the fitting polynomial. .in, ; These are the relative position coordinates within the window.

[0037] Calculate polynomials The function value y at the center of the window is used as the filtered output value of the data point.

[0038] Slide the window to the next data point and repeat the above process until all data has been traversed, generating the filtered and denoised output matrix. , .

[0039] For example, if the width of the moving window is set to 100, then for each data point in the moving window, a subsequence is formed by selecting 100 data points within the window centered on that data point, resulting in {xi−50,...,xi,...,xi+50}. Further, a 4-polynomial least squares fit is performed on the above subsequence to obtain the fitted polynomial as follows: Where t∈[-50,50] are the relative position coordinates within the window. Based on this, the function value y of the polynomial P(t) at the center position of the window (t=0) is calculated as the filtered output value of the data point. The window is slid to the next data point, and the above process is repeated until all data is traversed, generating a smoothed output matrix after filtering and noise reduction, as follows: , .

[0040] It is worth noting that by comparing the output matrices , Schematic diagram of the signal before and after filtering (for reference) Figure 5a , Figure 5b as well as Figure 6a , Figure 6b As shown; where, Figure 5a This is a schematic diagram before the axial stress signal is filtered. Figure 5b This is a schematic diagram of the filtered axial stress signal. Figure 6a This is a schematic diagram of the bending stress signal before filtering. Figure 6b (This is a schematic diagram of the filtered bending stress signal). It can be observed that, compared to the initial magnetic field signal matrix, by adjusting B... X1 B X2 The filtering and noise reduction processing of the magnetic field signal matrix effectively improves the signal-to-noise ratio of the magnetic field signal, makes the stress concentration characteristics clearer, eliminates false alarms caused by interference factors such as pipeline welds and noise, and prevents missed detections.

[0041] Step S5: Calculate the sensitivity S of each value in each column of the signal in the output matrices B1 and B2 respectively, to obtain the axial stress sensitivity matrix. and bending stress sensitivity matrix Among them, the sensitivity S satisfies, .

[0042] Based on the fitting relationship between the AC electromagnetic field array detection probe detection signal and stress obtained in step S2, the axial stress sensitivity matrix S X1 and bending stress sensitivity matrix S X2 Convert it into the corresponding stress matrix.

[0043] After eliminating the stress components caused by the internal pressure of the pipe in the stress matrix, the actual axial stress matrix of the pipe is obtained. and the actual pipe bending stress matrix .

[0044] Based on steps S3 and S4, step S5 is further implemented. It is worth noting that substituting the calculation results from the previous steps yields the sensitivity matrices, which satisfy: , Specifically, the stress matrix can be obtained by fitting the detection signal of the AC electromagnetic field array detection probe and the stress based on the relationship obtained in step S2. Stress matrix Furthermore, by converting the internal pressure data obtained during pipeline inspection into axial stress components based on thin film theory, the axial stress component matrix caused by internal pressure can be obtained. For the stress matrix respectively Stress matrix Subtract stress component matrix Finally, the actual axial stress matrix of the pipeline is obtained. and the actual pipe bending stress matrix .

[0045] It is worth noting that step S5 is specifically used to calculate the impact caused by the internal pressure (component) of the pipeline, eliminating the "internal pressure effect background noise" that is prevalent in existing technologies. (See reference...) Figure 7a , Figure 7b As shown; where, Figure 7a This is a schematic diagram of the network training results without subtracting axial stress. Figure 7b This is a schematic diagram of the network training results after subtracting axial stress.

[0046] Step S6: Establish a dual-branch stress analysis network based on the fusion of CNN-Transformer feature extraction and Stacking.

[0047] It should be further noted that the structure of this bi-branch stress analytical network model is as follows: Figure 8 As shown, its training process preferably adopts a phased collaborative training strategy, and the sample feature vectors used for model training are as follows: Figure 9As shown. Specifically, the upper branch of the dual-branch stress analysis network is a bending stress analysis module based on CNN-Transformer feature extraction, consisting of a multi-scale convolutional feature extraction module, a multi-dilation rate residual convolution module, and a Transformer encoder module; its input is the actual pipe bending stress matrix. The system outputs bending stress characteristic distribution and bending feature location, along with mileage information. The lower branch of the bi-branch stress analysis network is specifically an axial stress analysis module based on a stacking integration of XGBoost and random forest, composed of the base model XGBoost and the base model random forest. Its inputs are the bending feature location output from the upper branch of the bi-branch stress analysis network and various characteristic signals of the pipeline to be quantitatively evaluated; the output is the axial stress regression prediction value. Through the above bi-branch stress analysis network design, this invention decomposes the complex mixed problem in the pipeline stress non-destructive testing process into two relatively independent sub-tasks, thereby effectively reducing the signal aliasing problem caused by complex stress in the pipeline to be quantitatively evaluated, and specifically learning the feature representation of different stress modes based on this bi-branch stress analysis network.

[0048] As a preferred embodiment of the present invention, the independent training process of the branches on the bi-branch stress analytical network in step S6 is specifically described as follows: The actual pipe bending stress matrix is ​​extracted by a multi-scale convolution feature extraction module and a multi-dilation rate residual convolution module. Local features related to mileage information.

[0049] Specifically, three parallel convolutional feature extraction modules are constructed, with reference sizes of 1×1, 3×3, and 5×5, respectively. Pre-set padding parameters ensure that the length of the input sequence is preserved after convolution. Preprocessed stress data is simultaneously fed into these three parallel convolutional branches: the 1×1 convolution performs efficient cross-channel interaction and dimensionality reduction, the 3×3 convolution captures medium-scale local stress patterns, and the 5×5 convolution extracts macroscopic distribution trends. The operation of the multi-dilation rate residual convolution module can be described as follows: , , , , .

[0050] In the above formula, w and b represent the weight and bias, respectively. The superscripts in each formula indicate the layer number, and the subscripts indicate the size of the convolutional kernel in the capping layer. , , These represent connection operations. This residual connection mechanism ensures that even when capturing long-range dependencies using a large dilation rate, the "grid effect" that might be caused by sparse sampling of convolutional kernels is avoided, guaranteeing the stability of the gradient flow and alleviating the optimization challenges of deep networks. Furthermore, it acts as an information highway, directly transmitting low-level, high-precision local details to deeper layers, preventing the loss of crucial subtle signals in the pursuit of a wide receptive field. Ultimately, this provides assistance for the complex spatiotemporal context modeling performed by the subsequent Transformer encoder.

[0051] Long-range dependencies are captured by the Transformer encoder module. For example, let's assume at least six Transformer encoder layers are used. Through the self-attention mechanism of this Transformer encoder, the limitation of CNN mode, which can only see local receptive fields, is broken, allowing features at any point in the pipeline to interact with features at all positions along the entire pipeline, ultimately outputting a high-level feature sequence that deeply integrates global contextual information.

[0052] The bending stress feature distribution and bending feature location are output through spatial attention pooling, and then optimized using a composite loss function that combines classification and regression.

[0053] The independent training process of the lower branch of the dual-branch stress analytical network in step S6 is specifically described as follows: The bending feature positions output by the branches of the bi-branch stress analysis network are fused with the feature signals of the pipeline to be quantitatively evaluated, and then input into the two base learners, XGBoost and Random Forest, in parallel.

[0054] It is worth noting that XGBoost learners are better at capturing complex nonlinear relationships and reducing bias, while random forest learners are better at resisting noise and reducing variance. By utilizing these two different base learners (bias differences), it is possible to automatically identify which features are more stable and important globally.

[0055] The XGBoost learning process is first described as follows: starting with the objective function, the t-th tree is learned; then, combining the features of the input layer, the dataset is set. .in, Given actual stress data, the target of XGBoost in round t is: ; in, It is the sample error, and the regularization term is... Its main function is to control the leaf tree of the numbers. Leaf weight Size.

[0056] By using a second-order Taylor expansion, the objective of the previous equation can be approximated as: ; in, , .

[0057] It is worth noting that XGBoost's base models support not only decision trees but also linear models; for a tree with a fixed structure We can redefine a decision tree, and its optimal leaf weights can be expressed as: ; It is worth noting that in the actual training process, when building the t-th tree, a crucial issue is how to find the optimal split point of the leaf nodes. XGBoost supports two methods for splitting nodes - a greedy algorithm and an approximate algorithm. To ensure training accuracy, the greedy algorithm is chosen here. The process starts from a tree depth of 0: enumerate all available features for each leaf node; for each feature, sort the training samples belonging to that node in ascending order according to the feature value, determine the optimal split point for that feature by linear scanning, and record the splitting gain of that feature; select the feature with the largest gain as the splitting feature, use the optimal split point of that feature as the splitting position, split the node into two new leaf nodes, and associate each new node with the corresponding sample set; return to step 1, and recursively execute until a specific condition is met.

[0058] The corresponding minimum gain: ; The XGBoost process gradually corrects the residuals to reduce bias. It controls the complexity of the model through regularization and pruning mechanisms, avoiding overfitting. Furthermore, it can use a block storage structure to achieve parallel computation, resulting in high accuracy, strong flexibility, and support for parallelism.

[0059] The learning process of random forest can be further described as follows: Since each decision tree is trained independently, parallel computation can be performed using a multi-core processor; for training M trees under the same distribution, if the variance of a single tree is... The correlation coefficient between trees is The averaged variance is: ; when When the variance approaches Random sampling increases the correlation between trees As the variance decreases, the variance also decreases significantly.

[0060] XGBoost performs fine feature selection and regularization training through gradient boosting trees, and random forest obtains robust predictions through bagging ensemble.

[0061] It is worth noting that the learning process of this random forest has the advantage of reducing the tolerance of the model through Bagging, making it less sensitive to outliers and noise; and randomly sampling on a subset of features to discover locally useful feature combinations.

[0062] The outputs of the two base learners, XGBoost and Random Forest, are stacked and fused using a linear regression meta-model (a meta-learner optimally weights and integrates the outputs of the two base learners, XGBoost and Random Forest), to obtain the axial stress regression prediction. By combining and complementing the learning processes of XGBoost and Random Forest, the model retains its advantage in preventing overfitting while inheriting the high accuracy of XGBoost in fitting complex simulations. The overall model is highly adaptable to changes in data distribution and anomalies, helping to capture patterns in the data that are both "local" and "global."

[0063] As a preferred implementation method, the fusion model resulting from combining XGBoost and Random Forest can be described as follows: ; At this point, the deviation satisfies: ; Meanwhile, the variance is: ; At this point, the fusion weights of the two models are ; The optimal weights are obtained by minimizing the expected MSE: .

[0064] The analytical solution is equivalent to using their predicted values ​​as new features and fitting them to the validation set using linear regression. At this point, stacking the XGBoost and Random Forest models significantly improves the overall model's accuracy and generalization ability, fully utilizing the strengths and characteristics of both networks. In the final prediction process, more information can be used to interpret the test results, ultimately improving prediction performance, flexibility, and interpretability.

[0065] In addition, as a preferred embodiment of the present invention, the following steps are included before implementing step S7: Step S60: Through bi-branch loss weighting and multi-task learning, the feature representations of the bending stress analysis module and the axial stress analysis module are mutually coordinated and enhanced; The hyperparameters were optimized using early stopping and cross-validation.

[0066] Statistical analysis revealed that the axial stress after step S6 ( ) and bending stress ( The coefficients of determination (R²) between the predicted values ​​and the measured values ​​obtained from the tensile test calibration reached 0.94 and 0.92, respectively, with root mean square errors (RMSE) of 8.7 MPa and 6.3 MPa. Specifically, for the confirmed true stress concentration areas, the average probability of the model outputting the region's existence was 96.2%, the average distance positioning error was less than ±15 meters, and the average absolute error between the predicted principal direction angle of bending stress and the calculated value based on strain gauge measurements was less than 5°. Under the phased training strategy described above, the CNN-Transformer bending branch converged on the validation set loss after approximately 150 epochs. The base learners (XGBoost and Random Forest) of the Stacking ensemble branch both reached performance plateaus within 50 iterations in 10-fold cross-validation. End-to-end joint fine-tuning improved overall performance by approximately 2% within an additional 30 epochs.

[0067] Step S7: Based on the bending stress feature distribution, bending feature location, and axial stress regression prediction value output by the bi-branch stress analysis network trained in Step S6, complete the location and quantitative identification of the axial stress and bending stress of the pipeline to be quantitatively evaluated.

[0068] Based on completing step S6, proceed to step S7.

[0069] It is worth noting that, based on the calculation results obtained from the aforementioned steps, the axial stress and bending stress of the pipeline to be quantitatively evaluated are further located and quantitatively identified. The final tensile stress and bending stress values ​​are shown in Table 1 below.

[0070] ; Based on the above data analysis, we can identify areas of axial stress concentration, which can be found in the following reference: Figure 11 As shown. Furthermore, it can be further observed that the feature importance distribution is highly concentrated in the original signal from the stress probe itself, as detailed in the following example. Figure 12 As shown above, the stress anomaly in this region exhibits a strong causal relationship with the stress changes directly measured by the probe, indicating that the formation of this stress concentration region is mainly attributed to direct mechanical loading, rather than secondary factors such as material inhomogeneity or local geometric variations. The elevation distribution of the other segment of the bending stress concentration region output by the dual-branch stress analysis network can be referenced as follows: Figure 10 As shown, it can be observed that the areas of concentrated bending stress within this region are mainly concentrated at angles of 0-180° and 90-270°, and the elevation changes significantly in this area (from 1220.7 meters to 1549.0 meters, a drop of approximately 328 meters). The areas of concentrated output stress can be referenced as follows: Figure 13As shown, it illustrates the distribution of stress concentration areas on the original stress signal; the accurate mileage of the bending stress concentration area can be referenced as follows. Figure 14 As shown, the stress concentration area is located between pipeline mileages 149.1 and 159.1 kilometers. The final statistical analysis yielded the topographic parameters for this curved section, as follows: Figure 15 As shown, the detection length is approximately 10 kilometers, and a total of 9,166 valid data points were collected. The data density is high and can support detailed analysis.

[0071] Further calculations and analysis revealed 18 distinct "characteristic segments" (i.e., suspected stress concentration segments) within this region that met strict criteria. These characteristic segments contributed a total of 54 significant feature points, with the 90°-270° group (P3-P7) being the most active, indicating predominantly vertical bending. The elevation in this area varied significantly (from 1220.7 meters to 1549.0 meters, a drop of approximately 328 meters) and spanned a wide range of latitude and longitude. Therefore, the following conclusion can be drawn: the detected bending stress is highly likely directly related to the pipeline traversing complex terrain (such as hillsides and valleys). The terrain changes caused uneven stress on the pipeline, resulting in continuous bending stress, consistent with the results in the aforementioned images.

[0072] Ultimately, this invention demonstrates that the quantitative assessment method for pipeline axial and bending stress based on alternating electromagnetic fields successfully achieves quantitative assessment of pipeline axial and bending stress, enabling accurate tracing and qualitative identification of the root causes of stress concentration. This provides crucial scientific decision-making basis for subsequent targeted verification of stress distribution, formulation of repair strategies, and implementation of engineering interventions.

[0073] This invention provides a method for quantitatively assessing the axial and bending stresses of a pipeline based on an alternating current electromagnetic field. The method includes at least the following steps: Step S1: Arranging several AC electromagnetic field array detection probes at equal intervals along the Y-direction at locations on the surface of the pipeline to be quantitatively assessed using a double-row arrangement; Step S2: Performing stress calibration on the AC electromagnetic field array detection probes to establish a fitting relationship between the detection signals and stresses; Step S3: Using the AC electromagnetic field array detection probes, scanning the interior of the pipeline to be quantitatively assessed along the X-direction to obtain two sets of magnetic field signal matrices under excitation in different directions; Step S4: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] X1 B X2 Each data point in the magnetic field signal matrix undergoes filtering and noise reduction processing; Step S5: Calculate the sensitivity S of each value in each column of the output matrices B1 and B2 respectively to obtain the axial stress sensitivity matrix and the bending stress sensitivity matrix; After eliminating the stress components caused by the internal pressure of the pipe in the stress matrix, obtain the actual pipe axial stress matrix and the actual pipe bending stress matrix. Step S6: Establish a dual-branch stress analysis network based on CNN-Transformer feature extraction and Stacking fusion; Step S7: Based on the bending stress feature distribution, bending feature location and axial stress regression prediction value output by the dual-branch stress analysis network trained in Step S6, complete the localization and quantitative identification of the axial stress and bending stress of the pipeline to be quantitatively evaluated.

[0074] The quantitative assessment method for axial and bending stress of pipelines based on alternating electromagnetic fields, which has the above-described steps, has at least the following technical advantages compared to existing technologies: (1) The present invention provides a quantitative assessment method for axial stress and bending stress of pipeline based on AC electromagnetic field. Through the layout design of AC electromagnetic field array detection probe with double row orthogonal excitation, the independent synchronous detection of axial stress and bending stress of pipeline is realized. Furthermore, by constructing a training dual-branch stress analysis network, the complementary enhancement of bending stress analysis and axial stress analysis is realized. (2) The present invention provides a quantitative assessment method for axial stress and bending stress of pipelines based on AC electromagnetic field. By calibrating the fitting relationship between the detection signal and stress of the AC electromagnetic field array detection probe, and then combining the physical decoupling and fusion analysis of stress-related data, it effectively solves a series of technical problems that are difficult to quantify and analyze in the existing single excitation detection, such as inaccurate bending stress identification, large direction interpretation error and large amount of feature data. Finally, it provides highly reliable data support and intelligent analysis means for pipeline integrity management and safety assessment.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A quantitative assessment method for axial and bending stress in pipelines based on alternating electromagnetic fields, characterized in that, The steps include the following: Step S1: Define the axial direction of the pipeline to be quantitatively evaluated as the X direction, the circumferential direction of the pipeline to be quantitatively evaluated as the Y direction, and the radial direction of the pipeline to be quantitatively evaluated as the Z direction. A double-row arrangement is used, with several AC electromagnetic field array detection probes evenly spaced along the Y direction at the location on the pipe surface to be quantitatively evaluated. The first row of AC electromagnetic field array detection probes is used to detect the axial stress of the pipe, with the internal excitation coil and the magnetic sensor both positioned in the Y direction. The second row of AC electromagnetic field array detection probes is used to detect the bending stress of the pipe, with the internal excitation coil and the magnetic sensor both positioned in the X direction. Step S2: Perform stress calibration on the AC electromagnetic field array detection probe to establish a fitting relationship between the detection signal of the AC electromagnetic field array detection probe and the stress. Step S3: Using an AC electromagnetic field array detection probe, scan the inside of the pipe to be quantitatively evaluated along the X-axis to obtain two sets of magnetic field signal matrices under excitation in different directions, as follows: , ; Step S4: For B X1 B X2 Each data point in the magnetic field signal matrix is ​​filtered and denoised to obtain the filtered and denoised output matrix. , ; Step S5: Calculate the sensitivity S of each value in each column of the signal in the output matrices B1 and B2 respectively, to obtain the axial stress sensitivity matrix. and bending stress sensitivity matrix Among them, the sensitivity S satisfies, ; Based on the fitting relationship between the AC electromagnetic field array detection probe detection signal and stress obtained in step S2, the axial stress sensitivity matrix S X1 and bending stress sensitivity matrix S X2 Convert to the corresponding stress matrix; After eliminating the stress components caused by the internal pressure of the pipe in the stress matrix, the actual axial stress matrix of the pipe is obtained. and the actual pipe bending stress matrix ; Step S6: Establish a dual-branch stress analysis network based on the fusion of CNN-Transformer feature extraction and Stacking; Specifically, the upper branch of the dual-branch stress analysis network is a bending stress analysis module based on CNN-Transformer feature extraction, which consists of a multi-scale convolutional feature extraction module, a multi-expansion rate residual convolution module, and a Transformer encoder module; its input is the actual pipe bending stress matrix. The output includes the distribution and location of bending stress features, along with mileage information. The lower branch of the bi-branch stress analysis network is an axial stress analysis module based on Stacking integration of XGBoost and Random Forest, which consists of the base model XGBoost and the base model Random Forest. Its input is the bending feature location and various feature signals of the pipeline to be quantitatively evaluated output by the lower branch of the bi-branch stress analysis network, and the output is the axial stress regression prediction value. Step S7: Based on the bending stress feature distribution, bending feature location, and axial stress regression prediction value output by the bi-branch stress analysis network trained in Step S6, complete the location and quantitative identification of the axial stress and bending stress of the pipeline to be quantitatively evaluated.

2. The method for quantitative evaluation of axial and bending stress in pipelines based on alternating electromagnetic fields according to claim 1, characterized in that, The process of step S2 specifically includes the following steps: Step S21: Perform tensile stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for tensile stress calibration; place the AC electromagnetic field array detection probe in the middle of the standard test block for tensile stress calibration, and make the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor perpendicular to the tensile direction. Tensile stress is gradually applied starting from 0 MPa with a fixed stress increment; and the stress is paused at a fixed time interval after each tensile stress increment stabilizes. The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the tensile stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the tensile stress are calibrated at fixed intervals after each tensile stress increase stabilizes. Step S22: Perform bending stress calibration on the AC electromagnetic field array detection probe: Design a standard test block for bending stress calibration; place the AC electromagnetic field array detection probe on the upper or lower surface of the standard test block between the three bending pressure application point and the support point, so that the setting direction of the excitation coil of the AC electromagnetic field array detection probe and the sensitive direction of the magnetic sensor are parallel to the surface stress direction of the standard test block during the bending process. The bending stress is gradually applied starting from 0 MPa with a fixed stress increase; and a fixed time interval is set after each bending stress increase stabilizes. The detection signal of the AC electromagnetic field array detection probe is acquired in real time during the bending stress calibration process; and the detection signal of the AC electromagnetic field array detection probe and the bending stress are calibrated at fixed intervals after each bending stress increase stabilizes.

3. The method for quantitative evaluation of axial and bending stress in pipelines based on alternating electromagnetic fields according to claim 1, characterized in that, In step S4, B X1 B X2 The process of filtering and denoising each data point in the magnetic field signal matrix is ​​specifically described as follows: Set the width of the movable window to 2a+1; where a is a positive integer; In the moving window, for B X1 B X2 For each data point in the magnetic field signal matrix, select 2m+1 data points within a window centered on it to form a subsequence {xi−m,...,xi,...,xi+m}; where m is a positive integer and satisfies m>a; Perform k-th order polynomial least squares fitting on the subsequence to obtain the fitting polynomial. ;in, ; These are the relative position coordinates within the window; Compute polynomials The function value y at the center of the window is used as the filtered output value of the data point; Slide the window to the next data point and repeat the above process until all data has been traversed, generating the filtered and denoised output matrix. , .

4. The method for quantitative evaluation of axial and bending stress in pipelines based on alternating electromagnetic fields according to claim 1, characterized in that, The independent training process of the branches on the bi-branch stress analytical network in step S6 is specifically described as follows: The actual pipe bending stress matrix is ​​extracted by a multi-scale convolution feature extraction module and a multi-dilation rate residual convolution module. Local features related to mileage information; Long-range dependencies are captured by the Transformer encoder module; The bending stress feature distribution and bending feature location are output through spatial attention pooling, and then optimized using a composite loss function that combines classification and regression.

5. The method for quantitative evaluation of axial and bending stress in pipelines based on alternating electromagnetic fields according to claim 1, characterized in that, The independent training process of the lower branch of the bi-branch stress analytical network in step S6 is specifically described as follows: After fusing the bending feature positions of the branch outputs on the bi-branch stress analysis network with the feature signals of the pipeline to be quantitatively evaluated, the results are input in parallel into two base learners: XGBoost and Random Forest. XGBoost performs fine feature selection and regularized training through gradient boosting trees, and random forest obtains robust predictions through bagging ensemble. The outputs of the two base learners, XGBoost and Random Forest, are fused using a linear regression meta-model through stacking to obtain the axial stress regression prediction.

6. The method for quantitative evaluation of axial and bending stress in pipelines based on alternating electromagnetic fields according to claim 1, characterized in that, Before implementing step S7, the following steps are also included: Step S60: Through bi-branch loss weighting and multi-task learning, the feature representations of the bending stress analysis module and the axial stress analysis module are mutually coordinated and enhanced; The hyperparameters were optimized using early stopping and cross-validation.

Images