Method for quantifying metal loss from pipeline magnetic flux leakage signals

Abstract

The application relates to the technical field of pipeline magnetic flux leakage detection, and provides a pipeline magnetic flux leakage signal metal loss quantification method, sample sets and test sets are extracted from metal loss defect samples, and pipeline magnetic flux leakage signals are extracted. The defect boundary and the peak position are determined based on axial signals, radial signal column data subsets are divided, and a comprehensive distance feature value is calculated; the lateral effective subset is determined by analyzing the wave peak and wave trough indexes of the radial signals, the maximum difference position is calculated, the width feature value is calculated by combining the boundary coefficient and the channel number; the integral mean of the axial signals is a first depth feature, the subset of the circumferential signal containing the core magnetic field distortion information of the defect is extracted, and the integral mean thereof is a second depth feature, and the depth feature value is obtained by weighted fusion; the sample sets are used for training a deep neural network model, and the feature values of the test sets are input into the trained model to obtain a three-dimensional size quantification result of the defect. The method is processed through regionalized signals and multi-dimensional feature fusion, and the quantification precision and the anti-interference capability of complex defects are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipeline magnetic flux leakage detection, and particularly relates to a pipeline magnetic flux leakage signal metal loss quantification method. BACKGROUND

[0002] Pipelines are crucial transportation carriers in modern industrial and energy systems. During long-term operation, irregular wear is formed on the inner wall of the pipeline due to the scouring of solid impurities in the conveying medium, and the outer wall may be electrochemically corroded due to chloride ions and sulfides in the soil, and then local stress concentration caused by crustal movement easily breeds various defects from micron-level cracks to centimeter-level pits. As the mainstream means in the current pipeline nondestructive testing field, the core principle of the magnetic flux leakage signal detection technology is to magnetize the pipeline wall to a saturated state through a permanent magnet, and when the pipeline wall has defects, the magnetic force lines will be distorted at the defects and form a magnetic flux leakage field, and then the strength and distribution characteristics of the magnetic flux leakage signal are captured by a high-sensitivity Hall sensor.

[0003] This technology not only can complete the detection without damaging the integrity of the pipeline, but also can accurately identify various defect types such as cracks, pits and corrosion pits. However, even if the defects belong to the metal loss type, when the length increases from a few millimeters to tens of millimeters and the depth increases from 10% to 50% of the thickness of the pipeline wall, the peak strength, waveform width and gradient change of the magnetic flux leakage signal will all show significant differences. The existing pipeline defect quantification scheme is mostly based on the pipeline magnetic flux leakage signal and a neural network model, for example, the length is quantified by the axial peak-valley distance, the width is quantified by the circumferential median distance, and the depth is quantified by combining the radial peak-valley value with the neural network; or the global features are extracted and input into the neural network to realize quantification.

[0004] However, the above methods generally have the problems of missing signal regionalization processing and insufficient feature fine extraction, which limits the quantification accuracy of complex defects and has insufficient anti-interference ability, and it is difficult to meet the high-precision detection demand. SUMMARY

[0005] To solve the problems of missing signal regionalization processing and insufficient feature fine extraction in the existing pipeline defect quantification scheme, the present application provides a pipeline magnetic flux leakage signal metal loss quantification method, which comprises the following steps:

[0006] extracting a sample set and a test set from a metal loss defect sample;

[0007] extracting a pipeline magnetic flux leakage signal of the metal loss defect, wherein the pipeline magnetic flux leakage signal comprises an axial signal, a radial signal and a circumferential signal;

[0008] determine left and right boundaries and a peak position of the defect based on the axial signal, divide a region based on the peak position and calculate a comprehensive distance feature representing an extension degree of the defect based on a column data subset defined by the left and right boundaries to obtain a length feature value;

[0009] analyze a peak and a valley index of each row data of the radial signal to determine a transverse effective subset of the defect, calculate a peak and a valley difference of each row data in the effective subset and determine a maximum difference position, combine a boundary coefficient to calculate a channel number representing a width of the defect to obtain a width feature value;

[0010] calculate an integral mean of the axial signal as a first depth feature, extract a subset containing core magnetic field distortion information of the defect from the circumferential signal and calculate an integral mean thereof as a second depth feature, and perform weighted fusion on the first depth feature and the second depth feature to obtain a depth feature value;

[0011] input the length feature value, the width feature value and the depth feature value of the defects in the sample set and actual sizes thereof as training data into a deep neural network model to train the deep neural network model and establish a nonlinear mapping relationship from the feature values to the actual sizes;

[0012] input the feature values of the defects in the test set into the trained deep neural network model to obtain length, width and depth quantization results of the defects.

[0013] In a possible implementation, the determination of the left and right boundaries and the peak position of the defect based on the axial signal, the division of the region based on the peak position and the calculation of the comprehensive distance feature representing the extension degree of the defect based on the column data subset defined by the left and right boundaries to obtain the length feature value include the following steps:

[0014] calculate an accumulated sum of each column data based on the axial signal to determine the left and right boundaries and the peak position of the defect;

[0015] extract a column data subset defined by the left and right boundaries from the radial signal;

[0016] divide the column data subset into a left subset region and a right subset region based on the peak position as a boundary, and calculate distance arrays of the left subset region and the right subset region, respectively;

[0017] fuse the distance arrays of the left subset region and the right subset region into a comprehensive distance array;

[0018] optimize a stability coefficient according to a difference between actual lengths of defects in the sample set and the comprehensive distance array, and filter the length feature value based on the optimized stability coefficient.

[0019] In an implementable manner, the stability coefficient is optimized according to the difference between the actual length of the defects in the sample set and the comprehensive distance array, and the length characteristic value is screened by using the optimized stability coefficient, including the steps of:

[0020] For each defect in the sample set, the ratio of the index corresponding to the value closest to the actual length in the comprehensive distance array to the total number of rows is calculated as an individual stability coefficient;

[0021] The individual stability coefficients of all defects in the sample set are averaged to obtain the stability coefficient;

[0022] In the length quantification of the defects in the test set, the length characteristic value is directly calculated by using the stability coefficient.

[0023] In an implementable manner, the width characteristic value is obtained by analyzing the peak and valley indexes of each row of data of the radial signal to determine the transverse effective subset of the defects, calculating the peak and valley difference of each row of data in the effective subset and determining the position of the maximum difference, and calculating the channel number representing the width of the defects in combination with the boundary coefficient.

[0024] For each row of data of the radial signal, the peak index and the valley index are extracted;

[0025] The maximum value of the peak indexes of all rows is taken as the final peak index, and the minimum value of the valley indexes of all rows is taken as the final valley index;

[0026] The effective subset of the radial signal is extracted based on the final peak index and the final valley index;

[0027] The peak and valley difference sequence of each row of data in the effective subset is calculated, and the maximum difference and the corresponding row number are determined;

[0028] The width characteristic value is obtained by calculating the expanded row number by using the boundary coefficient and combining the fixed row distance.

[0029] In an implementable manner, the width characteristic value is obtained by calculating the expanded row number by using the boundary coefficient and combining the fixed row distance, including the steps of:

[0030] The actual width of each defect in the sample set, the boundary coefficient, and the corresponding width characteristic value are input into the deep neural network model, and the boundary coefficient includes the upper boundary coefficient and the lower boundary coefficient;

[0031] The optimal upper boundary coefficient and the optimal lower boundary coefficient that minimize the width quantification error are output by training iteration;

[0032] The optimal upper boundary coefficient and the optimal lower boundary coefficient are directly used in the width quantification of the defects in the test set to calculate the width feature value.

[0033] In a feasible implementation, the calculation of the integral mean of the axial signal as a first depth feature, the extraction of a subset containing defect core magnetic field distortion information from the circumferential signal and the calculation of the integral mean thereof as a second depth feature, and the weighted fusion of the first depth feature and the second depth feature to obtain a depth feature value, include the steps of:

[0034] The integral value of each row of data of the axial signal is calculated and the mean is taken as a first depth feature value;

[0035] The upper half of the data from the circumferential signal is extracted to form a first subset, and the data between the left trough index and the right trough index of each row of data of the first subset is extracted to reconstruct a second subset;

[0036] The integral value of each row of data of the second subset is calculated and the mean is taken as a second depth feature value;

[0037] A linear combination model is constructed through a correlation coefficient, and a weighted sum of the first depth feature value and the second depth feature value is used to predict the depth to obtain the depth feature value.

[0038] In a feasible implementation, the construction of a linear combination model through a correlation coefficient, and the prediction of the depth by a weighted sum of the first depth feature value and the second depth feature value to obtain the depth feature value, include the steps of:

[0039] The actual depth of each defect in the sample set, the first depth feature value, and the second depth feature value are input into the depth neural network model;

[0040] An optimal correlation coefficient that minimizes the depth quantification error is determined through training iterations, and a constraint condition that the weighted sum is 1 is satisfied;

[0041] The optimal correlation coefficient is directly used in the depth quantification of the defects in the test set to calculate the depth feature value.

[0042] In a feasible implementation, one-third of the defect points in the metal loss defect sample are randomly selected as the sample set, and the remaining one-fourth of the defect points in the metal loss defect sample are selected as the test set.

[0043] In a feasible implementation, the training of the depth neural network model uses a sample set and test set separation mode, the sample set is used for iterative optimization of model parameters and coefficients, and the test set is used for verification of quantification accuracy.

[0044] In a feasible implementation, the pipeline magnetic flux leakage signal is extracted based on a pipeline magnetic flux leakage detection original file.

[0045] The axial signal, the radial signal and the circumferential signal are all expressed in a matrix form, and the pipeline magnetic flux leakage signals corresponding to the same metal loss defect have consistent row numbers and column numbers.

[0046] From the above, the application provides a pipeline magnetic flux leakage signal metal loss quantification method. According to the morphological differences of defects, different defect morphologies are separated by a multi-dimensional feature extraction algorithm to obtain corresponding feature parameters; then, a training model is constructed based on a deep learning framework, and threshold optimization and parameter iteration are performed for different morphological defect samples to optimize the generalization ability of the model; finally, the trained model is used to realize accurate inversion of the defect length, depth and width, and the accuracy and robustness of the metal loss defect size quantification are significantly improved. The method of the application comprehensively utilizes three-axis signals, deeply excavates the relationship between signal features and defect size, and adopts a targeted quantitative calculation method to improve the accuracy of defect quantification. At the same time, the deep neural network is used for training and optimization to enhance the adaptability of the method to different samples, realize accurate quantification of the metal loss type defect size, and meet the high-precision requirements of pipeline defect detection in actual engineering. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings, which are incorporated into and form a part of the specification, illustrate an embodiment consistent with the present application and, together with the description, serve to explain the principles of the embodiments of the application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained from these drawings without creative labor for those skilled in the art.

[0048] Figure 1 A flowchart of the pipeline magnetic flux leakage signal metal loss quantification method shown in the exemplary embodiment of the application;

[0049] Figure 2 A schematic diagram of the defect axial signal shown in the exemplary embodiment of the application;

[0050] Figure 3 A schematic diagram of the defect radial signal shown in the exemplary embodiment of the application;

[0051] Figure 4 A schematic diagram of the defect circumferential signal shown in the exemplary embodiment of the application. DETAILED DESCRIPTION

[0052] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations, however, can be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art. The described features, acts, or characteristics can be combined in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of example implementations.

[0053] Pipelines are important transport carriers in modern industrial and energy systems. In long-term operation, irregular wear is formed on the inner wall of the pipeline due to the scouring of solid impurities in the transported medium, and the outer wall is prone to various defects due to electrochemical corrosion of chloride ions and sulfides in the soil and local stress concentration caused by crustal movement. Magnetic flux leakage signal detection technology is the current mainstream method for pipeline non-destructive testing. Its principle is to magnetize the pipeline wall to saturation with a permanent magnet, and when the pipeline wall has defects, the magnetic force lines are distorted to form a magnetic flux leakage field, which is captured by a high-sensitivity Hall sensor. This technology can complete the detection without damaging the integrity of the pipeline and accurately identify multiple defect types. However, even if the defects belong to the same metal loss type, the length increases from a few millimeters to tens of millimeters, and the depth increases from 10% to 50% of the pipeline wall thickness, the peak intensity, waveform width, and gradient change of the magnetic flux leakage signal will have significant differences.

[0054] Existing pipeline defect quantification methods include constructing a sample set by obtaining pipeline magnetic flux leakage signals, quantifying length by the peak-to-valley distance of the axial component differential signal, quantifying width by the median distance between peaks and valleys of the circumferential component, and combining neural networks and the peak-to-valley value of the radial component to quantify depth.

[0055] However, the above solutions have obvious defects and problems, including: 1. Coarse feature extraction: existing technologies have simple feature extraction of magnetic flux leakage signals, such as using only a single feature such as the peak-to-valley distance of the axial component differential or the median distance between peaks and valleys of the circumferential component, or not dividing the signal region according to the physical characteristics of each dimension of the defect, failing to deeply mine the subtle changes of the signal in different regions, resulting in incomplete capture of the characteristics of complex defects. 2. Weak anti-interference ability: the signal features relied on by existing technologies are easily affected by external factors, such as fluctuation in the lift-off distance and complex defect morphology, which can distort the linear relationship between a single feature and the size, and global features are also easily disturbed by irrelevant noise, limiting the model's generalization ability and reducing the quantification accuracy. 3. Inaccurate correlation between features and defects: existing technologies directly use peak-to-valley features to fit the size, or the correlation between the extracted features and the defect size is weak, failing to deeply mine the internal correlation between the features and the geometric parameters of the defect, resulting in difficulty in establishing an accurate mapping, especially for non-standard defects with large quantification errors.

[0056] To solve the above problems, the embodiment of the application provides a pipeline magnetic flux leakage signal metal loss quantification method, as shown in the figure, the method comprises the following steps: Figure 1

[0057] S100: extracting a sample set and a test set from a metal loss defect sample.

[0058] In some embodiments, one third of the defect points in the metal loss defect sample are randomly extracted as the sample set, and the remaining one fourth of the defect points are taken as the test set. In the implementation process, first, a metal loss defect sample library is established, and the metal loss defect sample library is the defect data of the actual pulled pipeline. The sample set is divided into a 75% training set and a 25% test set through a random sampling program. The separation design of the sample set and the test set ensures the objectivity of model evaluation and ensures data distribution consistency.

[0059] S200: extracting a pipeline magnetic flux leakage signal of a metal loss defect, the pipeline magnetic flux leakage signal comprising an axial signal, a radial signal and a circumferential signal.

[0060] Specifically, pipeline magnetic flux leakage signal data is read from a pipeline magnetic flux leakage detection original file. The axial signal matrix y_data, the radial signal matrix z_data and the circumferential signal matrix x_data are all stored in the form of an m row n column matrix, and the number of rows m and the number of columns n of the three-axis signal matrix are strictly aligned to ensure spatial position consistency. The synchronous extraction of the three-axis signal and the matrix alignment processing ensure the accuracy of feature extraction.

[0061] S300: extracting a length feature value, a width feature value and a depth feature value of the defect based on the pipeline magnetic flux leakage signal.

[0062] The method comprises the following steps:

[0063] S310: the length feature value is extracted using the axial signal and the radial signal. Specifically, the left and right boundaries and the peak position of the defect are determined based on the axial signal. The column data subset defined by the left and right boundaries is divided into regions with the peak position as the boundary, and the comprehensive distance feature representing the extension degree of the defect is calculated to obtain the length feature value.

[0064] S320: the width feature value is extracted using the radial signal. Specifically, the wave peak and wave trough indexes of each row of data of the radial signal are analyzed to determine the transverse effective subset of the defect. The wave peak and wave trough difference of each row of data in the effective subset is calculated, and the maximum difference position is determined. The number of channels representing the width of the defect is calculated in combination with the boundary coefficient to obtain the width feature value.

[0065] ​S330: The extraction of the depth feature value adopts the axial signal and the circumferential signal. Specifically, the integral mean of the axial signal is calculated as a first depth feature, a subset containing the core magnetic field distortion information of the defect is extracted from the circumferential signal, and the integral mean of the subset is calculated as a second depth feature. The first depth feature and the second depth feature are weighted and fused to obtain the depth feature value.

[0066] wherein the length feature extraction module determines the left and right boundaries y1 and y2 of the defect and the peak position y_max based on the cumulative sum of each column of data of the axial signal, and calculates a comprehensive distance array in combination with the corresponding column data subset of the radial signal; the width feature extraction module extracts an effective subset through the peak and valley index of each row of the radial signal, and calculates a feature value in combination with the upper boundary coefficient c_up and the lower boundary coefficient c_down; and the depth feature extraction module calculates the integral mean featureAs of the axial signal and the integral mean featureCs of the subset of the circumferential signal, respectively, and fuses the features through a linear combination model: Depthi=x1 featureAs+x2 featureCs, wherein x1 is the weight coefficient of the integral mean of the axial signal, and x2 is the weight coefficient of the integral mean of the subset of the circumferential signal.

[0067] S400: The length feature value, the width feature value and the depth feature value of the defects in the sample set are taken as training data together with the actual sizes thereof, and are input into a deep neural network model for training, so as to establish a nonlinear mapping relationship from the feature values to the actual sizes.

[0068] Further, the deep neural network model adopts a multi-layer full connection structure, the input layer receives a feature vector, the hidden layer introduces a ReLU activation function to enhance the nonlinear fitting capability, and the output layer linearly regresses to predict the defect size. This step realizes accurate mapping from the feature to the size through the combination of multi-dimensional feature extraction and the deep learning model, and the data-driven optimization of the key coefficients improves the adaptability of the method to the change of the defect morphology.

[0069] S500: The feature values of the defects in the test set are input into the trained deep neural network model, and the length, width and depth quantitative results of the defects are obtained through inversion.

[0070] In the training process, a mean square error loss function and an Adam optimizer are adopted, the key coefficients are iteratively optimized, and a nonlinear mapping relationship between the feature values and the actual sizes is established. In the test set verification stage, the extracted feature values are input into the trained model, and the quantitative results are obtained through inversion.

[0071] The embodiment realizes the whole process optimization from feature extraction to size prediction through systematic sample division, multi-scale feature extraction and deep neural network model. The separation design of sample set and test set ensures the objective evaluation of model generalization ability, the spatial alignment processing of three-axis signals improves the reliability of feature extraction, and the iterative optimization of key coefficients enhances the adaptability of the method to complex defect morphology. The mapping relationship formed by the deep neural network model through iterative optimization makes the quantitative results of the test set defects have high reliability, and the model shows good generalization ability to pipes of different materials, pipe diameters and defect types, meeting the demand of high-precision defect detection in actual engineering.

[0072] In some embodiments of the present application, the step S310 of extracting the length feature value using the axial signal and the radial signal specifically includes:

[0073] S311: Calculate the cumulative sum of each column of data based on the axial signal to determine the left and right boundaries of the defect and the peak position;

[0074] S312: Extract the column data subset defined by the left and right boundaries from the radial signal;

[0075] S313: Divide the radial signal subset into left and right subset regions with the peak position as the boundary, and calculate the distance array of the left and right subset regions respectively;

[0076] S314: Fuse the distance arrays of the left and right subset regions into a comprehensive distance array;

[0077] S315: Optimize the stability coefficient according to the difference between the actual length of the defect in the sample set and the comprehensive distance array, and select the length feature value with the optimized stability coefficient.

[0078] Specifically, as shown in Figure 2 and Figure 3 In the length feature extraction module, first, the column-by-column cumulative sum of the axial signal matrix y_data is calculated, and the index corresponding to the minimum two columns of the cumulative sum is determined as the left boundary y1 and the right boundary y2 through threshold comparison, and the peak position y_max corresponding to the maximum column of the cumulative sum is located.

[0079] Based on the index range of y1 and y2, the corresponding column data subset is extracted from the radial signal matrix z_data. The radial signal subset is divided into a left subset region (y1 to y_max columns) and a right subset region (y_max to y2 columns) with y_max as the boundary.

[0080] As shown in Figure 3 Because the change trend of the defect radial signal presents the characteristics of "first falling to the peak valley point, and then rising to the peak-peak value point", the following processing is performed:

[0081] In the left subset region, find the minimum value point row by row, calculate the point number difference between each minimum value point and the y_max index, form the left distance array. Among them, the difference point number of the first row is recorded as zl1, the difference point number of the second row is recorded as zl2, and the difference point number of the mth row is recorded as zlm, thereby obtaining an array [zl1, zl2, …, zlm] containing m values. After sorting the array in ascending order, the distance array [zldis1, zldis2, …, zldism] of the left subset region is calculated by combining the fixed distance dis_per_point between each point in the axial direction (each value is the product of the corresponding difference point number and dis_per_point).

[0082] In the right subset region, find the maximum value point row by row, calculate the point number difference between each maximum value point and the y_max index, form the right distance array. Among them, the difference point number of the first row is recorded as zr1, the difference point number of the second row is recorded as zr2, and the difference point number of the mth row is recorded as zrm, thereby obtaining an array [zr1, zr2, …, zrm] containing m values. The same processing method as the left subset region is used to calculate the distance array [zrdis1, zrdis2, …, zrdism] of the right subset region.

[0083] Fuse the left and right distance arrays by adding corresponding elements to obtain the comprehensive distance array. That is, add the distance arrays of the left and right subset regions corresponding elements, that is, add zldis1 and zrdis1, add zldis2 and zrdis2, …, add zldism and zrdism, to obtain a distance array containing m values.

[0084] Finally, the stability coefficient thre_value_stability is optimized by the minimum difference between the actual length of the defects in the sample set and the values in the comprehensive distance array, and the most representative length characteristic value is selected.

[0085] In this embodiment, the precise positioning of the defect boundary is realized by column analysis of the axial signal accumulation sum, which provides an accurate column range for the radial signal subset extraction. The determination of the peak position refines the division of the subset region, so that the distance array calculation can reflect the subtle changes of the defect morphology. Finally, the fusion of the comprehensive distance array and the optimization of the stability coefficient ensure the high representativeness and anti-interference ability of the length characteristic value.

[0086] The embodiment realizes accurate definition of the defect boundary by joint analysis of axial signal accumulation and peak positioning, distance array fusion of left and right subset regions and stability coefficient optimization, effectively improves the characterization ability of the characteristic value to the actual length of the defect, and solves the problems of noise interference and insufficient stability of the characteristic value in the traditional method. The precision and robustness of the length quantification of the metal loss defect are effectively improved.

[0087] In some embodiments of the present application, the determination of the stability coefficient specifically comprises:

[0088] S3151: For each defect in the sample set, calculate the ratio of the index corresponding to the value closest to the actual length in the comprehensive distance array to the total number of rows as a personal coefficient;

[0089] S3152: Average the personal coefficients of all defects in the sample set to obtain the stability coefficient;

[0090] S3153: In the length quantification of the test set defects, directly use the stability coefficient to calculate the length characteristic value.

[0091] In the stability coefficient optimization module, for each defect in the sample set, calculate the index position of the value in the comprehensive distance array with the minimum actual length difference, and define the ratio of the index to the total number of rows m of the matrix as the personal coefficient thre_value of the defect. By traversing all defects in the sample set, the arithmetic average of all personal coefficients is calculated to obtain the stability coefficient thre_value_stability. Therefore, when calculating the length characteristic value of the defect in the sample set, the value in the incremental array corresponding to the index "m x thre_value_stability" (integer) is taken as the most stable characteristic value.

[0092] In the length quantification process of the test set defects, the comprehensive distance array is directly filtered using the stability coefficient to extract the most representative length characteristic value. The global design of the coefficient ensures the consistency of the quantification of different morphological defects.

[0093] It can be understood that the calculation of the personal coefficient and the global average form the stability coefficient with statistical significance, avoiding frequent adjustment of the coefficient due to the difference in defect morphology, effectively solving the problem of dependence of the coefficient on a single defect morphology. At the same time, the application in the test set ensures the stability and consistency of the extraction of the length characteristic value, so that the length quantification of the test set defects is based on a unified coefficient standard, improving the universality and engineering applicability of the method.

[0094] In some embodiments of the present application, the step S320 of extracting the width characteristic value using the radial signal specifically comprises:

[0095] S321: Extract the peak index and the valley index for each row of data of the radial signal;

[0096] S322: Take the maximum value of the peak indexes of all rows as the final peak index, and take the minimum value of the valley indexes of all rows as the final valley index;

[0097] S323: Extract the effective subset of the radial signal based on the final peak index and the final valley index;

[0098] S324: Calculate the difference sequence between the peak and the valley of each row of data in the effective subset, and determine the maximum difference value and its corresponding row number;

[0099] S325: Calculate the number of extended rows by the upper boundary coefficient and the lower boundary coefficient, and obtain the width characteristic value combined with the fixed row distance.

[0100] Specifically, in the width feature extraction module, the radial signal data z_data of the defect is used for width quantization. The change feature of the defect radial signal is "firstly reduced to the valley, then rises to the peak, and then falls again", therefore, for each row of data in the radial signal z_data, the indexes corresponding to the valley and the peak are calculated respectively, wherein the index corresponding to the peak is denoted as index_p, and the index corresponding to the valley is denoted as index_v, that is, the peak and valley index extraction of each row of data is completed.

[0101] The valley indexes of m rows of data are collected to form a valley sequence [index_v1, index_v2,..., index_vm], and the minimum index is selected as the final valley index index_ve.

[0102] The peak indexes of m rows of data are collected in the same way as the valley indexes to form a peak sequence [index_p1, index_p2,..., index_pm], and the maximum index is selected as the final peak index index_pe.

[0103] Based on the index range of index_pe and index_ve, an effective subset is extracted from the radial signal. That is, for m rows of data of the radial signal z_data, only the column data defined by the valley index index_ve to the peak index index_pe is extracted to form a new m row data subset as the effective subset, denoted as subc.

[0104] In the obtained effective subset subc, for each row of data, the difference between the peak and the trough is calculated to obtain a difference sequence [difference1, difference2, …, differencem]. The element with the maximum difference in the sequence is found, and the row number corresponding to the element is denoted as differencemax_index, and the maximum difference is denoted as differencemax, where the value range of differencemax_index is 0 < differencemax_index < m.

[0105] According to the signal characteristics, the values in the difference sequence present the rule of "first increasing to the maximum value differencemax, and then gradually decreasing". Since the defect width is related to the number of rows affected by the defect, for each defect in the sample set, the width feature value is calculated in the following manner:

[0106] Taking the row number differencemax_index corresponding to differencemax as the reference, x_up rows are expanded upwards as the upper boundary of the defect, and y_down rows are expanded downwards as the lower boundary of the defect. The difference between x_up and y_down is calculated, which is the channel count channelcount affected by the defect. Since the distance between each row is a constant value channelvalue, the width feature value feature_width of the defect is the product of the channel count channelcount and the channel value, that is: feature_width = channelcount × channelvalue.

[0107] In this embodiment, the extraction of the peak and trough indexes realizes the accurate positioning of the effective subset, provides a data basis for the subsequent difference sequence calculation, and the positioning of the maximum difference and the corresponding row number reflects the key characteristics of the defect width; the application of the upper boundary coefficient and the lower boundary coefficient realizes the quantitative calculation of the feature value, which guarantees the accuracy of the width quantization in combination with the fixed row distance, and improves the representation ability of the feature value to the actual width through the optimization of the boundary coefficients.

[0108] In some embodiments of the present application, the optimization of the upper boundary coefficient and the lower boundary coefficient includes the following steps:

[0109] S3251: input the actual width, the upper boundary coefficient, the lower boundary coefficient and the corresponding width feature value of each defect in the sample set into a deep neural network model;

[0110] S3252: output the optimal upper boundary coefficient and the optimal lower boundary coefficient that minimize the width quantization error through training iteration;

[0111] S3253: The optimal upper boundary coefficient and the optimal lower boundary coefficient are directly used to calculate the width characteristic value in the wide quantification of the test set defects.

[0112] In the width feature optimization module, the actual width of each defect in the sample set, the upper boundary coefficient c up temp, the lower boundary coefficient c down temp and the corresponding width characteristic value feature width temp are input into the deep neural network model.

[0113] wherein the determination of the upper boundary expansion row number x up comprises the steps of setting the upper boundary coefficient as c up, calculating the upper boundary value: difference up = difference max * c up. In the difference value sequence [difference1, difference2, …, differencem], starting from the row number index differencemax_index where differencemax is located, traversing upwards to find the row number corresponding to the first element less than difference up, recorded as up_index. The upper boundary expansion row number x up is the difference between differencemax_index and up_index, i.e.: x up = differencemax_index - up_index.

[0114] The determination of the lower boundary expansion row number y down comprises the steps of setting the lower boundary coefficient as c down, calculating the lower boundary value: difference down = differencemax * c down. In the difference value sequence [difference1, difference2, …, differencem], starting from the row number index differencemax_index where differencemax is located, traversing downwards to find the row number corresponding to the first element less than difference down, recorded as down_index. The lower boundary expansion row number y down is the difference between down_index and differencemax_index, i.e.: y down = down_index - differencemax_index.

[0115] Further, the actual width of each defect in the sample set, the upper boundary coefficient c up, the lower boundary coefficient c down and the corresponding width characteristic value feature width are input into the deep neural network for training, and through iterative optimization, the upper boundary coefficient optimal_c_up and the lower boundary coefficient optimal_c_down in the sample set that make the width quantification accuracy highest are output.

[0116] For the defects in the test set, data processing is performed according to the above steps, wherein the optimal upper boundary coefficient optimal_c_up and the optimal lower boundary coefficient optimal_c_down obtained above are respectively used when calculating the upper boundary value difference_up and the lower boundary value difference_down, and then the width characteristic value of each defect in the test set is obtained, and the quantification of the defect width is completed in combination with actual requirements.

[0117] In this embodiment, the training of the deep neural network model realizes automatic optimization of the boundary coefficient, minimizes the width quantization error, and the application of the optimal coefficient improves the accuracy of the width quantization of the test set, so that the calculation of the width characteristic value is more accurate.

[0118] In some embodiments of the present application, the step S330 of extracting the depth characteristic value using the axial signal and the circumferential signal specifically includes:

[0119] S331: Calculate the integral value of each row of data of the axial signal and take the average as the first depth characteristic value;

[0120] S332: Extract the upper half of the data from the circumferential signal to form a first subset, and extract the data between the left valley index and the right valley index for each row of data of the first subset to reconstruct a second subset;

[0121] S333: Calculate the integral value of each row of data of the second subset and take the average as the second depth characteristic value; construct a linear combination model through the correlation coefficient, and predict the depth by the weighted sum of the first depth characteristic value and the second depth characteristic value to obtain the depth characteristic value.

[0122] For the defects in the sample, since the defect depth has the highest correlation with the peak and valley values of the axial signal and the circumferential signal, the axial signal y_data and the circumferential signal x_data of the defect are used when quantifying the defect depth.

[0123] For the axial signal, for each row mi (i = 1, 2, …, m) of data, the integral of the row is calculated to obtain an integral array {sa1, sa2, …, sam}. The average of the array is taken to obtain the integral average saave, which is the first depth characteristic value of the defect in the depth calculation, denoted as featureAs.

[0124] According to Figure 4As shown, for the circumferential signal x_data, according to its signal characteristics, the deeper the defect depth, the greater the peak and valley values of the circumferential signal. At the same time, the upper half of each curve of the circumferential signal presents a trend of "first increasing to a peak, then decreasing to a valley, and then increasing", and the lower half of the curve presents an opposite trend, that is, "first decreasing to a valley, then increasing to a peak, and then decreasing". Since the upper half can reflect the change of the depth, for the m rows of data, the data with the row index in the range of [0, m / 2] is selected as the research object, which is denoted as a first subset sub1.

[0125] For each row of data in the first subset sub1, two points with the smallest values are determined, and according to their morphological characteristics, the two minimum points correspond to the left valley point and the right valley point respectively, and the index of the left valley point is denoted as li, and the index of the right valley point is denoted as ri. For each row of data, only the data between the left valley index li and the right valley index ri is extracted, and a new subset is reconstructed by the above method, which is denoted as a second subset sub2.

[0126] For each row of data in the second subset sub2, the definite integral thereof is calculated to obtain an integral array {sc1, sc2, …, scm}. The average of the integral array is calculated to obtain an integral average scave, which is the second depth feature value of the defect in terms of depth, denoted as featureCs.

[0127] Finally, a linear combination model Depthi=x1featureAs+x2featureCs is constructed by using the correlation coefficients x1 and x2 to predict the depth by a weighted sum to obtain the depth feature value.

[0128] In this embodiment, the calculation of the axial signal integral average reflects the depth characteristics of the defect along the pipeline axis, and the reconstruction of the circumferential signal subset and the calculation of the integral average capture the depth distribution characteristics of the defect around the pipeline. Through the fusion of the axial signal integral average and the circumferential signal subset integral average, the multi-dimensional representation of the defect depth is realized. Moreover, the method adopts the construction of a linear combination model to realize the automatic allocation of feature weights through the optimization of correlation coefficients, thereby improving the accuracy of depth quantization.

[0129] In some embodiments of the present application, the optimization of the correlation coefficient specifically includes:

[0130] S3331: inputting the actual depth, the first depth feature value and the second depth feature value of each defect in the sample set into a depth neural network model;

[0131] S3332: determining the optimal correlation coefficient that minimizes the depth quantization error through training iteration, and satisfying the constraint condition that the weighted sum is 1;

[0132] S3333: In the depth quantification of the test set defects, the optimal correlation coefficient is directly used to calculate the depth feature value.

[0133] Specifically, in the depth feature optimization module, the actual depth of each defect in the sample set, the first depth feature value featureAs and the second depth feature value featureCs are input into the depth neural network model. The correlation coefficient of the first depth feature value is x1, and the correlation coefficient of the second depth feature value is x2. For each defect sample, there are:

[0134] Depthi=x1 featureAs+x2 featureCs, and x1+x2=1.

[0135] For each defect in the sample set, featureAs and featureCs are calculated respectively, and the actual depth of the defect, featureAs and featureCs are input into the depth neural network for training. Through iterative optimization, the specific values of the correlation coefficients x1 and x2 that make the quantization result optimal are determined.

[0136] According to the above steps, featureAs and featureCs are calculated, and the correlation coefficients x1 and x2 obtained by training are combined. In the depth quantification of the test set defects, the optimal correlation coefficient is directly used to calculate the depth feature value, and the quantization precision is improved.

[0137] The training of the depth neural network model in this embodiment realizes the automatic optimization of the correlation coefficient, minimizes the depth quantization error, and the weighted and constrained feature weight allocation is reasonable. The application of the optimal coefficient improves the accuracy of the depth quantification of the test set defects.

[0138] In some embodiments of the present application, the training of the depth neural network model adopts a sample set and test set separation mode. The sample set is used for iterative optimization of model parameters and coefficients, and the test set is used for verification of quantization precision.

[0139] It can be understood that in the model training phase, the sample set data is used for parameter optimization and key coefficient iteration of the depth neural network model. The predicted value is calculated by forward propagation, and the weight and coefficient are updated by back propagation, and the model performance is iteratively optimized. The test set data is independent of the training process, and is used to verify the generalization ability and quantization precision of the model. The separation design of the training set and the test set ensures the objectivity of the model evaluation and avoids the risk of overfitting.

[0140] The separation mode of the sample set and the test set ensures the independence of the model training and verification, improves the reliability of the model evaluation, and realizes the precise adjustment of the model parameters and coefficients through the iterative optimization process. The test set verification ensures the generalization ability of the quantization result.

[0141] The embodiment is aimed at the evaluation deviation problem caused by the overlap of validation data and training data in traditional model training. The independence of model training and validation is achieved through strict sample set and test set separation design. The iterative optimization process improves the fitting ability of the model to the sample set, and the test set validation guarantees the generalization ability of the model to unknown data.

[0142] In some embodiments of the present application, the extraction of pipeline magnetic flux leakage signals is based on pipeline magnetic flux leakage detection original files, ensuring the spatial alignment and consistency of signal data.

[0143] Specifically, in the signal extraction stage, pipeline magnetic flux leakage signal data is read from the pipeline magnetic flux leakage detection original file, and the axial signal, radial signal and circumferential signal are stored in the form of an m-row n-column matrix. Through synchronous acquisition technology and matrix alignment algorithm, the number of rows and columns of the three-axis signal matrix is strictly consistent, ensuring the consistency of the spatial position. In the signal extraction process, a filtering algorithm is used to eliminate noise interference and improve signal quality.

[0144] As can be known from the above embodiment content, the scheme realizes the whole process quantification of metal loss defects through systematic sample division, three-axis signal extraction, multi-dimensional feature extraction and deep neural network model. The separation design of sample set and test set guarantees the objectivity of model evaluation, and the spatial alignment processing of three-axis signals improves the reliability of feature extraction. Finally, the combination of multi-dimensional feature extraction and deep learning model realizes the accurate mapping from features to dimensions, and the data-driven optimization of key coefficients in the scheme improves the adaptability of the method to the change of defect morphology. The method meets the demand of high-precision defect detection in actual engineering, has strong engineering application value, effectively solves the problems of insufficient precision, weak anti-interference ability and single mapping relationship in traditional metal loss quantification methods, improves the precision and reliability of metal loss defect quantification, and provides a strong guarantee for the safe operation of pipelines.

[0145] Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the disclosure herein, with the sole proviso that the present application remain consistent with the principles of the present disclosure. The present application is intended to cover any variations, uses, or adaptations of the present disclosure that follow, in general, the principles of the present disclosure and include those that are known or customary practice in the art to which the present disclosure pertains. The specification and examples are to be regarded as exemplary only.

Claims

1. A method for quantifying metal loss in pipeline leakage magnetic field signals, characterized in that, The method includes: Sample sets and test sets are extracted from metal loss defect samples; Extracting magnetic flux leakage signals from pipes with metal loss defects, wherein the magnetic flux leakage signals include: axial signals, radial signals, and circumferential signals; Based on the axial signal, the left and right boundaries and peak positions of the defect are determined. Based on the radial signal, the column data subset defined by the left and right boundaries is divided into regions with the peak positions as the boundaries, and a comprehensive distance feature characterizing the extent of defect extension is calculated to obtain the length feature value. The peak and trough indices of each row of radial signal data are analyzed to determine the effective lateral subset of the defect. The peak and trough differences of each row of data in the effective subset are calculated and the location of the maximum difference is determined. The number of channels characterizing the defect width is calculated in combination with the boundary coefficient to obtain the width feature value. The integral mean of the axial signal is calculated as the first depth feature. A subset containing the magnetic field distortion information of the defect core is extracted from the circumferential signal and its integral mean is calculated as the second depth feature. The first depth feature and the second depth feature are weighted and fused to obtain the depth feature value. The length, width, and depth feature values ​​of the defects in the sample set, along with their actual dimensions, are used as training data and input into a deep neural network model for training, thereby establishing a nonlinear mapping relationship from feature values ​​to actual dimensions. The feature values ​​of the defects in the test set are input into the trained deep neural network model to invert and obtain the quantification results of the length, width and depth of the defects; The steps include: determining the left and right boundaries and peak positions of the defect based on the axial signal; dividing the region based on the column data subset defined by the radial signal at the left and right boundaries; and calculating a comprehensive distance feature characterizing the extent of defect extension to obtain a length feature value. The cumulative sum of each column of data is calculated based on the axial signal to determine the left and right boundaries and peak positions of the defect; Extract the column data subset defined by the left and right boundaries from the radial signal; The column data subset is divided into a left subset region and a right subset region using the peak position as the boundary, and the distance arrays of the left subset region and the right subset region are calculated respectively; The distance arrays of the left subset region and the right subset region are merged into a comprehensive distance array; The stability coefficient is optimized based on the difference between the actual length of the defects in the sample set and the comprehensive distance array, and the length feature value is selected using the optimized stability coefficient. This step includes: For each defect in the sample set, the ratio of the index corresponding to the value in the comprehensive distance array that is closest to the actual length to the total number of rows is calculated as the individual coefficient; The stability coefficient is obtained by averaging the individual coefficients of all defects in the sample set. In the length quantification of the defects in the test set, the length feature value is directly calculated using the stability coefficient; The analysis of the peak and trough indices of each row of radial signal data is used to determine the effective lateral subset of the defect. The difference between the peaks and troughs of each row of data in the effective subset is calculated, and the location of the maximum difference is determined. The number of channels representing the defect width is calculated using the boundary coefficient to obtain the width characteristic value. The steps include: For each row of data in the radial signal, extract the peak index and trough index; The maximum value of the peak indices of all rows is taken as the final peak index, and the minimum value of the trough indices of all rows is taken as the final trough index. Extract the effective subset of the radial signal based on the final peak index and the final trough index; Calculate the difference sequence between the peaks and troughs of each row of data in the effective subset, and determine the maximum difference and its corresponding row number; The step of calculating the number of expanded rows using the boundary coefficient and obtaining the width feature value by combining it with a fixed line spacing includes: The actual width, boundary coefficient, and corresponding width feature value of each defect in the sample set are input into the deep neural network model. The boundary coefficient includes the upper boundary coefficient and the lower boundary coefficient. The optimal upper and lower boundary coefficients that minimize the width quantization error are obtained by training iteratively. In the width quantization of the defects in the test set, the width feature value is calculated directly using the optimal upper boundary coefficient and the optimal lower boundary coefficient.

2. The method for quantifying metal loss in pipeline leakage magnetic field signals according to claim 1, characterized in that, The steps include: calculating the integral mean of the axial signal as a first depth feature; extracting a subset containing information about the magnetic field distortion of the defect core from the circumferential signal and calculating its integral mean as a second depth feature; and weighted fusing the first depth feature and the second depth feature to obtain a depth feature value. Calculate the integral value of each row of data in the axial signal and take the mean value, which is used as the first depth feature value; The upper half of the data is extracted from the circumferential signal to form a first subset. For each row of data in the first subset, the data between the left trough index and the right trough index is extracted to reconstruct a second subset. Calculate the integral value of each row of data in the second subset and take the mean, which is used as the second deep feature value; A linear combination model is constructed using correlation coefficients, and the depth is predicted by the weighted sum of the first depth feature value and the second depth feature value, thus obtaining the depth feature value.

3. The method for quantifying metal loss in pipeline leakage magnetic field signals according to claim 2, characterized in that, The step of constructing a linear combination model using correlation coefficients, and predicting the depth using a weighted sum of the first depth feature value and the second depth feature value, includes the following steps: The actual depth of each defect in the sample set, the first depth feature value, and the second depth feature value are input into the deep neural network model. The optimal correlation coefficient that minimizes the deep quantization error is determined through training iterations, while satisfying the constraint that the weighted sum is 1. In the depth quantization of defects in the test set, the optimal correlation coefficient is directly used to calculate the depth feature value.

4. The method for quantifying metal loss in pipeline leakage magnetic field signals according to claim 1, characterized in that, The sample set consists of three-quarters of the defect points randomly selected from the metal loss defect samples, and the test set consists of the remaining one-quarter of the defect points from the metal loss defect samples.

5. The method for quantifying metal loss in pipeline leakage magnetic field signals according to claim 1, characterized in that, The deep neural network model is trained using a sample set and a test set separation mode. The sample set is used for iterative optimization of model parameters and coefficients, while the test set is used to verify quantization accuracy.

6. The method for quantifying metal loss in pipeline leakage magnetic field signals according to claim 1, characterized in that, The extraction of the pipeline magnetic flux leakage signal is based on the original pipeline magnetic flux leakage detection file; The axial, radial, and circumferential signals are all represented in matrix form, and the pipeline leakage magnetic field signal corresponding to the same metal loss defect has the same number of rows and columns.

Images