Method for extracting and selecting features of quantifying defects in pipeline inspection

Abstract

The application provides a pipeline internal detection defect quantification feature extraction and feature selection probability determination method, comprising: obtaining the magnetic flux leakage internal detection data of a target pipeline, and constructing a pipeline magnetic flux leakage data matrix. According to a preset data amount, segmented magnetic flux leakage data is obtained. The axial magnetic flux leakage data, radial magnetic flux leakage data and circumferential magnetic flux leakage data in the segmented magnetic flux leakage data are marked for defects to obtain pipeline defect segment data. The pipeline defect segment data is denoised, and the data corresponding to the sensor channel with the largest peak-valley value difference is taken as the reference signal data for multi-layer discrete wavelet decomposition to obtain the approximate coefficient and the detail coefficient corresponding to each layer. The defect quantification features are extracted according to the denoised pipeline defect segment data and the approximate coefficient, and the selection probability of the defect quantification features is determined according to the detail coefficient. In this way, the importance of the defect quantification features can be accurately determined, and the accuracy and efficiency of the defect quantification algorithm are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipeline defect detection, and in particular to a pipeline internal detection defect quantification feature extraction and feature selection probability determination method. BACKGROUND

[0002] Pipeline transportation is widely used due to its low cost, high efficiency and good safety. However, due to the influence of factors such as long laying time, transportation medium and natural environment, some pipelines are prone to problems such as pipe wall damage, which has potential major hazards. Scientific and reasonable pipeline fault diagnosis is a key guarantee to ensure safe transportation of energy. As an important reference index for pipeline safety state evaluation, it is of great significance to realize the quantitative evaluation of pipeline defect information. At present, advanced methods mainly based on deep learning and machine learning are increasingly mature, providing a new direction for the development of pipeline fault diagnosis technology.

[0003] Using machine learning to realize defect size quantification first needs to accurately measure the geometric information such as length, width and depth of the pipeline defect to obtain defect data for the defect quantification algorithm to learn. Then, the signal data containing multi-directional and multi-scale defect information obtained by magnetic flux leakage internal detection is used to extract the features of multiple signals and match them with the measured values of the actual defects to realize the size prediction of new defects.

[0004] The preprocessing of the magnetic flux leakage internal detection signal data can improve the signal quality and thus improve the accuracy of the defect quantification algorithm. Improving the quality and diversity of the features can improve the generalization performance of the quantification model. In related technologies, threshold method can be used to denoise the magnetic flux leakage internal detection signal data. Generally, threshold method includes hard threshold method and soft threshold method. The hard threshold method may introduce artifacts or enhance the non-noise part in some signals when removing noise, which may cause the quality of the final signal to decrease or distort. The soft threshold method tends to introduce a smoothing effect near the signal change, which may cause the signal edge to become blurred and lose some detailed information. In this way, it is impossible to accurately extract the defect quantification features from the magnetic flux leakage internal detection signal data, which reduces the accuracy of the defect quantification algorithm. SUMMARY

[0005] The technical problem to be solved by the present application is that the defect quantification features cannot be accurately extracted from the magnetic flux leakage internal detection signal data, which reduces the accuracy of the defect quantification algorithm.

[0006] To solve the above technical problems, the present application provides a pipeline internal detection defect quantification feature extraction and feature selection probability determination method, which specifically adopts the following technical solutions:

[0007] The application provides a pipeline internal detection defect quantification feature extraction and feature selection probability determination method, which comprises the following steps: first, obtaining the magnetic flux leakage internal detection data of a target pipeline, and constructing a pipeline magnetic flux leakage data matrix according to the magnetic flux leakage internal detection data, wherein the magnetic flux leakage internal detection data comprises magnetic flux leakage internal detection signal data collected by multiple sensor channels. Then, the pipeline magnetic flux leakage data matrix is intercepted according to a preset interception data amount, and segmented magnetic flux leakage data is obtained, wherein the segmented magnetic flux leakage data comprises axial magnetic flux leakage data, radial magnetic flux leakage data and circumferential magnetic flux leakage data. The axial magnetic flux leakage data, the radial magnetic flux leakage data and the circumferential magnetic flux leakage data are labeled respectively, and pipeline defect segment data is obtained by interception. Secondly, the pipeline defect segment data is denoised to obtain denoised pipeline defect segment data; the peak-valley value difference corresponding to each sensor channel is determined according to the denoised pipeline defect segment data, and the pipeline defect segment data corresponding to the sensor channel with the largest peak-valley value difference is taken as reference signal data. Next, the reference signal data is subjected to multi-layer discrete wavelet decomposition according to a first decomposition parameter, and the first approximation coefficient and the first detail coefficient corresponding to each layer are obtained. Finally, defect quantification features are extracted according to the denoised pipeline defect segment data and the first approximation coefficient corresponding to each layer, and the selection probability of the defect quantification features is determined according to the first detail coefficient corresponding to each layer.

[0008] In an alternative implementation, the defect quantification features include one or more of the following features: peak value, valley value, peak-valley difference, peak-peak interval, length threshold, width threshold, surface energy and volume energy.

[0009] In an alternative implementation, the expression of the surface energy is as follows:

[0010]

[0011] wherein S represents the surface energy, X l represents the left valley point of the pipeline defect, X r represents the right valley point of the pipeline defect, F D represents the magnetic field intensity curve of the sensor, F rl represents the straight line formed by the left valley point and the right valley point.

[0012] In an alternative implementation, the selection probability of the defect quantification features is determined according to the first detail coefficient corresponding to each layer, which comprises the following steps: first, the curve inflection point corresponding to the first detail coefficient is determined according to the difference value of the first detail coefficient corresponding to each layer. Then, the selection probability of the defect quantification features is determined according to the curve inflection point.

[0013] In an alternative implementation, the defect quantification features include original signal features and decomposition features corresponding to each layer; wherein the expression of the selection probability of the decomposition features is as follows:

[0014] Pfn = a n × (1 - F kn );

[0015] wherein,

[0016]

[0017] In the formula, P fn represents the selection probability of the decomposition feature corresponding to the nth layer, n is an integer greater than or equal to 1, a n is the preset reference probability coefficient corresponding to the nth layer, F kn represents the oscillation frequency corresponding to the nth layer, F ni represents the number of curve inflection points corresponding to the nth layer, D n represents the length of the first detail coefficient corresponding to the nth layer. The selection probability expression of the original signal feature is:

[0018] P O = 1 - (P f1 + P f2 + … + P fn );

[0019] In the formula, P O represents the selection probability of the original signal feature, P f1 represents the selection probability of the decomposition feature corresponding to the first layer, P f2 represents the selection probability of the decomposition feature corresponding to the second layer, and P fn represents the selection probability of the decomposition feature corresponding to the nth layer.

[0020] In an alternative implementation, the pipeline defect segment data is denoised to obtain denoised pipeline defect segment data, specifically including: first, the reference signal data is subjected to multi-layer discrete wavelet decomposition according to the second decomposition parameter to obtain the second approximation coefficient and the second detail coefficient corresponding to each layer. Then, the second detail coefficient is denoised by a threshold method according to a preset threshold function to obtain denoised second detail coefficient. Finally, the denoised second detail coefficient and the second approximation coefficient are subjected to inverse transformation to obtain the denoised pipeline defect segment data.

[0021] In an alternative implementation, the expression of the denoised second detail coefficient is:

[0022]

[0023] wherein,

[0024]

[0025]

[0026] In the formula, y represents the second detail coefficient after noise reduction, x represents the second detail coefficient corresponding to each layer before noise reduction, represents a preset threshold function, a represents a preset variable parameter, y represents the average value of the second detail coefficient corresponding to each layer, N represents the length of the original signal data, and D n represents the length of the second detail coefficient corresponding to the nth layer.

[0027] In an alternative implementation, the first decomposition parameter includes the coefficient length of the first approximation coefficient and the first detail coefficient, and the first decomposition layer number. Then, the multi-layer discrete wavelet decomposition of the reference signal data according to the first decomposition parameter includes: performing the multi-layer discrete wavelet decomposition of the reference signal data according to the coefficient length and the first decomposition layer number through a wavelet base function.

[0028] In an alternative implementation, the expression of the coefficient length is:

[0029] L n = floor ((a + b * (n - 1)) * L n-1 );

[0030] In the formula, L n represents the coefficient length corresponding to the nth layer, floor represents the down rounding of the numerical value, and a and b are preset constants.

[0031] In an alternative implementation, the peak-valley value difference corresponding to each sensor channel is determined according to the pipeline defect segment data after noise reduction, which specifically includes: first, performing outlier rejection and missing value compensation on the pipeline defect segment data after noise reduction to obtain the pipeline defect segment data after data preprocessing. Then, the peak-valley value difference corresponding to each sensor channel is determined according to the pipeline defect segment data after data preprocessing.

[0032] In summary, by using the internal defect quantification feature extraction and feature selection probability determination method provided in the present application, the method uses the soft threshold denoising based on the discrete wavelet analysis method to realize the signal denoising of the defect feature region, improves the restoration degree of the magnetic flux leakage internal detection signal, reduces the influence of signal distortion caused by denoising, and thus can accurately determine the importance of the defect quantification feature. At the same time, based on the importance of the defect quantification feature, the method provides a method for determining the selection probability of the defect quantification feature when using the defect quantification algorithm, effectively improves the accuracy and efficiency of the defect quantification algorithm, and optimizes the inversion efficiency of the algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The flowchart of the pipeline internal detection defect quantification feature extraction and feature selection probability determination method provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0034] The embodiments will be described in detail below with reference to examples thereof as illustrated in the accompanying drawings. In the following description, like numbers refer to like elements throughout the description. The embodiments described in the following description are not meant to be inclusive of all embodiments consistent with the present application. Rather, the following description is presented by way of example only with reference to the accompanying drawings.

[0035] Pipeline transportation is widely used due to its low cost, high efficiency and good safety. However, due to the long laying time, transportation medium and natural environment and other factors, some pipelines are prone to problems such as pipe wall damage, which has potential major hazards. The leakage of oil pipelines not only brings huge economic losses to industrial production and people's life, but also causes serious damage to the environment and affects the health and safety of residents along the line. High-precision pipeline defect inversion helps to accurately evaluate the defect shape and provide quantitative analysis and data support for the operating state of the pipeline. This enables workers to identify serious defects in time and replace them, thereby reducing the major economic losses and environmental pollution that may be caused by leakage accidents. At the same time, for defects that are not serious and will not cause leakage, unnecessary pipeline replacement is avoided, and accurate prediction ability is provided for pipeline operation and maintenance.

[0036] Scientific and reasonable pipeline fault diagnosis is the key guarantee to ensure the safe transportation of energy. Defect size is an important reference index for evaluating the safety state of the pipeline, so it is of great significance to realize the quantitative evaluation of pipeline defect information. At present, advanced methods mainly based on deep learning and machine learning are increasingly mature, providing a new direction for the development of pipeline fault diagnosis technology.

[0037] Using machine learning to realize defect size quantification first needs to accurately measure the geometric information such as length, width and depth of the pipeline defect to obtain the defect data for the defect quantification algorithm learning. Then, the signal data containing multi-directional and multi-scale defect information obtained by magnetic flux leakage internal detection is used to extract the features of multiple signals and match them with the measured values of actual defects to realize the size prediction of new defects.

[0038] The preprocessing of the magnetic flux leakage internal detection signal data can improve the signal quality, and further improve the accuracy of the defect quantification algorithm. The quality and diversity of the features can improve the generalization performance of the quantification model. Since the pipeline is usually in a complex environment, there are many noise disturbances in the magnetic flux leakage internal detection signal. Among them, the wavelet denoising method is widely used in magnetic flux leakage internal detection signal denoising due to its multi-resolution analysis, energy concentration, adaptability, time-frequency localization and computational efficiency. The wavelet threshold denoising reduces the low-amplitude components in the signal by setting a threshold for low-amplitude signals, and then reconstructs the magnetic flux leakage internal detection signal. Generally, the threshold method is divided into hard threshold method and soft threshold method. The hard threshold method may introduce artifacts or enhance some non-noise parts in the signal when removing noise, which may cause the quality of the final signal to decrease or distort. The soft threshold method tends to introduce a smoothing effect near the signal change, which may cause the signal edge to become blurred and lose some detailed information. As a result, it is difficult to accurately extract defect quantification features from the magnetic flux leakage internal detection signal data, and the accuracy of the defect quantification algorithm is reduced.

[0039] To solve the above problems, the embodiment of the application provides a pipeline internal detection defect quantification feature extraction and feature selection probability determination method. The method uses a soft threshold denoising method based on discrete wavelet analysis to denoise the defect feature region signal, thereby improving the restoration degree of the magnetic flux leakage internal detection signal, reducing the influence of signal distortion caused by denoising, and accurately determining the importance of the defect quantification features. At the same time, based on the importance of the defect quantification features, a method for determining the selection probability of the defect quantification features when using the defect quantification algorithm is provided, which effectively improves the accuracy of the defect quantification algorithm and optimizes the inversion efficiency of the algorithm.

[0040] The scheme provided by the embodiment of the application will be described below with reference to the accompanying drawings.

[0041] Specifically, referring to Figure 1 , the flowchart of the internal detection defect quantification feature extraction and feature selection probability determination method provided by the embodiment of the application is shown in Figure 1 , the internal detection defect quantification feature extraction and feature selection probability determination method provided by the application comprises the following steps S101-S107:

[0042] S101, acquiring the magnetic flux leakage internal detection data of the target pipeline, and constructing a pipeline magnetic flux leakage data matrix according to the magnetic flux leakage internal detection data.

[0043] In the embodiment of the application, the magnetic flux leakage internal detection data of the target pipeline can be collected by a multi-sensor magnetic flux leakage internal detection device. That is, the magnetic flux leakage internal detection data includes magnetic flux leakage internal detection signal data collected by multiple sensor channels.

[0044] Further, the pipeline magnetic flux leakage data matrix can be constructed according to the magnetic flux leakage internal detection data. Specifically, the discrete sampling point magnetic flux leakage data collected by different channels of the sensor can be constructed into a column vector matrix in the following form:

[0045]

[0046] In the formula, D represents the pipeline magnetic flux leakage data matrix, x i，j represents the magnetic flux leakage signal data of the i-th discrete sampling point collected by the j-th sensor channel, i = 1, 2, …, m, j = 1, 2, …, n, n is the number of magnetic flux leakage sensor channels, and m is the number of discrete sampling points of each channel of the sensor.

[0047] S102, intercepting the pipeline magnetic flux leakage data matrix according to a preset interception data amount to obtain segmented magnetic flux leakage data.

[0048] Next, the pipeline magnetic flux leakage data matrix can be intercepted according to a preset interception data amount to obtain segmented magnetic flux leakage data for defect labeling. The segmented magnetic flux leakage data includes axial magnetic flux leakage data, radial magnetic flux leakage data, and circumferential magnetic flux leakage data.

[0049] The preset interception data amount can be preset according to the sampling mode of the equipment and the actual application requirement. For example, in the case of using the collection mode of the multi-sensor magnetic flux leakage internal detection equipment, 1000 discrete sampling points are taken as the preset interception data amount. For example, the sampling sensor equipment has a sampling mileage of 0.2 mm, and the data mileage of each segment is 0.2*1000 = 200 mm, that is, 20 cm is taken as a segmenting period. The same applies to the time sampling sensor equipment.

[0050] S103, respectively labeling defects in the axial magnetic flux leakage data, the radial magnetic flux leakage data, and the circumferential magnetic flux leakage data, and intercepting to obtain pipeline defect segment data.

[0051] Further, the axial magnetic flux leakage data, the radial magnetic flux leakage data, and the circumferential magnetic flux leakage data in the segmented magnetic flux leakage data intercepted in S102 can be labeled defects by using a unified defect labeling standard, and the defect data can be extracted from the magnetic flux leakage data to obtain the pipeline defect segment data.

[0052] In an implementation manner, the defect labeling standard can be preset according to prior knowledge. When labeling defects in the magnetic flux leakage data, the axial magnetic flux leakage data can be mainly labeled defects, and the radial magnetic flux leakage data and the circumferential magnetic flux leakage data can be labeled defects at the same mileage position and sensor serial number.

[0053] S104, denoising the pipeline defect segment data to obtain denoised pipeline defect segment data.

[0054] Further, the pipeline defect segment data obtained in S103 can be subjected to noise reduction processing to further improve the accuracy of determining the importance of the defect quantification features.

[0055] In S105, the peak-valley difference of each sensor channel is determined according to the pipeline defect segment data after noise reduction, and the pipeline defect segment data corresponding to the sensor channel with the largest peak-valley difference is taken as the reference signal data.

[0056] Specifically, among all the sensor channels, the pipeline defect segment data corresponding to the sensor channel with the largest peak-valley difference is taken as the reference signal data, which is used to extract the defect quantification features and determine the selection probability of the defect quantification features.

[0057] The peak-valley difference of the sensor channel can be determined by the following expression:

[0058] F fg = F max - F min ;

[0059] In the formula, F fg represents the peak-valley difference, F max represents the maximum value of the pipeline defect segment data after noise reduction corresponding to the sensor channel, and F min represents the minimum value of the pipeline defect segment data after noise reduction corresponding to the sensor channel.

[0060] In some embodiments, before determining the peak-valley difference of each sensor channel, the pipeline defect segment data after noise reduction can also be subjected to data preprocessing to eliminate abnormal values and compensate for missing values, thereby improving the accuracy of determining the importance of the defect quantification features. Then, in S105, the peak-valley difference of each sensor channel is determined according to the pipeline defect segment data after noise reduction, specifically including:

[0061] First, the pipeline defect segment data after noise reduction is subjected to abnormal value elimination and missing value compensation to obtain the pipeline defect segment data after data preprocessing.

[0062] Then, the peak-valley difference of each sensor channel is determined according to the pipeline defect segment data after data preprocessing.

[0063] Specifically, the abnormal values in the pipeline defect segment data after noise reduction can be determined by the following method:

[0064] F cal = abs(F i - F average );

[0065] The meaning expressed by the above formula is to calculate the absolute value of the difference between the maximum value Fi the average value F of all values on the sensor channel average , and then taking the absolute value F cal , to obtain a discrimination value F cal . If the F cal is greater than 1.5 times the F average , the data F i is judged as an abnormal value. Wherein 1.5 is a given reference value, for example, the multiple value can be preset according to the signal quality of the pipeline defect segment data after noise reduction and actual application requirements.

[0066] Further, the above abnormal value data F i may be processed as follows to eliminate abnormal values and replace abnormal values, i.e. missing value compensation:

[0067] F err = 0.5 * (F i-1 + F i+1 );

[0068] In the formula, F err represents the compensated data value, F i-1 and F i+1 represent the data on both sides of the data F i , i.e. the abnormal value can be replaced by the average value of the values on both sides.

[0069] In an implementation mode, if there are continuous abnormal values in one sensor channel, the adjacent sensor channel can be selected as the reference signal data.

[0070] S106, according to the first decomposition parameter, performing multi-layer discrete wavelet decomposition on the reference signal data to obtain first approximation coefficients and first detail coefficients corresponding to each layer.

[0071] In some embodiments, the above first decomposition parameter includes: a coefficient length of the first approximation coefficient and the first detail coefficient, and a first decomposition layer number. Then, according to the first decomposition parameter, performing multi-layer discrete wavelet decomposition on the reference signal data includes:

[0072] Performing multi-layer discrete wavelet decomposition on the reference signal data according to the coefficient length and the first decomposition layer number through a wavelet basis function.

[0073] In some embodiments, the expression of the coefficient length is:

[0074] L n = floor ((a + b * (n-1)) * L n-1 );

[0075] In the formula, L nwherein n represents the number of the layer, and n=0 represents the original reference signal data. floor represents the down rounding of the logarithm value, and a and b are preset constants. For example, a can be 0.6, and b can be 0.5.

[0076] In some embodiments, the wavelet base function described above can be a db4 wavelet, and the first decomposition layer number can be 3, for example.

[0077] In S107, the defect quantitative features are extracted according to the pipeline defect segment data after noise reduction and the first approximation coefficients corresponding to each layer, and the selection probability of the defect quantitative features is determined according to the first detail coefficients corresponding to each layer.

[0078] Finally, the defect quantitative features can be extracted according to the pipeline defect segment data after noise reduction and the first approximation coefficients corresponding to each layer. In some embodiments, the defect quantitative features can include one or more of the following features: peak value, valley value, peak-valley difference, peak-peak distance, length threshold, width threshold, surface energy, volume energy.

[0079] In an example, taking the axial magnetic flux leakage data as an example, the peak value is the maximum value in the pipeline defect segment data after noise reduction of the axial magnetic flux leakage data, the valley value is the minimum value in the pipeline defect segment data after noise reduction of the axial magnetic flux leakage data, and the peak-valley difference is the difference between the peak value and the valley value. For the pipeline defect segment data after noise reduction with two maximum value points, the peak-peak distance is calculated as the difference between the mileage corresponding to the two maximum value points. The length threshold is the number of rows of the pipeline defect segment data after noise reduction, i.e., the number of mileage direction points extracted. The width threshold is the number of columns of the pipeline defect segment data after noise reduction, i.e., the number of sensor channels included in the defect region.

[0080] In another example, the expression of the extracted surface energy is:

[0081]

[0082] wherein S represents the surface energy, X l represents the left valley point of the pipeline defect, X r represents the right valley point of the pipeline defect, F D represents the magnetic field intensity curve of the sensor, F rl represents the straight line formed by the left valley point and the right valley point.

[0083] In another example, the method of extracting the volume energy is: calculating the surface energy of the column vector with the largest peak-valley difference and the two column vectors on the left and right sides in the pipeline magnetic flux leakage data matrix and summing up the surface energies.

[0084] In the embodiments of the present application, the axial leakage magnetic data, the radial leakage magnetic data and the circumferential leakage magnetic data are subjected to wavelet analysis to obtain a plurality of first approximation coefficients, and the defect quantization features are extracted from the pipeline defect segment data after noise reduction and the first approximation coefficients of each layer. The number of original defect quantization features can be expanded by a corresponding multiple according to the number of wavelet decomposition layers.

[0085] In some embodiments, the random forest algorithm can be used to inverse the defect quantization features to inverse the geometric features such as length, width and depth of the pipeline defects. Specifically, the random forest algorithm is used to randomly select the defect quantization features for calculation, and then the size of the pipeline defects is derived.

[0086] In an implementation manner, a random forest model with feature selection weight can be used, that is, the importance of the defect quantization features is calculated by preprocessing the defect signal data, and the weight of the selection probability of the defect quantization features is determined according to the importance.

[0087] In some embodiments, the selection probability of the defect quantization features determined according to the first detail coefficient of each layer specifically includes:

[0088] First, the inflection point of the curve corresponding to the first detail coefficient of each layer is determined according to the difference value of the first detail coefficient of each layer.

[0089] Specifically, the calculation method of the inflection point corresponding to the first detail coefficient of each layer is as follows: the difference sequence of the first detail coefficient is calculated, that is, the difference between the subsequent value and the previous value in the first detail coefficient is calculated. Then, the inflection point can be determined by the following method:

[0090] M i =(C i-1 ×C i ) (i≥2);

[0091] In the formula, M i is the inflection point judgment result, and C i represents the value in the difference sequence of the first detail coefficient. Specifically, the above formula is the product of each value and the previous value in the difference sequence of the first detail coefficient starting from the second value. If the result M i is negative, it is determined that the point is an inflection point of the first detail coefficient.

[0092] Then, the selection probability of the defect quantization features is determined according to the inflection point of the curve.

[0093] In some embodiments, the defect quantization features can include original signal features and decomposition features corresponding to each layer. The selection probability expression of the decomposition features is:

[0094] P fn =a n ×(1-Fkn );

[0095] wherein,

[0096]

[0097] in the formula, P fn represents the selection probability of the decomposition feature corresponding to the nth layer, n is an integer greater than or equal to 1, a n is the preset reference probability coefficient corresponding to the nth layer, F kn represents the oscillation frequency corresponding to the nth layer, F ni represents the number of curve inflection points corresponding to the nth layer, D n represents the length of the first detail coefficient corresponding to the nth layer.

[0098] The selection probability expression of the original signal feature is:

[0099] P O = 1-(P f1 + P f2 + … + P fn );

[0100] in the formula, P O represents the selection probability of the original signal feature, P f1 represents the selection probability of the decomposition feature corresponding to the first layer, P f2 represents the selection probability of the decomposition feature corresponding to the second layer, and P fn represents the selection probability of the decomposition feature corresponding to the nth layer.

[0101] For example, taking the first decomposition layer number as 3 layers as an example, the defect quantification feature can include: the original signal feature, the one-layer decomposition feature, the two-layer decomposition feature and the three-layer decomposition feature.

[0102] For example, the probability of the one-layer decomposition feature can be represented as:

[0103] P f1 = a1x (1-F k1 ).

[0104] in the formula, P f1 represents the selection probability of the decomposition feature corresponding to the first layer (i.e. the one-layer decomposition feature), a1 is the preset reference probability coefficient corresponding to the first layer, for example, a1 can be 0.2.

[0105] The original signal feature selection probability can be represented as:

[0106] P O = 1-(P f1 + P f2 + P f3 ).

[0107] In some embodiments, in S104, the pipeline defect segment data can be denoised by threshold method. Specifically, the pipeline defect segment data can be decomposed by wavelet to obtain the detail coefficients and the approximation coefficients of each layer corresponding to the pipeline defect segment data. Then, a threshold function is set for the detail coefficients of each layer, and the denoised detail coefficients are obtained according to the threshold function. The denoised detail coefficients and the approximation coefficients are inverse transformed to obtain the denoised pipeline defect segment data. S104 can specifically include:

[0108] S1041, performing multi-layer discrete wavelet decomposition on the reference signal data according to a second decomposition parameter to obtain second approximation coefficients and second detail coefficients corresponding to each layer.

[0109] S1042, denoising the second detail coefficients by threshold method according to a preset threshold function to obtain denoised second detail coefficients.

[0110] S1043, inverse transforming the denoised second detail coefficients and the second approximation coefficients to obtain denoised pipeline defect segment data.

[0111] In some embodiments, in the denoising of the pipeline defect segment data by threshold method, the expression of the denoised second detail coefficients is:

[0112]

[0113] wherein,

[0114]

[0115]

[0116] wherein, y represents the denoised second detail coefficients, x represents the second detail coefficients of each layer before denoising, represents the preset threshold function, a represents the preset variable parameter, y represents the average value of the second detail coefficients of each layer, N represents the length of the original signal data, and D n represents the length of the second detail coefficients of the nth layer.

[0117] In summary, by using the internal defect quantitative feature extraction and feature selection probability determination method provided in the above embodiments of the present application, the method uses a soft threshold denoising method based on discrete wavelet analysis to denoise the defect feature region signal, improves the restoration degree of the magnetic flux leakage internal detection signal, reduces the influence of signal distortion caused by denoising, and thus can accurately determine the importance of the defect quantitative feature. At the same time, based on the importance of the defect quantitative feature, a method for determining the selection probability of the defect quantitative feature when using the defect quantitative algorithm is provided, which effectively improves the accuracy and efficiency of the defect quantitative algorithm and optimizes the inversion efficiency of the algorithm.

[0118] In the description of the present application, it should be understood that the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0119] In the present application, unless otherwise specifically defined and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise specifically limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0120] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of different embodiments or examples without contradiction.

[0121] The similar parts among the embodiments provided in the application can be referred to each other, the specific embodiments provided above are only several examples under the general concept of the application, and do not constitute the limitation of the protection scope of the application. Any other embodiments extended according to the application scheme without creative labor for those skilled in the art shall fall within the protection scope of the application.

Claims

1. A method for quantitative feature extraction and feature selection probability determination of defects detected inside pipelines, characterized in that, include: The internal magnetic flux leakage detection data of the target pipeline is acquired, and a pipeline magnetic flux leakage data matrix is ​​constructed based on the internal magnetic flux leakage detection data. The internal magnetic flux leakage detection data includes internal magnetic flux leakage detection signal data collected by multiple sensor channels. The pipeline magnetic flux leakage data matrix is ​​extracted according to a preset data extraction amount to obtain segmented magnetic flux leakage data, which includes: axial magnetic flux leakage data, radial magnetic flux leakage data and circumferential magnetic flux leakage data; Defects were labeled on the axial magnetic flux leakage data, the radial magnetic flux leakage data, and the circumferential magnetic flux leakage data respectively, and the pipeline defect segment data was extracted. The pipeline defect segment data is denoised to obtain denoised pipeline defect segment data; the peak-valley difference corresponding to the multiple sensor channels is determined based on the denoised pipeline defect segment data, and the pipeline defect segment data corresponding to the sensor channel with the largest peak-valley difference is used as the reference signal data. The reference signal data is subjected to multi-level discrete wavelet decomposition based on the first decomposition parameters to obtain the first approximation coefficient and the first detail coefficient corresponding to each level. Defect quantization features are extracted based on the denoised pipeline defect fragment data and the first approximation coefficient corresponding to each layer, and the selection probability of the defect quantization features is determined based on the first detail coefficient corresponding to each layer.

2. The method according to claim 1, characterized in that, The defect quantification features include one or more of the following features: Peak value, valley value, peak-to-valley difference, peak-to-peak spacing, length threshold, width threshold, surface energy, volume energy.

3. The method according to claim 2, characterized in that, The expression for the surface energy is: Where S represents surface energy, X l X represents the left valley point of the pipeline defect. r F is the valley point on the right side of the pipeline defect. D The magnetic field strength curve of the sensor, F rl The line formed by connecting the valley point on the left and the valley point on the right.

4. The method according to any one of claims 1-3, characterized in that, The step of determining the selection probability of the defect quantification feature based on the first detail coefficient corresponding to each layer includes: The curve inflection point corresponding to the first detail coefficient is determined based on the difference value of the first detail coefficient corresponding to each layer; The selection probability of the defect quantification feature is determined based on the inflection point of the curve.

5. The method according to claim 4, characterized in that, The defect quantification features include: original signal features and decomposition features corresponding to each layer; wherein, the probability expression for selecting the decomposition features is: P fn =a n ×(1-F kn ); in, In the formula, P fn This represents the probability of selecting the decomposition feature corresponding to the nth layer, where n is an integer greater than or equal to 1, and a n F is the preset reference probability coefficient corresponding to the nth layer. kn F represents the oscillation frequency corresponding to the nth layer. ni D represents the number of inflection points of the curve corresponding to the nth layer. n This represents the length of the first detail coefficient corresponding to the nth layer; The probability expression for selecting the original signal features is as follows: P O =1-(P f1 +P f2 +…+P fn ); In the formula, P O P represents the probability of selecting features from the original signal. f1 P represents the probability of selecting the decomposition feature corresponding to the first layer. f2 P represents the probability of selecting the decomposition feature corresponding to the second layer. fn This represents the probability of selecting the decomposition feature corresponding to the nth layer.

6. The method according to claim 1, characterized in that, The process of denoising the pipeline defect segment data to obtain denoised pipeline defect segment data specifically includes: The reference signal data is subjected to multi-level discrete wavelet decomposition based on the second decomposition parameters to obtain the second approximation coefficient and the second detail coefficient corresponding to each level. The second detail coefficients are denoised using a thresholding method based on a preset threshold function to obtain the denoised second detail coefficients. The second detail coefficients and the second approximation coefficients after noise reduction are inversely transformed to obtain the noise-reduced pipeline defect segment data.

7. The method according to claim 6, characterized in that, The expression for the second detail coefficient after noise reduction is: in, In the formula, y represents the second detail coefficient after denoising, and x represents the second detail coefficient corresponding to each layer before denoising. α represents the preset threshold function, γ represents the preset variable parameter, γ represents the average value of the second detail coefficients corresponding to each layer, N represents the length of the original signal data, and D represents the value of ... n This represents the length of the second detail coefficient corresponding to the nth layer.

8. The method according to claim 1, characterized in that, The first decomposition parameters include: the coefficient lengths of the first approximation coefficient and the first detail coefficient, and the number of the first decomposition layers; The step of performing multi-level discrete wavelet decomposition on the reference signal data according to the first decomposition parameters includes: The reference signal data is subjected to multi-level discrete wavelet decomposition using wavelet basis functions based on the coefficient length and the first decomposition level.

9. The method according to claim 8, characterized in that, The expression for the length of the coefficient is: L n =floor((a+b*(n-1))*L n-1 ); In the formula, L n This indicates the coefficient length corresponding to the nth layer, floor indicates rounding down the value, and a and b are preset constants.

10. The method according to claim 1, characterized in that, The step of determining the peak-valley difference corresponding to each of the multiple sensor channels based on the noise-reduced pipeline defect segment data specifically includes: The denoised pipeline defect segment data is subjected to outlier removal and missing value compensation to obtain preprocessed pipeline defect segment data. The peak-valley difference corresponding to each of the multiple sensor channels is determined based on the preprocessed pipeline defect segment data.

Images