Distribution line insulation defect type identification method and device, terminal equipment and storage medium

Abstract

The invention discloses a distribution line insulation defect type identification method and device, terminal equipment and a storage medium. The method comprises the following steps: firstly, obtaining a detection signal of a distribution line insulation part; then decomposing structural feature data and texture feature data from the detection signal; constructing a feature matrix according to the structural feature data and the texture feature data; performing feature weighting operation on the feature matrix according to a preset attention weight distribution mechanism and the feature matrix to obtain a feature matrix after feature weighting; inputting the feature matrix subjected to feature weighting into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs defect type probability distribution of the insulation part of the distribution line according to input data; and finally, determining the defect type of the insulation part of the distribution line according to the defect type probability distribution. According to the invention, the identification efficiency of the insulation defect type of the distribution line can be improved.

Description

Technical Field

[0001] The present invention relates to the field of distribution line defect detection, and particularly to a method, device, terminal device and storage medium for identifying insulation defect types of distribution lines. Background Art

[0002] The insulation performance of distribution lines is directly related to the safe and stable operation of the power system. Traditional insulation defect detection mainly relies on manual inspections. Technicians obtain detection data such as electromagnetic signals and acoustic signals of the insulated parts by carrying equipment such as infrared thermometers and ultrasonic detectors, and then identify defects by visually observing waveform maps. However, this method of manually detecting defects has the drawbacks of low detection efficiency and detection accuracy depending on experience. Manual inspections need to check the lines section by section, and it takes hours or even days to detect a single distribution line, which is difficult to meet the real-time monitoring requirements of smart grids. Summary of the Invention

[0003] The present invention provides a method, device, terminal device and storage medium for identifying insulation defect types of distribution lines, and the method effectively improves the identification efficiency of insulation defect types of distribution lines.

[0004] An embodiment of the present invention provides a method for identifying insulation defect types of distribution lines, including:

[0005] Obtaining detection signals of the insulated parts of the distribution line;

[0006] Decomposing structural feature data and texture feature data from the detection signals;

[0007] Constructing a feature matrix according to the structural feature data and the texture feature data;

[0008] Performing a feature weighting operation on the feature matrix according to a preset attention weight distribution mechanism and the feature matrix to obtain a feature-weighted feature matrix;

[0009] Inputting the feature-weighted feature matrix into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs the probability distribution of the defect types of the insulated parts of the distribution line according to the input data;

[0010] Determining the defect types of the insulated parts of the distribution line according to the probability distribution of the defect types.

[0011] Further, the detection signals include time-domain electromagnetic detection signals and ultrasonic detection signals;

[0012] The decomposing of the structural feature data and the texture feature data from the detection signals includes:

[0013] Perform wavelet transform on the time-domain electromagnetic detection signal to obtain electromagnetic characteristic coefficients;

[0014] Perform wavelet transform on the ultrasonic detection signal to obtain ultrasonic characteristic coefficients;

[0015] Calculate the local energy distribution based on the electromagnetic characteristic coefficients and the ultrasonic characteristic coefficients;

[0016] Perform band decomposition on the electromagnetic characteristic coefficients to obtain high-frequency structure characteristic coefficients;

[0017] Construct an energy matrix based on the high-frequency structure characteristic coefficients and perform principal component analysis on the energy matrix to obtain principal eigenvectors;

[0018] Perform band decomposition on the ultrasonic characteristic coefficients to obtain low-frequency texture characteristic coefficients;

[0019] Construct a texture matrix based on the low-frequency texture characteristic coefficients and decompose the texture matrix to obtain eigenbasis vectors;

[0020] Perform orthogonal transformation on the principal eigenvectors and the eigenbasis vectors to obtain an orthogonal feature space containing defect feature information;

[0021] Calculate the structure feature data and the texture feature data based on the orthogonal feature space, a preset feature boundary frequency, the high-frequency structure characteristic coefficients, and the low-frequency texture characteristic coefficients.

[0022] Further, the calculating the structure feature data and the texture feature data based on the orthogonal feature space, a preset feature boundary frequency, the high-frequency structure characteristic coefficients, and the low-frequency texture characteristic coefficients includes:

[0023] Form complementary feature pairs by combining high-frequency structure characteristic coefficients and low-frequency texture characteristic coefficients that meet the preset complementary judgment conditions;

[0024] Map the complementary feature pairs to multi-scale feature representations in the orthogonal feature space;

[0025] Construct a feature description matrix based on the multi-scale feature representations and divide the feature description matrix into the structure feature data and the texture feature data according to the feature boundary frequency.

[0026] Further, the forming complementary feature pairs by combining high-frequency structure characteristic coefficients and low-frequency texture characteristic coefficients that meet the preset complementary judgment conditions includes:

[0027] Calculate the overlapping interval in the frequency domain between the high-frequency structure characteristic coefficients and the low-frequency texture characteristic coefficients;

[0028] Calculate the ratio between the overlapping interval and a preset total bandwidth to obtain a band overlap degree;

[0029] Calculate the cross-correlation coefficient between the high-frequency structure feature coefficient and the low-frequency texture feature coefficient;

[0030] Judge whether the band overlap degree is less than a preset first band threshold, and judge whether the cross-correlation coefficient is less than a preset second band threshold.

[0031] If both judgment results are yes, determine that the high-frequency structure feature coefficient and the low-frequency texture feature coefficient meet the complementary judgment condition, and form a complementary feature pair with the high-frequency structure feature coefficient and the low-frequency texture feature coefficient.

[0032] Otherwise, determine that the high-frequency structure feature coefficient and the low-frequency texture feature coefficient do not meet the complementary judgment condition.

[0033] Further, the performing a feature weighting operation on the feature matrix according to a preset attention weight allocation mechanism and the feature matrix to obtain a feature-weighted feature matrix includes:

[0034] Decompose the feature matrix with the spatial resolution as the decomposition reference standard to obtain a feature sequence; wherein, the feature sequence contains scale features with different spatial resolutions.

[0035] According to the attention weight allocation mechanism, assign corresponding attention weights to each scale feature in the feature sequence to obtain a feature-weighted feature sequence.

[0036] Fuse the feature-weighted feature sequence to obtain a feature-weighted feature matrix.

[0037] Further, the process of constructing the defect diagnosis network includes:

[0038] Obtain a distribution network fault sample set; wherein, the distribution network fault sample set includes: a number of distribution lines with defects and corresponding true defect types.

[0039] Construct a double-branch collaborative diagnosis network, use the distribution network fault sample set as the input, and use the predicted defect type probability distribution as the output, and iteratively train the double-branch collaborative diagnosis network until the loss function value of the double-branch collaborative diagnosis network converges.

[0040] Use the trained double-branch collaborative diagnosis network as the defect diagnosis network.

[0041] During each training process, the probability distribution of the defect types predicted by the current dual-branch collaborative diagnosis network is compared with the corresponding true defect types in the power distribution line fault sample set, the loss function value is calculated according to the comparison result, and the network parameters of the dual-branch collaborative diagnosis network are adjusted according to the calculated loss function value.

[0042] An embodiment of the present invention further provides a device for identifying the insulation defect types of a power distribution line, including: a signal acquisition module, a signal decomposition module, a feature matrix construction module, a feature weighting module, a probability distribution calculation module, and a defect type determination module;

[0043] The signal acquisition module is used to acquire the detection signal of the insulated part of the power distribution line;

[0044] The signal decomposition module is used to decompose the structural feature data and the texture feature data from the detection signal;

[0045] The feature matrix construction module is used to construct a feature matrix according to the structural feature data and the texture feature data;

[0046] The feature weighting module is used to perform a feature weighting operation on the feature matrix according to a preset attention weight distribution mechanism and the feature matrix to obtain a feature-weighted feature matrix;

[0047] The probability distribution calculation module is used to input the feature-weighted feature matrix into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs the probability distribution of the defect types of the insulated part of the power distribution line according to the input data;

[0048] The defect type determination module is used to determine the defect type of the insulated part of the power distribution line according to the probability distribution of the defect types.

[0049] Further, the construction process of the defect diagnosis network includes:

[0050] Obtain a power distribution line fault sample set; wherein, the power distribution line fault sample set includes: a number of power distribution lines with defects and corresponding true defect types;

[0051] Construct a dual-branch collaborative diagnosis network, use the power distribution line fault sample set as the input, and use the predicted probability distribution of the defect types as the output, and perform iterative training on the dual-branch collaborative diagnosis network until the loss function value of the dual-branch collaborative diagnosis network converges;

[0052] Use the trained dual-branch collaborative diagnosis network as the defect diagnosis network;

[0053] Among them, in each training process, the probability distribution of the defect types predicted by the current dual-branch collaborative diagnosis network is compared with the corresponding true defect types in the power distribution line fault sample set, the loss function value is calculated according to the comparison result, and the network parameters of the dual-branch collaborative diagnosis network are adjusted according to the calculated loss function value.

[0054] This application also provides a terminal device, including:

[0055] One or more processors;

[0056] A memory, coupled to the processor, for storing one or more programs;

[0057] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for identifying the insulation defect types of the power distribution line as described in the above-mentioned invention embodiment.

[0058] This application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for identifying the insulation defect types of the power distribution line as described in the above-mentioned invention embodiment is implemented.

[0059] By implementing the present invention, the following beneficial effects are achieved:

[0060] The present invention provides a method, apparatus, terminal device and storage medium for identifying insulation defect types of a distribution line. The method first obtains a detection signal of an insulated part of the distribution line; decomposes the detection signal into structural feature data and texture feature data; constructs a feature matrix according to the structural feature data and the texture feature data; performs a feature weighting operation on the feature matrix according to a preset attention weight distribution mechanism and the feature matrix to obtain a feature-weighted feature matrix; inputs the feature-weighted feature matrix into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs a probability distribution of the defect types of the insulated part of the distribution line according to the input data; determines the defect type of the insulated part of the distribution line according to the probability distribution of the defect types. Thus, the present application abandons the method of manually checking the defects of the line section by section. Through an automated signal processing process, it can quickly analyze the detection signal of the insulated part of the distribution line. From obtaining the detection signal to decomposing the feature data, constructing and processing the feature matrix, and then using the defect diagnosis network to output the result, the whole process can be completed in a short time, greatly shortening the detection time of a single distribution line, meeting the requirements of real-time monitoring of the smart grid, being able to detect potential insulation defects in time, and providing a strong guarantee for the stable operation of the power system. In addition, by accurately decomposing the structural feature data and texture feature data from the detection signal, constructing a feature matrix for analysis, more comprehensive defect information can be mined, and finally, based on the defect diagnosis network, a probability distribution of the defect types is output, so as to more accurately determine the defect type, effectively avoiding the problems of missed detection and misdetection caused by insufficient manual experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the present application, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0062] Figure 1 is a schematic flowchart of a method for identifying insulation defect types of a distribution line provided by an embodiment of the present application;

[0063] Figure 2 is a schematic structural diagram of a device for identifying insulation defect types of a distribution line provided by an embodiment of the present application;

[0064] Figure 3 is a schematic structural diagram of a terminal device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following will clearly and completely describe the technical solutions in this application in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above description of the drawings are intended to cover non-exclusive inclusion.

[0067] In the description of the embodiments of this application, technical terms such as "first" and "second" are only used to distinguish different objects and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity, specific order, or primary-secondary relationship of the indicated technical features. In the description of the embodiments of this application, the meaning of "a plurality of" is more than two, unless otherwise specifically defined.

[0068] Referring to "embodiments" herein means that the specific features, structures, or characteristics described in connection with the embodiments may be included in at least one embodiment of this application. The phrase appears in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein may be combined with other embodiments.

[0069] In the description of the embodiments of this application, the term "and / or" is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this text generally represents an "or" relationship between the associated objects before and after.

[0070] In the description of the embodiments of this application, the term "a plurality of" refers to more than two (including two). Similarly, "a plurality of groups" refers to more than two groups (including two groups), and "a plurality of pieces" refers to more than two pieces (including two pieces).

[0071] In the description of the embodiments of the present application, unless otherwise clearly specified and limited, technical terms such as "installation", "connection", "connection", "fixation", etc. should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can also be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and can be the communication inside two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific circumstances.

[0072] See Figure 1 , which is a schematic flow chart of a method for identifying insulation defect types of a distribution line provided by an embodiment of the present invention, including:

[0073] S1. Obtain the detection signal of the insulated part of the distribution line;

[0074] Specifically, an electromagnetic ultrasonic probe is used to scan the insulated part of the distribution line to obtain the detection signal.

[0075] S2. Decompose the structural feature data and texture feature data from the detection signal;

[0076] In a preferred embodiment, the detection signal includes a time-domain electromagnetic detection signal and an ultrasonic detection signal;

[0077] The decomposing of the structural feature data and texture feature data from the detection signal includes:

[0078] Perform wavelet transform on the time-domain electromagnetic detection signal to obtain electromagnetic feature coefficients;

[0079] Perform wavelet transform on the ultrasonic detection signal to obtain ultrasonic feature coefficients;

[0080] Calculate the local energy distribution according to the electromagnetic feature coefficients and the ultrasonic feature coefficients;

[0081] Perform band decomposition on the electromagnetic feature coefficients to obtain high-frequency structural feature coefficients;

[0082] Construct an energy matrix according to the high-frequency structural feature coefficients, and perform principal component analysis on the energy matrix to obtain the principal eigenvector;

[0083] Perform band decomposition on the ultrasonic feature coefficients to obtain low-frequency texture feature coefficients;

[0084] Construct a texture matrix according to the low-frequency texture feature coefficients, and decompose the texture matrix to obtain the eigenbasis vector;

[0085] Perform an orthogonal transformation on the main eigenvector and the eigenbasis vector to obtain an orthogonal feature space containing defect feature information;

[0086] Based on the orthogonal feature space, a preset feature demarcation frequency, the high-frequency structure feature coefficient, and the low-frequency texture feature coefficient, calculate the structure feature data and the texture feature data;

[0087] Schematically, an electromagnetic ultrasonic probe includes an electromagnetic excitation unit and an ultrasonic receiving unit; the electromagnetic excitation unit and the ultrasonic receiving unit of the electromagnetic ultrasonic probe are arranged in a circular array, and the excitation frequency of the electromagnetic excitation unit and the sampling frequency of the ultrasonic receiving unit satisfy a cooperative matching relationship;

[0088] Specifically, the time-domain electromagnetic detection signal s em (t) is represented by the following formula:

[0089] s em (t) = A e (t)cos(2πf e t + φ e (t)) + n e (t);

[0090] The ultrasonic detection signal s us (t) is represented by the following formula:

[0091] s us (t) = A u (t)cos(2πf u t + φ u (t)) + n u (t);

[0092] Among them, S em (t) is the time-domain electromagnetic detection signal, representing the electromagnetic waveform received by the electromagnetic ultrasonic probe at time t; s us (t) is the ultrasonic detection signal, representing the ultrasonic waveform received by the electromagnetic ultrasonic probe at time t; A e (t) is the instantaneous amplitude of the time-domain electromagnetic detection signal, reflecting the energy change during the propagation of the electromagnetic wave; A u (t) is the instantaneous amplitude of the ultrasonic detection signal, reflecting the energy change during the propagation of the ultrasonic wave; f e is the operating frequency of the electromagnetic excitation unit; f u is the operating frequency of the ultrasonic receiving unit; φ e (t) is the instantaneous phase of the time-domain electromagnetic detection signal, reflecting the phase change during the propagation of the electromagnetic wave; φ u (t) is the instantaneous phase of the ultrasonic detection signal, reflecting the phase change during the propagation of the ultrasonic wave; ne (t) is the environmental noise of the electromagnetic channel, including electromagnetic interference and system noise; n u (t) is the environmental noise of the ultrasonic channel, including mechanical vibration and acoustic interference;

[0093] In the present invention, the electromagnetic ultrasonic probe is arranged in a circular array. The special design of the electromagnetic excitation unit and the ultrasonic receiving unit enables the electromagnetic wave and the ultrasonic wave to form a dual-mode excitation-reception mechanism at the insulated part of the power distribution line. A frequency complementary relationship is formed between the high-frequency electromagnetic field generated by the electromagnetic excitation unit and the ultrasonic response frequency of the ultrasonic receiving unit. This design is different from the traditional single-frequency detection method. Experiments show that when the electromagnetic wave frequency is in the MHz order of magnitude and the ultrasonic wave frequency is in the kHz order of magnitude, they can respectively correspond to the surface and internal defect characteristics of the insulated part, thus realizing the collaborative detection of macroscopic and microscopic defects.

[0094] Schematically, perform wavelet transform on the time-domain electromagnetic detection signal to obtain an electromagnetic characteristic coefficient, perform wavelet transform on the ultrasonic detection signal to obtain an ultrasonic characteristic coefficient, and calculate the local energy distribution according to the electromagnetic characteristic coefficient and the ultrasonic characteristic coefficient;

[0095] Specifically, perform wavelet transform on the time-domain electromagnetic detection signal through the following formula:

[0096]

[0097] Specifically, perform wavelet transform on the ultrasonic detection signal through the following formula:

[0098]

[0099] Specifically, calculate the local energy distribution through the following formula:

[0100] E(a,b) = α|W em (a,b)| 2 +β|W us (a,b)| 2 ;

[0101] Among them, W em (a, b) is the electromagnetic characteristic coefficient, obtained by performing wavelet transform on the time-domain electromagnetic detection signal, representing the time-frequency characteristics at the scale parameter a and the translation parameter b; W us(a, b) is the ultrasonic characteristic coefficient, which is obtained by performing wavelet transform on the ultrasonic detection signal, representing the time-frequency characteristics at the scale parameter a (the scale parameter of wavelet transform, controlling the stretching of the wavelet basis function and determining the frequency resolution) and the translation parameter b (the translation parameter of wavelet transform, controlling the translation of the wavelet basis function and determining the time resolution); ψ is the wavelet basis function, and ψ* represents its conjugate complex number, which is used for the multi-resolution analysis of the signal; α is the weight coefficient of electromagnetic energy, used to balance the contributions of electromagnetic and ultrasonic signals; β is the weight coefficient of electromagnetic energy, used to balance the contributions of electromagnetic and ultrasonic signals. It refers to the translation and scale transformation of the mother wavelet on the time axis in wavelet transform; E(a, b) is the local energy distribution on the time-frequency plane, which is used for feature extraction and defect recognition.

[0102] In the present invention, the adaptive processing of the bimodal signal is realized through wavelet transform. Compared with the traditional Fourier transform, wavelet transform can simultaneously obtain the time-domain and frequency-domain information of the signal, and is suitable for processing non-stationary signals. In practical applications, the electromagnetic characteristic coefficient reflects the change of dielectric characteristics on the surface of the insulator, while the ultrasonic characteristic coefficient corresponds to the change of acoustic impedance inside the material. This collaborative processing mechanism of the bimodal signal significantly improves the accuracy of defect detection.

[0103] Schematically, the electromagnetic characteristic coefficient is decomposed into high-frequency structure characteristic coefficients by frequency band decomposition, and the ultrasonic characteristic coefficient is decomposed into low-frequency texture characteristic coefficients by frequency band decomposition.

[0104] Specifically, the electromagnetic characteristic coefficient is decomposed into frequency bands by the following formula:

[0105]

[0106] Specifically, the ultrasonic characteristic coefficient is decomposed into frequency bands by the following formula:

[0107]

[0108] In the formula, C h (k) is the high-frequency structure characteristic coefficient of the kth frequency band, reflecting the surface defect characteristics of the insulator; C l (k) is the low-frequency texture characteristic coefficient of the kth frequency band, reflecting the internal defect characteristics of the insulator; f h is the central frequency of high-frequency analysis, used for surface defect detection; f l is the central frequency of low-frequency analysis, used for internal defect detection; δ is the Dirac δ function; a j is the jth scale parameter; b iis the i-th translation parameter; k is the frequency band index; i is the sampling point index, where i = 1, 2,..., N and N is the number of signal sampling points; j is the frequency band division index, where j = 1, 2,..., J and J is the number of frequency band divisions.

[0109] Schematically, an energy matrix is constructed according to the high-frequency structure feature coefficients, and a texture matrix is constructed according to the low-frequency texture feature coefficients;

[0110] Specifically, the energy matrix of the high-frequency structure feature coefficients is calculated by the following formula:

[0111]

[0112] Specifically, the texture matrix of the low-frequency texture feature coefficients is calculated by the following formula:

[0113]

[0114] where M E is an N×K-dimensional energy matrix, which describes the energy distribution characteristics of high-frequency features; M T is an N×K-dimensional texture matrix, which describes the texture distribution characteristics of low-frequency features; K is the number of frequency band divisions, which determines the accuracy of frequency analysis; N is the number of signal sampling points, which determines the accuracy of time analysis;

[0115] Schematically, principal component analysis is performed on the energy matrix to obtain the principal eigenvector, and the texture matrix is decomposed to obtain the eigenbasis vector;

[0116] Specifically, the following steps are used to perform principal component analysis on the energy matrix: First, the energy matrix is centered to obtain the processed energy matrix, then the covariance matrix is calculated according to the processed energy matrix, the covariance matrix is decomposed into eigenvalues and corresponding eigenvectors, and then the eigenvectors are sorted according to the corresponding eigenvalue magnitudes to obtain the principal eigenvector; where the ratio of each eigenvalue to the sum of all eigenvalues is the energy contribution rate of the principal component;

[0117] Specifically, the following steps are used to decompose the texture matrix: The texture matrix is decomposed into the form of U∑V T where U is the left singular matrix, V is the right singular matrix, and ∑ is the singular value matrix; the singular values are sorted from large to small, and the corresponding left singular vectors are the eigenbasis vectors; where the ratio of the square of each singular value to the sum of the squares of all singular values is the energy contribution rate of the eigenbasis vector;

[0118] Specifically, the following steps are used to perform singular value decomposition on the texture matrix: First, calculate the texture matrix T and its transpose matrix T TThe product of which gives the matrix T·T T ; Perform eigenvalue decomposition on the matrix T·T T to obtain the eigenvalues λ i and the corresponding eigenvectors u i ; Use the eigenvectors u i as the column vectors of the left singular matrix U; The square root of the eigenvalue λ i is the singular value σ i , which constitutes the diagonal elements of the singular value matrix Σ; Calculate the matrix T T ·T, and use the eigenvectors v i obtained by performing eigenvalue decomposition on it as the column vectors of the right singular matrix V; The singular values σ i are sorted from largest to smallest, and the corresponding left singular vectors are the eigenbasis vectors; The magnitude of the singular value reflects the importance of the corresponding eigenbasis vector in the expression of texture information. The larger the singular value, the greater the contribution of the texture feature in that direction;

[0119] Preferably, according to the energy contribution rate of the main eigenvectors and the eigenbasis vectors, select the main eigenvectors and eigenbasis vectors with an energy contribution rate reaching a preset ratio (such as 95%, which is not limited here) for subsequent orthogonal transformation analysis. This can reduce the data dimension while retaining the main feature information and improve the calculation efficiency;

[0120] Schematically, perform orthogonal transformation on the main eigenvectors and the eigenbasis vectors to obtain an orthogonal feature space containing defect feature information;

[0121] Specifically, perform orthogonal transformation operations on the selected main eigenvectors and eigenbasis vectors to generate an orthogonal feature space F:

[0122] It should be noted that in the present invention, the column vectors of the left singular matrix U obtained by performing singular value decomposition on the texture matrix T are the eigenbasis vectors. These eigenbasis vectors form a set of orthogonal bases and can represent the basic patterns of texture features from different angles. Specifically, each eigenbasis vector u i corresponds to a singular value σ i . The ratio of the square of this singular value to the sum of the squares of all singular values reflects the importance of this eigenbasis vector in texture expression. When the texture matrix is projected onto these eigenbasis vectors, the component sizes of the texture features in each basic pattern can be obtained, thereby realizing the low-dimensional expression of texture features; And the eigenbasis vectors obtained by decomposing the texture matrix exactly refer to the column vectors of the left singular matrix U obtained during the singular value decomposition process. In the subsequent construction process of the orthogonal feature space, these eigenbasis vectors, together with the main eigenvectors obtained by principal component analysis of the energy matrix, construct a complete defect feature representation space through orthogonal transformation, thereby realizing the effective fusion of structural features and texture features.

[0123] In a preferred embodiment, calculating the structural feature data and the texture feature data based on the orthogonal feature space, a preset feature demarcation frequency, the high-frequency structural feature coefficients, and the low-frequency texture feature coefficients includes:

[0124] Forming complementary feature pairs from the high-frequency structural feature coefficients and the low-frequency texture feature coefficients that satisfy a preset complementary judgment condition;

[0125] In the orthogonal feature space, mapping the complementary feature pairs into multi-scale feature representations;

[0126] Constructing a feature description matrix according to the multi-scale feature representation, and dividing the feature description matrix into the structural feature data and the texture feature data according to the feature demarcation frequency;

[0127] Schematically, forming complementary feature pairs from the high-frequency structural feature coefficients and the low-frequency texture feature coefficients that satisfy a preset complementary judgment condition;

[0128] In a preferred embodiment, forming the complementary feature pairs from the high-frequency structural feature coefficients and the low-frequency texture feature coefficients that satisfy a preset complementary judgment condition includes:

[0129] Calculating the overlapping interval in the frequency domain between the high-frequency structural feature coefficients and the low-frequency texture feature coefficients;

[0130] Calculating the ratio between the overlapping interval and a preset total frequency band width to obtain a frequency band overlap degree;

[0131] Calculating the cross-correlation coefficient between the high-frequency structural feature coefficients and the low-frequency texture feature coefficients;

[0132] Judging whether the frequency band overlap degree is less than a preset first frequency band threshold, and judging whether the cross-correlation coefficient is less than a preset second frequency band threshold;

[0133] If both judgment results are yes, determining that the high-frequency structural feature coefficients and the low-frequency texture feature coefficients satisfy the complementary judgment condition, and forming complementary feature pairs from the high-frequency structural feature coefficients and the low-frequency texture feature coefficients;

[0134] Otherwise, determining that the high-frequency structural feature coefficients and the low-frequency texture feature coefficients do not satisfy the complementary judgment condition;

[0135] Schematically, the complementarity between the high-frequency structural feature coefficients and the low-frequency texture feature coefficients is determined by calculating their frequency band overlap degree;

[0136] Specifically, determine the overlapping interval of the high-frequency structure feature coefficient and the low-frequency texture feature coefficient in the frequency domain, and use the ratio between the overlapping interval and the preset total frequency band width as the frequency band overlap degree;

[0137] When the frequency band overlap degree is lower than the preset first frequency band threshold, it indicates that the high-frequency structure feature coefficient mainly contains high-frequency band information while the low-frequency texture feature coefficient mainly contains low-frequency band information, and the two are complementary in frequency domain distribution;

[0138] Meanwhile, calculate the cross-correlation coefficient of the two feature coefficients. When the cross-correlation coefficient is lower than the preset second frequency band threshold and the frequency band overlap degree is lower than the preset first frequency band threshold, it indicates that the information carried by the two feature coefficients has strong complementarity. At this time, combine them into a complementary feature pair;

[0139] It should be noted that the first frequency band threshold is used to measure the overlapping degree of the high-frequency band and the low-frequency band. Considering the typical frequency feature distributions of surface defects and internal defects in electromagnetic ultrasonic testing, this threshold is usually set to 0.2 - 0.3, that is, 20% - 30% of frequency band overlap is allowed, which can not only ensure the integrity of feature information but also maintain the effectiveness of frequency band division; the second frequency band threshold is used to evaluate the correlation between the high-frequency structure feature coefficient and the low-frequency texture feature coefficient. Considering the characteristics that the surface cracks of the insulated part of the distribution line are usually in the MHz magnitude while the internal defect response is in the kHz magnitude, this threshold is preferably 0.4 - 0.5, and this range can effectively distinguish the feature manifestations of different types of defects;

[0140] By adjusting the first frequency band threshold and the second frequency band threshold, the detection sensitivity of different types of insulation defects can be optimized. For example, when it is necessary to improve the detection ability for tiny surface cracks, the first frequency band threshold can be appropriately reduced to enhance the extraction effect of high-frequency structure features;

[0141] Specifically, form a complementary feature pair with the high-frequency structure feature coefficient and the low-frequency texture feature coefficient that meet the complementary judgment conditions. The calculation expression of the complementary feature pair is as follows:

[0142]

[0143] Among them, P(k) is the complementary feature pair of the k-th frequency band, which contains the complete information of surface and internal defects; γ is the feature fusion weight coefficient (0 < γ < 1), which is dynamically adjusted according to the signal quality; is the feature fusion operator, which realizes the non-linear combination of high-frequency and low-frequency features.

[0144] Schematically, in the orthogonal feature space, the complementary feature pairs are mapped into a multi-scale feature representation, a feature description matrix is constructed based on the multi-scale feature representation, and according to a preset feature demarcation frequency, the feature description matrix is divided into the structural feature data and the texture feature data;

[0145] Specifically, the complementary feature pairs are first subjected to scale decomposition through multi-level wavelet transform, and the feature information is decomposed into multiple scale levels corresponding to defects of different sizes; then the features of each scale are mapped into the orthogonal feature space composed of the main eigenvector and the eigenbasis vector through inner product calculation to achieve the unification of feature expression; then adaptive sub-scale subdivision is performed according to the feature complexity, using more sub-scales to describe complex regions and fewer sub-scales to express simple regions; finally, the features of all scales and sub-scales are combined to form a complete multi-scale feature representation, covering from the macroscopic structural features (such as large-area insulation deterioration regions) at the coarse scale to the microscopic texture features (such as tiny cracks and bubbles) at the fine scale, realizing the comprehensive analysis of defect features at different scales and laying a foundation for the subsequent construction of the feature description matrix;

[0146] Specifically, the complementary feature pairs are mapped into a multi-scale feature representation, and the high-frequency structural feature data and the low-frequency texture feature data in the multi-scale feature representation are non-linearly combined through a preset feature fusion operator to construct a feature description matrix containing complete defect information, and according to the feature demarcation frequency, the feature data in the feature description matrix with a frequency higher than the feature demarcation frequency is divided into structural feature data, corresponding to macroscopic defect features; the feature data with a frequency lower than the feature demarcation frequency is divided into texture feature data, corresponding to microscopic defect features;

[0147] Specifically, the process of constructing a feature description matrix containing complete defect information includes: First, classify the obtained multi-scale feature representations, and divide them into a high-frequency structure feature group and a low-frequency texture feature group according to frequency characteristics; then calculate the importance weights of each feature group, where the importance weights are determined based on the mutual information between the features and the defect types and the feature stability; next, perform non-linear combination using a feature fusion operator based on an adaptive kernel function. For high-frequency features, a narrow-band Gaussian kernel function is used to retain fine structure information, and for low-frequency features, a wide-band radial basis kernel function is used to highlight the overall texture distribution. Non-linear fusion of features is achieved through the combination of weighted summation and a non-linear activation function; finally, the fused features are arranged and organized according to spatial position and frequency relationship to construct a feature description matrix, where the rows of the matrix represent features at different spatial positions and the columns represent features at different frequencies or scales; to improve the discrimination ability, sparse optimization processing is performed on the constructed feature description matrix, and redundant elements in the matrix are made to approach zero through L1 regularization constraints, retaining the most representative feature elements, so as to obtain a feature description matrix that contains both complete defect information and high discrimination ability. This matrix contains the information of both high-frequency structure features and low-frequency texture features and can comprehensively reflect various defect features of the insulation part of the distribution line;

[0148] It should be noted that the feature demarcation frequency is an empirical value and has been set in advance; the structure feature data in the feature description matrix corresponds to high-frequency components, and the texture feature data corresponds to low-frequency components, realizing the collaborative detection of macroscopic and microscopic defects in the insulation part of the distribution line.

[0149] S3. Construct a feature matrix based on the structure feature data and the texture feature data;

[0150] Schematically, the structure feature data and the texture feature data are respectively normalized using a preset adaptive normalization method, and then hierarchical sampling is respectively performed on the normalized structure feature data and the texture feature data:

[0151] Specifically, for the normalized structure feature data, calculate statistics such as the mean, variance, skewness, and kurtosis of the normalized structure feature data, and based on these statistics of the normalized structure feature data, divide the normalized structure feature data into three layers: high, medium, and low according to the energy value; for the normalized texture feature data, calculate statistics such as the mean, variance, skewness, and kurtosis of the normalized texture feature data, and based on these statistics of the normalized texture feature data, divide the texture feature data into three layers: complex, medium, and simple according to the texture complexity;

[0152] For each layer in the structural feature data after stratification, sample structural feature data is randomly selected according to a preset ratio (such as 7:3) to ensure that the distribution ratios of the sample structural feature data for each layer are consistent, thereby obtaining a training data set and a validation data set; for each layer in the texture feature data after stratification, sample texture feature data is randomly selected according to a preset ratio (such as 7:3) to ensure that the distribution ratios of the sample texture feature data for each layer are consistent, thereby obtaining a training data set and a validation data set;

[0153] The sample structural feature data and the sample texture feature data are subjected to standardization processing, and the standardized sample structural feature data and the sample texture feature data are aligned and combined according to the corresponding spatial positions and frequency band information to construct a feature matrix; wherein, the rows of the feature matrix represent different samples, the columns represent different feature dimensions, and the matrix element values are the standardized feature quantization values. This method for constructing a feature matrix based on stratified sampling ensures the representativeness of the samples and the integrity of the features;

[0154] It should be noted that the adaptive normalization method provided in this embodiment is a processing method that automatically selects normalization parameters according to the statistical distribution characteristics of different features. The z-score normalization based on the mean and standard deviation is used for the structural feature data, and the min-max normalization based on the maximum and minimum values is used for the texture feature data. By comparing the data distribution consistency of the two normalization methods, the optimal normalization strategy is selected.

[0155] S4. According to a preset attention weight allocation mechanism and the feature matrix, perform a feature weighting operation on the feature matrix to obtain a feature matrix after feature weighting;

[0156] In a preferred embodiment, the step of performing a feature weighting operation on the feature matrix according to a preset attention weight allocation mechanism and the feature matrix to obtain a feature matrix after feature weighting includes:

[0157] Taking the spatial resolution as the decomposition reference standard, decompose the feature matrix to obtain a feature sequence; wherein, the feature sequence contains scale features of different spatial resolutions;

[0158] According to the attention weight allocation mechanism, assign corresponding attention weights to each scale feature in the feature sequence to obtain a feature sequence after feature weighting;

[0159] Fuse the feature sequence after feature weighting to obtain a feature matrix after feature weighting;

[0160] Schematically, taking the spatial resolution as the decomposition reference standard, decompose the feature matrix to obtain a feature sequence;

[0161] Specifically, features with a spatial resolution at the original resolution scale are used to capture local features such as micron-scale surface cracks and micro-bubbles; features with a spatial resolution at the 1 / 2 scale, with the resolution reduced by half, are used to detect defect features such as medium-sized insulation aging areas, millimeter-scale surface spalling, and internal cavities; features with a spatial resolution at the 1 / 4 scale, with the resolution further reduced, are used to identify defect features such as large-area insulation degradation areas, centimeter-scale large-area insulation aging, and severe corrosion; features at each scale retain the electromagnetic and ultrasonic information at the corresponding scale.

[0162] Schematically, according to the attention weight assignment mechanism, corresponding attention weights are assigned to each scale feature in the feature sequence to obtain a feature sequence after feature weighting.

[0163] First, channel dimension compression is performed on the feature matrix: average pooling and max pooling operations are performed on the feature matrix along the channel dimension, the obtained features are concatenated, then feature fusion is performed through the convolutional layer of a preset attention model, and finally, a spatial attention feature is generated through a preset sigmoid activation function; then, spatial dimension compression is performed on the feature matrix: average pooling and max pooling are respectively performed on the feature matrix along the spatial dimension to generate vectors, the generated vectors are concatenated, and dimensionality reduction is performed through the fully connected layer of a preset attention model, and finally, a channel attention feature is generated through the sigmoid activation function; this dual compression mechanism can improve the detection sensitivity for different types of defects in the insulation part of the distribution line.

[0164] It should be noted that the attention weight assignment mechanism aims to calculate the weight coefficient dynamically according to the importance of the features, that is, by calculating the activation intensity of the feature sequence in the spatial domain and the channel domain, evaluating the contribution degree of different features, assigning higher weights to important features, and reducing the influence of redundant features.

[0165] Specifically, the contribution degree of each scale feature in the feature sequence is determined through the following steps: according to the proportion of the spatial attention feature corresponding to each scale feature in the overall spatial attention feature, determine the spatial domain contribution degree of each scale feature in the spatial domain; according to the proportion of the channel attention feature corresponding to each scale feature in the overall channel attention feature, determine the channel domain contribution degree of each scale feature in the channel domain; perform weighted summation on the spatial domain contribution degree and the channel domain contribution degree corresponding to each scale feature to obtain the contribution degree corresponding to each scale feature.

[0166] After determining the contribution degree of each of the scale features, the attention weight allocation mechanism will adopt a cascaded manner. According to the contribution degree of each scale feature and the connection relationship of each scale feature, corresponding attention weights are assigned to each scale feature. That is, when the contribution degree of the previous scale feature is relatively high, the attention weight of the scale feature connected to the previous scale feature will be correspondingly increased through a preset cascaded allocation method.

[0167] Schematically, after obtaining the feature sequence after feature weighting, fuse the feature sequence after feature weighting to obtain a feature matrix after feature weighting;

[0168] Specifically, fuse the feature sequence after feature weighting through a preset feature fusion method. It should be noted that the feature fusion method adopts methods such as weighted summation and feature splicing to integrate each scale feature to obtain a feature matrix after feature weighting.

[0169] S5. Input the feature matrix after feature weighting into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs the probability distribution of the defect types of the insulated parts of the distribution line according to the input data;

[0170] In a preferred embodiment, the construction process of the defect diagnosis network includes:

[0171] Obtain a distribution line fault sample set; wherein, the distribution line fault sample set includes: a number of distribution lines with defects and corresponding true defect types;

[0172] Construct a dual-branch collaborative diagnosis network. Using the distribution line fault sample set as the input and the predicted probability distribution of the defect types as the output, iteratively train the dual-branch collaborative diagnosis network until the loss function value of the dual-branch collaborative diagnosis network converges;

[0173] Take the trained dual-branch collaborative diagnosis network as the defect diagnosis network;

[0174] Among them, in each training process, compare the probability distribution of the defect types predicted by the current dual-branch collaborative diagnosis network with the corresponding true defect types in the distribution line fault sample set, calculate the loss function value according to the comparison result, and adjust the network parameters of the dual-branch collaborative diagnosis network according to the calculated loss function value;

[0175] Specifically, using the power distribution line fault sample set as the training samples, input them into the initial dual-branch collaborative diagnosis network. The main branch of the dual-branch collaborative diagnosis network processes the feature data related to the structural feature data in the training samples through a preset convolutional layer and a preset pooling layer to capture the overall structural characteristics of the insulation part of the power distribution line, such as macroscopic features like large-area insulation aging areas and obvious structural damages. The auxiliary branch of the dual-branch collaborative diagnosis network is used to process the feature data related to the texture feature data in the training samples to capture microscopic defect features such as micron-level surface cracks and tiny air bubbles.

[0176] During each iterative training process, the dual-branch collaborative diagnosis network compares the probability distribution of the defect types predicted by the main branch with the label vector of the corresponding true defect type in the power distribution line fault sample set, and thus calculates the value of the cross-entropy loss function of the main branch according to the comparison result. Similarly, the dual-branch collaborative diagnosis network compares the probability distribution of the defect types predicted by the auxiliary branch with the label vector of the corresponding true defect type in the power distribution line fault sample set, and thus calculates the value of the cross-entropy loss function of the auxiliary branch according to the comparison result.

[0177] Perform a weighted sum of the cross-entropy loss of the auxiliary branch and the value of the cross-entropy loss function of the main branch to determine the value of the loss function of the dual-branch collaborative diagnosis network. Then, adjust the network parameters of the dual-branch collaborative diagnosis network according to the value of the loss function of the dual-branch collaborative diagnosis network, and repeat the iterative training operation until the pre-constructed defect diagnosis network is finally obtained.

[0178] Input the feature matrix after feature weighting into the pre-constructed defect diagnosis network, so that the defect diagnosis network outputs the probability distribution of the defect types of the insulation part of the power distribution line.

[0179] S6. Determine the defect type of the insulation part of the power distribution line according to the probability distribution of the defect types.

[0180] Schematically, the defect types include surface cracks, partial discharges, internal cavities, internal air bubbles, etc.

[0181] Specifically, assume that the probability distribution of defect types output by step S5 is (taking surface cracks, partial discharge, internal voids, and internal bubbles as examples only, and the probability distribution of the remaining defect types is omitted here): surface cracks (35.2%), partial discharge (10.7%), internal voids (40.1%), internal bubbles (4.0%); the defect types exceeding the preset ratio are taken as the defect types of the insulated part of the distribution line. Assume that the preset ratio is 30%. Since the proportion of internal voids is 40.1% and the proportion of surface cracks is 35.2%, both are greater than the preset ratio. Therefore, it is determined that the defect types of the insulated part of the distribution line are internal voids and surface cracks.

[0182] To verify the effectiveness of the present invention, an experimental verification was carried out on the distribution line of a 10 kV substation in this embodiment. The experimental samples included 300 groups of different types of insulation defect samples, including typical defect types such as surface cracks, internal bubbles, and partial discharge. The samples were randomly divided into a training set and a test set. The method of the present invention was compared with the traditional single detection method. The experimental platform used a standard electromagnetic ultrasonic detection device equipped with a high-precision sensor array and a data acquisition system. The experimental results were comprehensively evaluated from multiple dimensions such as detection accuracy rate, missed detection rate, and false detection rate.

[0183] Table 1 Comparison table of detection methods

[0184]

[0185] As shown in Table 1, in the detection of surface cracks, the accuracy rate of the present invention reached 94.3%, which was 11.8 percentage points higher than that of the traditional electromagnetic detection; in the detection of internal bubbles, the accuracy rate of the present invention reached 92.8%, which was 7.1 percentage points higher than that of the traditional ultrasonic detection; for the detection of partial discharge, the accuracy rate of the present invention was 91.5%, which was significantly higher than all the comparison methods.

[0186] Table 2 Performance comparison table under complex scenarios

[0187]

[0188]

[0189] As shown in Table 2, in the case where multiple types of defects exist simultaneously, the recognition accuracy rate of the present invention reached 89.5%, which was 10.9 percentage points higher than that of the best comparison method (simple feature fusion 78.6%); the position positioning error was reduced to 4.2 mm, showing a significant improvement compared with other methods; in the anti-interference ability score, the present invention reached 8.7 points, far exceeding the traditional methods; the system stability score reached 8.9 points, indicating that the present invention has excellent robustness.

[0190] Refer to Figure 2, which is a device for identifying insulation defect types of a distribution line provided by an embodiment of the present invention, includes: a signal acquisition module, a signal decomposition module, a feature matrix construction module, a feature weighting module, a probability distribution calculation module, and a defect type determination module;

[0191] The signal acquisition module is used to acquire detection signals of the insulated part of the distribution line;

[0192] The signal decomposition module is used to decompose structural feature data and texture feature data from the detection signals;

[0193] The feature matrix construction module is used to construct a feature matrix according to the structural feature data and the texture feature data;

[0194] The feature weighting module is used to perform a feature weighting operation on the feature matrix according to a preset attention weight distribution mechanism and the feature matrix to obtain a feature-weighted feature matrix;

[0195] The probability distribution calculation module is used to input the feature-weighted feature matrix into a pre-constructed defect diagnosis network, so that the defect diagnosis network outputs the defect type probability distribution of the insulated part of the distribution line according to the input data;

[0196] The defect type determination module is used to determine the defect type of the insulated part of the distribution line according to the defect type probability distribution.

[0197] In a preferred embodiment, the construction process of the defect diagnosis network includes:

[0198] Obtain a distribution line fault sample set; wherein, the distribution line fault sample set includes: a number of distribution lines with defects and corresponding true defect types;

[0199] Construct a double-branch collaborative diagnosis network, use the distribution line fault sample set as the input, and the predicted defect type probability distribution as the output, and perform iterative training on the double-branch collaborative diagnosis network until the loss function value of the double-branch collaborative diagnosis network converges;

[0200] Use the trained double-branch collaborative diagnosis network as the defect diagnosis network;

[0201] Wherein, in each training process, compare the defect type probability distribution predicted by the current double-branch collaborative diagnosis network with the corresponding true defect type in the distribution line fault sample set, calculate the loss function value according to the comparison result, and adjust the network parameters of the double-branch collaborative diagnosis network according to the calculated loss function value.

[0202] SeeFigure 3 , an embodiment of the present application further provides a terminal device, including:

[0203] One or more processors;

[0204] A memory, coupled to the processor, for storing one or more programs;

[0205] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for identifying insulation defect types of a power distribution line.

[0206] The processor is used to control the overall operation of the terminal device to complete all or part of the steps of the above-described method for identifying insulation defect types of a power distribution line. The memory is used to store various types of data to support the operation of the terminal device. These data may include, for example, instructions for any application program or method operating on the terminal device, as well as application program-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0207] In an exemplary embodiment, the terminal device can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components, and is used to execute the method for identifying the insulation defect type of the distribution line as described in any one of the above embodiments, and achieve the same technical effect as the above method.

[0208] In another exemplary embodiment, there is also provided a computer-readable storage medium including a computer program. When the computer program is executed by a processor, the steps of the method for identifying the insulation defect type of the distribution line as described in any one of the above embodiments are implemented. For example, the computer-readable storage medium can be the above-mentioned memory including the computer program, and the above computer program can be executed by the processor of the terminal device to complete the method for identifying the insulation defect type of the distribution line as described in any one of the above embodiments, and achieve the same technical effect as the above method.

[0209] The above is the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for identifying insulation defect types of distribution lines, characterized in that: include: Obtain detection signals from the insulation parts of distribution lines; Decomposing structural feature data and texture feature data from the detection signal; Constructing a feature matrix according to the structural feature data and the texture feature data; According to a preset attention weight allocation mechanism and the feature matrix, a feature weighting operation is performed on the feature matrix to obtain a feature weighted feature matrix; Inputting the weighted feature matrix into a pre-built defect diagnosis network, so that the defect diagnosis network outputs a probability distribution of defect types at the insulation part of the distribution line according to the input data; The defect type of the insulation portion of the power distribution line is determined according to the defect type probability distribution.

2. The method for identifying the type of insulation defect of a distribution line according to claim 1, characterized in that: The detection signal includes a time domain electromagnetic detection signal and an ultrasonic detection signal; Decomposing the structural feature data and the texture feature data from the detection signal comprises: Performing wavelet transform on the time-domain electromagnetic detection signal to obtain electromagnetic characteristic coefficients; Performing wavelet transform on the ultrasonic detection signal to obtain ultrasonic characteristic coefficients; Calculating local energy distribution according to the electromagnetic characteristic coefficient and the ultrasonic characteristic coefficient; Performing frequency band decomposition on the electromagnetic characteristic coefficients to obtain high-frequency structural characteristic coefficients; According to the high-frequency structural characteristic coefficients, an energy matrix is ​​constructed, and a principal component analysis is performed on the energy matrix to obtain a principal eigenvector; Performing frequency band decomposition on the ultrasonic characteristic coefficient to obtain a low-frequency texture characteristic coefficient; Constructing a texture matrix according to the low-frequency texture feature coefficients, and decomposing the texture matrix to obtain feature basis vectors; Performing an orthogonal transformation on the main feature vector and the feature basis vector to obtain an orthogonal feature space containing defect feature information; The structural feature data and the texture feature data are calculated based on the orthogonal feature space, a preset feature boundary frequency, the high-frequency structural feature coefficient, and the low-frequency texture feature coefficient.

3. The method for identifying the type of insulation defect of a distribution line according to claim 2, characterized in that: The step of calculating the structural feature data and the texture feature data based on the orthogonal feature space, the preset feature boundary frequency, the high-frequency structural feature coefficient, and the low-frequency texture feature coefficient comprises: The high-frequency structural feature coefficients and the low-frequency texture feature coefficients that meet the preset complementarity judgment conditions are combined into a complementary feature pair; In the orthogonal feature space, mapping the complementary feature pair into a multi-scale feature representation; A feature description matrix is ​​constructed according to the multi-scale feature representation, and the feature description matrix is ​​divided into the structural feature data and the texture feature data according to the feature boundary frequency.

4. The method for identifying the type of insulation defect of a distribution line according to claim 3, characterized in that: The high-frequency structural feature coefficients and low-frequency texture feature coefficients that meet the preset complementarity judgment condition form a complementary feature pair, including: Calculating the overlapping interval of the high-frequency structural feature coefficient and the low-frequency texture feature coefficient in the frequency domain; Calculating the ratio between the overlapping interval and the preset total frequency bandwidth to obtain the frequency band overlap; Calculating the mutual correlation coefficient between the high-frequency structural feature coefficient and the low-frequency texture feature coefficient; determining whether the frequency band overlap is less than a preset first frequency band threshold, and determining whether the mutual correlation coefficient is less than a preset second frequency band threshold, If the judgment results are all yes, it is determined that the high-frequency structure feature coefficient and the low-frequency texture feature coefficient meet the complementarity judgment condition, and the high-frequency structure feature coefficient and the low-frequency texture feature coefficient form a complementary feature pair, On the contrary, it is determined that the high-frequency structural feature coefficient and the low-frequency texture feature coefficient do not satisfy the complementarity judgment condition.

5. The method for identifying the type of insulation defect of a distribution line according to claim 1, characterized in that: The step of performing a feature weighting operation on the feature matrix according to the preset attention weight allocation mechanism and the feature matrix to obtain a feature weighted feature matrix includes: Decomposing the feature matrix using spatial resolution as a decomposition reference standard to obtain a feature sequence; wherein the feature sequence includes scale features of different spatial resolutions; According to the attention weight allocation mechanism, each scale feature in the feature sequence is assigned a corresponding attention weight to obtain a feature sequence after feature weighting; The feature sequences after feature weighting are fused to obtain a feature matrix after feature weighting.

6. The method for identifying the type of insulation defect of a distribution line according to claim 1, characterized in that: The construction process of the defect diagnosis network includes: Acquire a distribution line fault sample set; wherein the distribution line fault sample set includes: a number of distribution lines with defects and corresponding real defect types; Constructing a dual-branch collaborative diagnosis network, taking the distribution line fault sample set as input and the predicted defect type probability distribution as output, and iteratively training the dual-branch collaborative diagnosis network until the loss function value of the dual-branch collaborative diagnosis network converges; The trained dual-branch collaborative diagnosis network is used as the defect diagnosis network; In each training process, the probability distribution of defect types predicted by the current dual-branch collaborative diagnosis network is compared with the corresponding actual defect types in the distribution line fault sample set, the loss function value is calculated according to the comparison result, and the network parameters of the dual-branch collaborative diagnosis network are adjusted according to the calculated loss function value.

7. A device for identifying insulation defect types of distribution lines, characterized in that: include: Signal acquisition module, signal decomposition module, feature matrix construction module, feature weighting module, probability distribution calculation module and defect type determination module; The signal acquisition module is used to acquire the detection signal of the insulation part of the distribution line; The signal decomposition module is used to decompose the structural feature data and texture feature data from the detection signal; The feature matrix construction module is used to construct a feature matrix according to the structural feature data and the texture feature data; The feature weighting module is used to perform a feature weighting operation on the feature matrix according to a preset attention weight allocation mechanism and the feature matrix to obtain a feature weighted feature matrix; The probability distribution calculation module is used to input the feature matrix after the feature weighting into the pre-built defect diagnosis network, so that the defect diagnosis network outputs the probability distribution of the defect type of the insulation part of the distribution line according to the input data; The defect type determination module is used to determine the defect type of the insulation part of the distribution line according to the defect type probability distribution.

8. The device for identifying insulation defect types of power distribution lines according to claim 7, characterized in that: The construction process of the defect diagnosis network includes: Acquire a distribution line fault sample set; wherein the distribution line fault sample set includes: a number of distribution lines with defects and corresponding real defect types; Constructing a dual-branch collaborative diagnosis network, taking the distribution line fault sample set as input and the predicted defect type probability distribution as output, and iteratively training the dual-branch collaborative diagnosis network until the loss function value of the dual-branch collaborative diagnosis network converges; The trained dual-branch collaborative diagnosis network is used as the defect diagnosis network; In each training process, the probability distribution of defect types predicted by the current dual-branch collaborative diagnosis network is compared with the corresponding actual defect types in the distribution line fault sample set, the loss function value is calculated according to the comparison result, and the network parameters of the dual-branch collaborative diagnosis network are adjusted according to the calculated loss function value.

9. A terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for identifying the type of insulation defects in a distribution line as described in any one of claims 1 to 6.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying the type of insulation defects in a distribution line according to any one of claims 1 to 6 is implemented.

Images