GIS partial discharge type identification method and device, terminal and medium

Abstract

The invention discloses a GIS partial discharge type identification method and device, a terminal and a medium. The method comprises the following steps: acquiring a statistical characteristic quantity data set of a historical PRPD atlas of partial discharge under each defect type of a GIS; preprocessing the obtained statistical characteristic quantity data set to obtain a sample set after dimension reduction; training a constructed deep residual shrinkage network DRSN recognition model by using the sample set, and optimizing network parameters of the model by taking the minimum SLoss loss function as a target; and loading the obtained statistical characteristic quantity real-time data of the GIS partial discharge PRPD atlas into the trained and optimized DRSN identification model to obtain an identification result of the GIS partial discharge type. According to the method, the type of GIS partial discharge can be quickly and accurately identified, and a powerful guarantee is provided for safe operation of GIS equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment detection, and in particular to a method, device, terminal and medium for identifying local discharge types of GIS. Background Art

[0002] Gas-insulated switchgear (GIS) is increasingly widely used in power systems due to its outstanding advantages such as compact structure, maintenance-free and low environmental impact. Various insulation defects may occur in GIS equipment during production, transportation, assembly and long-term operation, such as metal spikes generated inside, insulator surface abnormalities and floating potential insulation defects. If these defects are not discovered and handled in time, they will bring different degrees of safety hazards to the normal operation of the equipment. Therefore, in order to reduce safety hazards, it is necessary to identify the type of partial discharge and accurately evaluate the insulation state of GIS. At present, there are many studies on partial discharge identification: in some related technologies, the sample data is feature extracted by using statistical operators, and the principal component factors are extracted by PCA principal component analysis to reduce the dimension of the sample, and then the multi-classification support vector machine SVM is used to classify the discharge defect type; the main disadvantages of the PCA+SVM method are: its dimension reduction result does not fully consider the sample feature quantity, is easily affected by manual selection, and the accuracy of the recognition result is relatively low. In other related technologies, pattern recognition and discharge severity classification are performed by using or improving neural network algorithms. For example, some existing studies use single DRSN deep residual network model, MobileNetV3 lightweight network model, fuzzy neural network and MIV-MEA-LVQ network and other neural network algorithms for state recognition; although the results of these neural networks rely on debugging and training to make their recognition effect more accurate than the use of unified operators, these algorithms ignore the complex relationship and mutual influence characteristics of sample data, and have problems such as instability, large amount of calculation, and overfitting. In addition, some studies have provided a method for using SCG (scaled conjugate gradient) algorithm for local discharge pattern recognition of GIS equipment. However, the conjugate gradient algorithm is suitable for solving large coefficient matrices, and there is a possibility of too many iterations, which may lead to errors, and it is easy to have problems such as poor adaptability and insufficient robustness. Summary of the invention

[0003] In order to solve the deficiencies in the prior art and further improve the effect of pattern recognition, the present invention provides a GIS local discharge type identification method, device, terminal and medium, which can quickly and accurately identify the type of GIS local discharge, and provide a strong guarantee for the safe operation of GIS equipment.

[0004] The present invention adopts the following technical solution.

[0005] In a first aspect, the present invention provides a method for identifying a GIS partial discharge type, the method comprising:

[0006] Obtain the statistical characteristic data set of the historical PRPD spectrum of partial discharge under various defect types of GIS;

[0007] Preprocessing the acquired statistical feature data set to obtain a sample set after dimensionality reduction;

[0008] The sample set is used to train the constructed deep residual shrinkage network DRSN recognition model, and the network parameters of the model are optimized with the minimum SLoss loss function as the goal;

[0009] The real-time data of the statistical characteristics of the GIS partial discharge PRPD spectrum is loaded into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

[0010] Optionally, the statistical characteristics of the PRPD spectrum include the maximum discharge amplitude of the positive / negative half cycle, the average discharge amplitude of the positive / negative half cycle, the positive / negative half cycle pulse repetition rate, the total pulse repetition rate, the positive / negative half cycle skewness, the positive / negative half cycle steepness, the phase asymmetry, the discharge asymmetry and the cross-correlation factor.

[0011] Optionally, the step of preprocessing the acquired statistical feature quantity data set includes:

[0012] The obtained statistical feature quantity data set is tested for missing values ​​and outliers. If there are missing values, they are interpolated and supplemented. If there are outliers, the outliers are replaced by adjacent normal values ​​or directly eliminated to obtain a complete and usable initial data set.

[0013] Perform Z-score standardization on the initial data set to obtain an initial feature sample set X;

[0014] The improved t-SNE algorithm is used to perform dimension reduction processing on the initial feature sample set X to obtain a sample set X' after dimension reduction.

[0015] Optionally, the step of performing dimensionality reduction processing on the initial feature sample set X using the improved t-SNE algorithm includes:

[0016] Calculate the covariance matrix S of the initial feature sample set X;

[0017] Based on the covariance matrix S, the Mahalanobis distance between each sample in the initial feature sample set X is calculated respectively;

[0018] Combined with the Mahalanobis distance, the similarity between each sample in the initial feature sample set X in its original high-dimensional space is calculated by Gaussian distribution, and its expression is as follows:

[0019]

[0020] Where D ij represents the initial feature sample set X = (X1, X2, ..., X n ) in the i-th sample X i With the jth sample X j Mahalanobis distance between two samples; i = 1, 2…n, j = 1, 2…n, i≠j, n represents the total number of samples; D il Represents sample X i With sample X l The Mahalanobis distance between them, l≠i; σ is a hyperparameter that controls the width of the Gaussian distribution; p ij Represents sample X i With sample X j The similarity between

[0021] Use X'=(X'1, X'2, ..., X' n ) is used to characterize the dimensionality reduction target sample set mapped from the initial feature sample set X to the low-dimensional space, and the similarity between the samples in the sample set X' in the low-dimensional space is represented by a t distribution with a degree of freedom of 1, which is expressed as follows:

[0022]

[0023] Where D i ' j Represents the i-th sample X' in the sample set X' i With the jth sample X j 'The Mahalanobis distance between il Represents sample X' i With sample X' l The Mahalanobis distance between ij Represents sample X' i With sample X j ' Similarity between;

[0024] Based on the similarity p ij and q ij , KL divergence is used to characterize the difference in probability distribution between high-dimensional space and low-dimensional space, and the dimensionality reduction model with the goal of minimizing KL divergence is constructed as follows:

[0025]

[0026] minL(D i ' j )

[0027]

[0028] In the formula, L(D i 'j ) represents the Mahalanobis distance D in low-dimensional space i ' j is the KL divergence objective function of the variable to be solved;

[0029] The dimension reduction model is solved to obtain the final sample set X' after dimension reduction.

[0030] Optionally, the calculation formula for calculating the Mahalanobis distance between each sample in the initial feature sample set X based on the covariance matrix S is as follows:

[0031]

[0032] In the formula, S -1 represents the inverse matrix of the covariance matrix S; T represents the transpose;

[0033] Correspondingly, sample X' i With sample X j The Mahalanobis distance D i ' j The calculation formula is as follows:

[0034]

[0035] In the formula, S' -1 Represents the inverse matrix of the covariance matrix S'.

[0036] Optionally, the algorithm used to solve the dimensionality reduction model is the Adam optimization algorithm;

[0037] The Adam optimization algorithm includes: based on the objective function L(D i ' j )’s first-order moment estimation and second-order moment estimation are used to iteratively update the learning parameter α to optimize the distribution of each sample in the low-dimensional space and thus obtain the optimal solution.

[0038] Optionally, the dimensionality reduction model is solved using an Adam optimization algorithm, and the specific steps include:

[0039] Step A: Initialize the first-order moment estimate u1, the second-order moment estimate u2 and the learning parameter α of the dimensionality reduction model;

[0040] Step B: According to the current objective function L(D i ' j ), updating the first-order moment estimate u1 and the second-order moment estimate u2;

[0041] Step C: perform bias correction on the updated first-order moment estimate u1 and second-order moment estimate u2;

[0042] Step D: Based on the corrected first-order estimate u1 and second-order moment estimate u2, the learning parameter α is updated to obtain an updated value of the Mahalanobis distance of each sample in the low-dimensional space;

[0043] Step E: According to the updated value of the Mahalanobis distance of each sample in the low-dimensional space, determine the current objective function L(D i ' j ) converges or reaches the maximum number of iterations; if not, return to step B; if so, output the distribution of each sample in the low-dimensional space based on the currently updated Mahalanobis distance, so as to obtain the final sample set X' after dimensionality reduction.

[0044] Optionally, the formula for updating the first-order moment estimate u1 and the second-order moment estimate u2 is as follows:

[0045] u1 t =β1·u1 t-1 +(1-β1)·g t

[0046]

[0047] Where u1 t and u1 t-1 Respectively represent the first-order moment estimates at the t-th and t-1-th iterations; u2 t and u2 t-1 They represent the second-order moment estimates at the t-th and t-1-th iterations, respectively; t is the number of iterations; β1 and β2 are the coefficients of the first-order moment estimate and the second-order moment estimate, respectively; g t is the objective function L(D i ' j )’s gradient.

[0048] Optionally, the expressions for bias correction of the updated first-order moment estimate u1 and second-order moment estimate u2 are as follows:

[0049]

[0050] In the formula, and They represent the first-order moment estimate and the second-order moment estimate after bias correction at the t-th iteration, respectively. and They represent the t-th power of coefficients β1 and β2 respectively.

[0051] Optionally, in step D, the learning parameter α is updated to obtain the calculation formulas of the updated value of the Mahalanobis distance of each sample in the low-dimensional space as follows:

[0052]

[0053] In the formula, αt and α t-1 denote the learning parameters at the t-th and t-1-th iterations respectively; ε is a set constant; D i ' j (t) and D i ' j (t-1) represents the sample X' at the t-th and t-1-th iterations respectively i With sample X j ' is the Mahalanobis distance between .

[0054] Optionally, the constructed deep residual shrinkage network DRSN recognition model includes a convolutional layer, multiple residual shrinkage modules, a batch processing layer, a ReLU activation layer, a global mean pooling layer and a fully connected layer connected in sequence; the structure of each residual shrinkage module includes two branches, the first branch includes a convolutional layer with a convolution kernel size of 1, and the second branch includes a first batch processing layer, a first ReLU activation layer, a first convolutional layer, a second batch processing layer, a second ReLU activation layer, a first convolutional layer, a global average pooling layer, a first fully connected layer, a second fully connected layer and a Sigmoid activation layer;

[0055] Among them, the input data of the residual shrinkage module is processed by the first batch processing layer, the first ReLU activation layer, the first convolution layer, the second batch processing layer, the second ReLU activation layer and the first convolution layer in sequence, and the first feature output is multiplied by the second feature output after the first feature is processed by the global average pooling layer, the first fully connected layer, the second fully connected layer and the Sigmoid activation layer in sequence, and then added to the convolution feature of the input data after being processed by another convolution layer to obtain the output result of the residual shrinkage module.

[0056] Optionally, the expression of the SLoss loss function is as follows:

[0057]

[0058] In the formula, is the discharge type recognition result of the kth input sample; y k is the actual discharge type corresponding to the kth input sample; n is the total number of samples.

[0059] In a second aspect, the present invention provides a GIS partial discharge type identification device, which runs the steps of any method described in the first aspect of the present invention, and the device comprises:

[0060] The data acquisition module is used to obtain the statistical characteristic data set of the historical PRPD spectrum of partial discharge under various defect types of GIS;

[0061] A preprocessing module is used to preprocess the acquired statistical feature data set to obtain a sample set after dimensionality reduction;

[0062] A training optimization module is used to train the constructed deep residual shrinkage network DRSN recognition model using the sample set, and optimize the network parameters of the model with the goal of minimizing the SLoss loss function;

[0063] The discharge type recognition module is used to load the real-time data of the statistical characteristic quantities of the acquired GIS partial discharge PRPD spectrum into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

[0064] In a third aspect, the present invention provides a terminal, including a processor and a storage medium;

[0065] The storage medium is used to store instructions;

[0066] The processor is used to operate according to the instructions to execute the steps of any one of the methods described in the first aspect of the present invention.

[0067] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect of the present invention.

[0068] The beneficial effects of the present invention are as follows:

[0069] 1. The present invention extracts statistical feature quantities of the PRPD spectrum of GIS partial discharge, uses the t-SNE algorithm to reduce the dimension of the feature quantities, and adopts the Adam optimization algorithm to solve the minimization objective function. According to the results of Adam optimization, the learning parameters are updated, redundant information is removed, and the most representative features are retained. A sample set of GIS partial discharge spectrum feature dimension reduction is established, and the deep residual shrinkage network DRSN recognition model is optimized and trained, thereby realizing rapid and accurate recognition of GIS partial discharge types.

[0070] 2. The current t-SNE algorithm usually uses Euclidean distance to characterize the similarity between samples. There is only the straight-line distance between two points calculated, ignoring the correlation between each feature, which is not conducive to the complexity of the PRPD atlas data of GIS partial discharge, and is easy to affect the problem of identification results; the present invention uses Mahalanobis distance to calculate the similarity between samples, which can more comprehensively reflect the relationship between features and reduce the complexity of calculation. Since a sample includes the feature quantity of four signals, and there is a certain connection between the same feature of different signals of the same voltage level sample and the different features of the same signal, the Mahalanobis distance weights the difference between each feature by the inverse matrix of the covariance matrix, takes the connection between various features into consideration, and can more comprehensively retain the connection between sample features. Therefore, the t-SNE algorithm improved based on the Mahalanobis distance of the present invention still retains the advantages of the standard algorithm. In low-dimensional space, it can effectively retain the local structure of the data, and give greater weight to the data points far away, plus the use of Mahalanobis distance to more accurately display the structural characteristics of the data, which helps to alleviate the crowding problem of high-dimensional reduction to low-dimensionality, and helps to improve the learning efficiency and accuracy of the feature quantity sample set in the neural network model.

[0071] 3. In the process of solving the KL divergence minimization objective function of the t-SNE algorithm, gradient descent is a commonly used method, but it has the problems of slow convergence speed and large memory requirement; the Adam optimization algorithm provided by the present invention only calculates the first-order derivative of the KL divergence, which greatly reduces the amount of calculation, improves the calculation efficiency, and the effect is also improved compared with the gradient descent algorithm; the Adam optimization algorithm is used to perform t-SNE dimension reduction, remove redundant information, and reduce the instability of subsequent deep residual shrinkage network DRSN recognition results; the present invention trains the deep residual shrinkage network DRSN based on the sample set after dimensionality reduction to remove the influence of artificial selection and improve the robustness of the features. To a certain extent, it overcomes the problems of low efficiency, low accuracy and excessive iterations of the neural network in the existing methods using only the unified operator. The recognition accuracy of the present invention is improved compared with the existing methods using only the neural network model or using the dimensionality reduction clustering method, and can realize fast and accurate identification of the type of GIS local discharge, providing important theoretical support for the optimal design and safe operation of GIS equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a schematic diagram of the flow of the method for identifying the type of GIS partial discharge in the present invention;

[0073] Figure 2 Schematic diagram of the structure of the residual shrinkage module in the present invention;

[0074] Figure 3It is a comparison chart of the prediction accuracy of the test set by the method of the present invention and the existing methods of PCA+K-means, PCA+SVM and DRSN respectively;

[0075] Figure 4 This is a structural principle block diagram of the GIS partial discharge type identification device in the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0077] Embodiment 1:

[0078] Reference Figure 1 The embodiment of the present invention provides a method for identifying the type of partial discharge in a GIS, which specifically includes the following steps:

[0079] Step S1, obtaining a statistical feature data set of historical PRPD spectra of partial discharge under various defect types of GIS;

[0080] Step S2: preprocessing the acquired statistical feature data set to obtain a sample set after dimensionality reduction;

[0081] Step S3, using the sample set to train the constructed deep residual shrinkage network DRSN recognition model, and optimizing the network parameters of the model with the goal of minimizing the SLoss loss function;

[0082] Step S4: Load the acquired real-time data of the statistical characteristic quantity of the GIS partial discharge PRPD spectrum into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

[0083] A preferred but non-limiting embodiment, in step S1, a SiPM sensor is used to collect the optical signals of the ultraviolet band, the visible light band and the infrared band generated during the partial discharge process, the voltage sensor is used to capture the voltage phase signal of the partial discharge, and the high-frequency current sensor is responsible for obtaining the HFCT signal of the partial discharge, thereby obtaining the PRPD spectrum of the three-band optical signal and the HFCT signal-phase. In the stage of extracting the statistical feature of the PRPD spectrum, the collected pulse current signal data and the optical signal data of the three bands are first filtered and processed, and then the statistical feature of the PRPD spectrum of the pulse current signal and the optical signal is extracted to obtain a data set of each defect type of GIS. Among them, the statistical feature of the PRPD spectrum includes the maximum discharge amplitude of the positive / negative half cycle, the average discharge amplitude of the positive / negative half cycle, the positive / negative half cycle pulse repetition rate, the total pulse repetition rate, the positive / negative half cycle skewness, the positive / negative half cycle steepness, the phase asymmetry, the discharge asymmetry and the cross-correlation factor. In this embodiment, the matrix X of the original data set n×m is obtained = (X1, X2,…, X n ), a voltage level sample X in X i =(x i1 ,x i2 ,…,x im ) T , where x im For voltage level sample X i The mth statistical feature quantity of the species, the number of samples in the original data set is n, the dimension of each voltage level sample is m, and m=14 in this embodiment.

[0084] It is further explained that the maximum discharge amplitude ±Amax of the positive / negative half cycle and the average discharge amplitude Aav can reflect the severity of local discharge. The maximum discharge amplitude selects the maximum value of the absolute value of the positive half cycle pulse and the negative half cycle pulse according to the positive and negative half cycles. The average discharge amplitude divides all the discharge pulse amplitudes within the measurement time range of an integer multiple of the power frequency cycle into positive and negative half cycles, and calculates the average value respectively; the pulse repetition rate N is the ratio of the total number of discharge pulses occurring within the measurement time range to the time range, that is, the number of discharge pulses occurring per second, which includes the positive / negative half cycle pulse repetition rate. The positive / negative half-cycle skewness ±Sk describes the left-right skewness of the sample contour when the data set is compared with the normal distribution; the positive / negative half-cycle steepness ±Ku describes the upward convexity of the sample contour of the data set relative to the normal distribution; the phase asymmetry φ reveals the distribution of the discharge starting phase in the positive and negative half-cycles; the discharge asymmetry Q reflects the characteristics of the discharge energy distribution in the positive and negative half-cycles of the AC voltage; the cross-correlation factor Cc reflects the correlation between the shapes of the discharge pulse clusters in the positive and negative half-cycles of the AC voltage.

[0085] As an embodiment of the present invention, the step of preprocessing the acquired statistical feature quantity data set in step S2 includes:

[0086] S2.1. Detect missing values ​​and outliers on the acquired statistical feature data set. If missing values ​​exist, interpolate and supplement them. If outliers exist, replace them with adjacent normal values ​​or directly remove them, so as to obtain a complete and usable initial data set.

[0087] S2.2, performing Z-score normalization processing on the initial data set to obtain an initial feature sample set X;

[0088] S2.3, using the improved t-SNE algorithm to perform dimensionality reduction processing on the initial feature sample set X to obtain a sample set X' after dimensionality reduction; specifically including:

[0089] S2.3.1. Calculate the covariance matrix S of the initial feature sample set X:

[0090]

[0091] In the formula, μ i Represents the i-th sample X in the initial feature sample set X i The expected value of , i = 1, 2…n; T represents transpose.

[0092] S2.3.2. Based on the covariance matrix S, the Mahalanobis distance between each sample in the initial feature sample set X is calculated respectively. The calculation formula is as follows:

[0093]

[0094] In the formula, S -1 represents the inverse matrix of the covariance matrix S; T represents the transpose;

[0095] S2.3.3, combined with the Mahalanobis distance, the similarity between each sample in the initial feature sample set X in its original high-dimensional space is calculated by Gaussian distribution, and the expression is as follows:

[0096]

[0097] Where D ij represents the initial feature sample set X = (X1, X2, ..., X n ) in the i-th sample X i With the jth sample X j Mahalanobis distance between two samples; i = 1, 2…n, j = 1, 2…n, i≠j, n represents the total number of samples; D il Represents sample X i With sample X lThe Mahalanobis distance between them, l≠i; σ is a hyperparameter that controls the width of the Gaussian distribution; p ij Represents sample X i With sample X j The similarity between

[0098] S2.3.4. Use X'=(X'1, X'2,...,X' n ) is used to characterize the dimensionality reduction target sample set mapped from the initial feature sample set X to the low-dimensional space, and the similarity between the samples in the sample set X' in the low-dimensional space is represented by a t distribution with a degree of freedom of 1, which is expressed as follows:

[0099]

[0100] Where D i ' j Represents the i-th sample X' in the sample set X' i With the jth sample X j 'The Mahalanobis distance between il Represents sample X' i With sample X' l The Mahalanobis distance between ij Represents sample X' i With sample X j ' similarity between; where sample X' i With sample X j The Mahalanobis distance D i ' j The calculation formula is as follows:

[0101]

[0102] In the formula, S' -1 Represents the inverse matrix of the covariance matrix S'; wherein, the covariance matrix S' is calculated in the same way as the matrix S, which will not be repeated here.

[0103] S2.3.5, based on the similarity p ij and q ij , KL divergence is used to characterize the difference in probability distribution between high-dimensional space and low-dimensional space, and the dimensionality reduction model with the goal of minimizing KL divergence is constructed as follows:

[0104]

[0105] minL(D i ' j )

[0106]

[0107] In the formula, L(D i ' j) represents the Mahalanobis distance D in low-dimensional space i ' j is the KL divergence objective function of the variable to be solved;

[0108] S2.3.5. Solve the dimension reduction model to obtain the final sample set X' after dimension reduction.

[0109] Furthermore, this embodiment uses KL divergence to compare the probability distribution difference between the high-dimensional space and the low-dimensional space, and uses the Adam algorithm to update the learning parameter α to obtain the minimum KL divergence, thereby outputting a sample set after dimensionality reduction.

[0110] As an embodiment of the present invention, step S2.3.5 uses the Adam optimization algorithm to solve the dimension reduction model, based on the objective function L(D i ' j ) is used to iteratively update the learning parameter α to optimize the distribution of each sample in the low-dimensional space, thereby obtaining the optimal solution; the specific steps include:

[0111] Step A: Initialize the first-order moment estimate u1, the second-order moment estimate u2 and the learning parameter α of the dimensionality reduction model; in this embodiment, the initial values ​​of the first-order moment estimate u1 and the second-order moment estimate u2 are 0 and biased toward 0; the initial value of the learning parameter α is set to 0.001.

[0112] Step B: According to the current objective function L(D i ' j ), the first-order moment estimate u1 and the second-order moment estimate u2 are updated; the update formulas are as follows:

[0113] u1 t =β1·u1 t-1 +(1-β1)·g t

[0114]

[0115] Where u1 t and u1 t-1 Respectively represent the first-order moment estimates at the t-th and t-1-th iterations; u2 t and u2 t-1 They represent the second-order moment estimation at the t-th and t-1-th iterations respectively; t is the number of iterations; β1 and β2 are the coefficients of the first-order moment estimation and the second-order moment estimation respectively; in this embodiment, the values ​​of β1 and β2 are 0.9 and 0.999 respectively. t is the objective function L(D i ' j ), its expression is:

[0116]

[0117] In the formula, L(D i ' j ) t is the objective function value at the tth iteration; δ represents the derivative of the objective function with respect to the sample points after the current dimensionality reduction.

[0118] Step C: perform bias correction on the updated first-order moment estimate u1 and second-order moment estimate u2; their expressions are as follows:

[0119]

[0120] In the formula, and They represent the first-order moment estimate and the second-order moment estimate after bias correction at the t-th iteration, respectively. and They represent the t-th power of coefficients β1 and β2 respectively.

[0121] Step D: Based on the corrected first-order estimate u1 and second-order moment estimate u2, the learning parameter α is updated to obtain the updated value of the Mahalanobis distance of each sample in the low-dimensional space; the calculation formulas are as follows:

[0122]

[0123] In the formula, α t and α t-1 denote the learning parameters at the t-th and t-1-th iterations respectively; ε is a set constant, and in this implementation, ε is set to 1.0×10 -8 ;D i ' j (t) and D i ' j (t-1) represents the sample X' at the t-th and t-1-th iterations respectively i With sample X j ' is the Mahalanobis distance between .

[0124] Step E: According to the updated value of the Mahalanobis distance of each sample in the low-dimensional space, determine the current objective function L(D i ' j ) converges or reaches the maximum number of iterations; if not, return to step B; if so, output the distribution of each sample in the low-dimensional space based on the currently updated Mahalanobis distance, so as to obtain the final sample set X' after dimensionality reduction.

[0125] Reference Figure 1 and Figure 2As shown, the deep residual shrinkage network DRSN recognition model constructed in this embodiment includes a convolutional layer, multiple residual shrinkage modules, a batch processing layer, a ReLU activation layer, a global mean pooling layer and a fully connected layer connected in sequence.

[0126] In this embodiment, the number of residual shrinkage modules is three; the structure of each residual shrinkage module includes two branches, the first branch includes a convolution layer with a convolution kernel size of 1, and the second branch includes a first batch processing layer, a first ReLU activation layer, a first convolution layer, a second batch processing layer, a second ReLU activation layer, a first convolution layer, a global average pooling layer, a first fully connected layer, a second fully connected layer and a Sigmoid activation layer; wherein, the first feature of the input data is processed by the first batch processing layer, the first ReLU activation layer, the first convolution layer, the second batch processing layer, the second ReLU activation layer and the first convolution layer in sequence, and the first feature is multiplied by the second feature of the first feature after being processed by the global average pooling layer, the first fully connected layer, the second fully connected layer and the Sigmoid activation layer in sequence, and then added to the convolution feature of the input data after being processed by the other convolution layer to obtain the output result of the residual shrinkage module.

[0127] As an embodiment of the present invention, the process of training and optimizing the constructed deep residual shrinkage network DRSN recognition model in step S3 includes:

[0128] S3.1, divide the sample set X' after dimension reduction into training set and test set in a ratio of 9:1;

[0129] S3.2, training the constructed DRSN recognition model based on the training set;

[0130] S3.3, based on the training results of the training set, combined with the test set, the network parameters of the DRSN recognition model are optimized with the goal of minimizing the SLoss loss function to obtain the final trained DRSN recognition model; wherein the expression of the SLoss loss function is as follows:

[0131]

[0132] In the formula, is the discharge type recognition result of the kth input sample; y k is the actual discharge type corresponding to the kth input sample; n is the total number of samples.

[0133] Reference Figure 3 In order to verify the effectiveness of the method of the present invention, this embodiment uses the GIS partial discharge type recognition method provided by the present invention and the PCA+K-means, PCA+SVM, and DRSN in the prior art to perform discharge type recognition experiments on the test set. Figure 3 It can be seen from the comparison results that the recognition accuracy of the present invention is significantly higher than that of other methods, which is increased by 8.73%, 4.80% and 3.08% respectively, ensuring the recognition accuracy and providing a strong guarantee for the safe operation of GIS equipment.

[0134] The beneficial effects of the present invention are as follows:

[0135] 1. The present invention extracts statistical feature quantities of the PRPD spectrum of GIS partial discharge, uses the t-SNE algorithm to reduce the dimension of the feature quantities, and adopts the Adam optimization algorithm to solve the minimization objective function. According to the results of Adam optimization, the learning parameters are updated, redundant information is removed, and the most representative features are retained. A sample set of GIS partial discharge spectrum feature dimension reduction is established, and the deep residual shrinkage network DRSN recognition model is optimized and trained, thereby realizing rapid and accurate recognition of GIS partial discharge types.

[0136] 2. The current t-SNE algorithm usually uses Euclidean distance to characterize the similarity between samples. There is only the straight-line distance between two points calculated, ignoring the correlation between each feature, which is not conducive to the complexity of the PRPD atlas data of GIS partial discharge, and is easy to affect the problem of identification results; the present invention uses Mahalanobis distance to calculate the similarity between samples, which can more comprehensively reflect the relationship between features and reduce the complexity of calculation. Since a sample includes the feature quantity of four signals, and there is a certain connection between the same feature of different signals of the same voltage level sample and the different features of the same signal, the Mahalanobis distance weights the difference between each feature by the inverse matrix of the covariance matrix, takes the connection between various features into consideration, and can more comprehensively retain the connection between sample features. Therefore, the t-SNE algorithm improved based on the Mahalanobis distance of the present invention still retains the advantages of the standard algorithm. In low-dimensional space, it can effectively retain the local structure of the data, and give greater weight to the data points far away, plus the use of Mahalanobis distance to more accurately display the structural characteristics of the data, which helps to alleviate the crowding problem of high-dimensional reduction to low-dimensionality, and helps to improve the learning efficiency and accuracy of the feature quantity sample set in the neural network model.

[0137] 3. In the process of solving the KL divergence minimization objective function of the t-SNE algorithm, gradient descent is a commonly used method, but it has the problems of slow convergence speed and large memory requirement; the Adam optimization algorithm provided by the present invention only calculates the first-order derivative of the KL divergence, which greatly reduces the amount of calculation, improves the calculation efficiency, and the effect is also improved compared with the gradient descent algorithm; the Adam optimization algorithm is used to perform t-SNE dimension reduction, remove redundant information, and reduce the instability of subsequent deep residual shrinkage network DRSN recognition results; the present invention trains the deep residual shrinkage network DRSN based on the sample set after dimensionality reduction to remove the influence of artificial selection and improve the robustness of the features. To a certain extent, it overcomes the problems of low efficiency, low accuracy and excessive iterations of the neural network in the existing methods using only the unified operator. The recognition accuracy of the present invention is improved compared with the existing methods using only the neural network model or using the dimensionality reduction clustering method, and can realize fast and accurate identification of the type of GIS local discharge, providing important theoretical support for the optimal design and safe operation of GIS equipment.

[0138] Embodiment 2:

[0139] like Figure 4 As shown, the present invention provides a GIS partial discharge type identification device, which is used to implement the steps of the method in the above embodiment 1, and the device specifically includes:

[0140] The data acquisition module is used to obtain the statistical characteristic data set of the historical PRPD spectrum of partial discharge under various defect types of GIS;

[0141] A preprocessing module is used to preprocess the acquired statistical feature data set to obtain a sample set after dimensionality reduction;

[0142] A training optimization module is used to train the constructed deep residual shrinkage network DRSN recognition model using the sample set, and optimize the network parameters of the model with the goal of minimizing the SLoss loss function;

[0143] The discharge type recognition module is used to load the real-time data of the statistical characteristic quantities of the acquired GIS partial discharge PRPD spectrum into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

[0144] The GIS partial discharge type identification device provided in the embodiment of the present invention and the GIS partial discharge type identification method provided in Example 1 are based on the same technical concept, and can produce the beneficial effects as described in Example 1. For the contents not described in detail in this embodiment, please refer to Example 1.

[0145] Embodiment three:

[0146] A terminal provided by an embodiment of the present invention includes a processor and a storage medium;

[0147] The storage medium is used to store instructions;

[0148] The processor is used to operate according to the instruction to execute the steps of the method according to any one of the first embodiments.

[0149] Embodiment 4:

[0150] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of any method described in Embodiment 1 are implemented.

[0151] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0152] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0153] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0154] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying GIS partial discharge types, characterized in that the method include: Obtain the statistical characteristic data set of the historical PRPD spectrum of partial discharge under various defect types of GIS; Preprocessing the acquired statistical feature data set to obtain a sample set after dimensionality reduction; The sample set is used to train the constructed deep residual shrinkage network DRSN recognition model, and the network parameters of the model are optimized with the minimum SLoss loss function as the goal; The real-time data of the statistical characteristics of the GIS partial discharge PRPD spectrum is loaded into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

2. The GIS partial discharge type identification method according to claim 1 is characterized in that: The statistical characteristics of the PRPD spectrum include the maximum discharge amplitude of the positive / negative half cycle, the average discharge amplitude of the positive / negative half cycle, the positive / negative half cycle pulse repetition rate, the total pulse repetition rate, the positive / negative half cycle skewness, the positive / negative half cycle steepness, the phase asymmetry, the discharge asymmetry and the cross-correlation factor.

3. The GIS partial discharge type identification method according to claim 1 is characterized in that: The steps of preprocessing the acquired statistical feature data set include: The obtained statistical feature quantity data set is tested for missing values ​​and outliers. If there are missing values, they are interpolated and supplemented. If there are outliers, the outliers are replaced by adjacent normal values ​​or directly eliminated to obtain a complete and usable initial data set. Perform Z-score standardization on the initial data set to obtain an initial feature sample set X; The improved t-SNE algorithm is used to perform dimension reduction processing on the initial feature sample set X to obtain a sample set X' after dimension reduction.

4. The GIS partial discharge type identification method according to claim 3 is characterized in that: The step of using the improved t-SNE algorithm to perform dimensionality reduction processing on the initial feature sample set X includes: Calculate the covariance matrix S of the initial feature sample set X; Based on the covariance matrix S, the Mahalanobis distance between each sample in the initial feature sample set X is calculated respectively; Combined with the Mahalanobis distance, the similarity between each sample in the initial feature sample set X in its original high-dimensional space is calculated by Gaussian distribution, and its expression is as follows: Where D ij represents the initial feature sample set X = (X1, X2, ..., X n ) in the i-th sample X i With the jth sample X j Mahalanobis distance between two samples; i = 1, 2…n, j = 1, 2…n, i≠j, n represents the total number of samples; D il Represents sample X i With sample X l The Mahalanobis distance between them, l≠i; σ is a hyperparameter that controls the width of the Gaussian distribution; p ij Represents sample X i With sample X j The similarity between Use X'=(X'1,X'2,…,X' n ) is used to characterize the dimensionality reduction target sample set mapped from the initial feature sample set X to the low-dimensional space, and the similarity between the samples in the sample set X' in the low-dimensional space is represented by a t distribution with a degree of freedom of 1, which is expressed as follows: Where D′ ij Represents the i-th sample X' in the sample set X' i With the jth sample X j 'The Mahalanobis distance between il Represents sample X' i With sample X' l The Mahalanobis distance between ij Represents sample X' i With sample X j 'Similarity between; Based on the similarity p ij and q ij , KL divergence is used to characterize the difference in probability distribution between high-dimensional space and low-dimensional space, and the dimensionality reduction model with the goal of minimizing KL divergence is constructed as follows: minL(D i ' j ) In the formula, L(D i ' j ) represents the Mahalanobis distance D in low-dimensional space i ' j is the KL divergence objective function of the variable to be solved; The dimension reduction model is solved to obtain the final sample set X' after dimension reduction.

5. The GIS partial discharge type identification method according to claim 4 is characterized in that: The calculation formula for calculating the Mahalanobis distance between each sample in the initial feature sample set X based on the covariance matrix S is as follows: In the formula, S -1 represents the inverse matrix of the covariance matrix S; T represents the transpose; Correspondingly, sample X' i With sample X j The Mahalanobis distance D′ ij The calculation formula is as follows: In the formula, S' -1 Represents the inverse matrix of the covariance matrix S'.

6. The GIS partial discharge type identification method according to claim 4 is characterized in that: The algorithm used to solve the dimensionality reduction model is the Adam optimization algorithm; The Adam optimization algorithm includes: based on the objective function L(D i ' j )’s first-order moment estimation and second-order moment estimation are used to iteratively update the learning parameter α to optimize the distribution of each sample in the low-dimensional space and thus obtain the optimal solution.

7. The GIS partial discharge type identification method according to claim 4 or 6, characterized in that: The Adam optimization algorithm is used to solve the dimension reduction model. The specific steps include: Step A: Initialize the first-order moment estimate u1, the second-order moment estimate u2 and the learning parameter α of the dimensionality reduction model; Step B: According to the current objective function L(D′ ij ), updating the first-order moment estimate u1 and the second-order moment estimate u2; Step C: perform bias correction on the updated first-order moment estimate u1 and second-order moment estimate u2; Step D: Based on the corrected first-order estimate u1 and second-order moment estimate u2, the learning parameter α is updated to obtain an updated value of the Mahalanobis distance of each sample in the low-dimensional space; Step E: According to the updated value of the Mahalanobis distance of each sample in the low-dimensional space, determine the current objective function L(D i ' j ) converges or reaches the maximum number of iterations; if not, return to step B; if so, output the distribution of each sample in the low-dimensional space based on the currently updated Mahalanobis distance, so as to obtain the final sample set X' after dimensionality reduction.

8. The GIS partial discharge type identification method according to claim 7 is characterized in that: The formula for updating the first-order moment estimate u1 and the second-order moment estimate u2 is as follows: u1 t =β1·u1 t-1 +(1-β1)·g t Where u1 t and u1 t-1 Respectively represent the first-order moment estimates at the t-th and t-1-th iterations; u2 t and u2 t-1 They represent the second-order moment estimates at the t-th and t-1-th iterations, respectively; t is the number of iterations; β1 and β2 are the coefficients of the first-order moment estimate and the second-order moment estimate, respectively; g t is the objective function L(D′) at the tth iteration ij )’s gradient.

9. The GIS partial discharge type identification method according to claim 8, characterized in that: The expressions for bias correction of the updated first-order moment estimate u1 and second-order moment estimate u2 are as follows: In the formula, and They represent the first-order moment estimate and the second-order moment estimate after bias correction at the t-th iteration, respectively. and They represent the t-th power of coefficients β1 and β2 respectively.

10. The GIS partial discharge type identification method according to claim 9, characterized in that: In step D, the learning parameter α is updated to obtain the calculation formulas of the updated value of the Mahalanobis distance of each sample in the low-dimensional space as follows: In the formula, α t and α t-1 Represent the learning parameters at the tth and t-1th iterations respectively; ε is a set constant; D′ ij (t) and D′ ij (t-1) represents the sample X' at the t-th and t-1-th iterations respectively i With sample X j ' is the Mahalanobis distance between .

11. The GIS partial discharge type identification method according to claim 1, characterized in that: The constructed deep residual shrinkage network DRSN recognition model includes sequentially connected convolutional layers, multiple residual shrinkage modules, batch processing layers, ReLU activation layers, global mean pooling layers and fully connected layers; the structure of each residual shrinkage module includes two branches, the first branch includes a convolutional layer with a convolution kernel size of 1, and the second branch includes a first batch processing layer, a first ReLU activation layer, a first convolutional layer, a second batch processing layer, a second ReLU activation layer, a first convolutional layer, a global average pooling layer, a first fully connected layer, a second fully connected layer and a Sigmoid activation layer; Among them, the input data of the residual shrinkage module is processed by the first batch processing layer, the first ReLU activation layer, the first convolution layer, the second batch processing layer, the second ReLU activation layer and the first convolution layer in sequence, and the first feature output is multiplied by the second feature output after the first feature is processed by the global average pooling layer, the first fully connected layer, the second fully connected layer and the Sigmoid activation layer in sequence, and then added to the convolution feature of the input data after being processed by another convolution layer to obtain the output result of the residual shrinkage module.

12. The GIS partial discharge type identification method according to claim 1, characterized in that: The expression of the SLoss loss function is as follows: In the formula, is the discharge type recognition result of the kth input sample; y k is the actual discharge type corresponding to the kth input sample; n is the total number of samples.

13. A GIS partial discharge type identification device, which runs the GIS partial discharge type identification method according to any one of claims 1 to 12, characterized in that: The device includes The data acquisition module is used to obtain the statistical characteristic data set of the historical PRPD spectrum of partial discharge under various defect types of GIS; A preprocessing module is used to preprocess the acquired statistical feature data set to obtain a sample set after dimensionality reduction; A training optimization module is used to train the constructed deep residual shrinkage network DRSN recognition model using the sample set, and optimize the network parameters of the model with the goal of minimizing the SLoss loss function; The discharge type recognition module is used to load the real-time data of the statistical characteristic quantities of the acquired GIS partial discharge PRPD spectrum into the trained and optimized DRSN recognition model to obtain the recognition result of the GIS partial discharge type.

14. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-8.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images