High-frequency partial discharge feature extraction and fault type identification method

Abstract

The invention relates to the technical field of electric power, in particular to a high-frequency partial discharge feature extraction and fault type identification method. According to the technical scheme, the high-frequency partial discharge feature extraction and fault type identification method comprises the following steps: a, carrying out data processing on a partial discharge signal, including singular value decomposition feature extraction based on a Hankel matrix, feature parameter optimization based on a particle swarm algorithm and normalization processing; b, dynamically optimizing the weight, bias and hyper-parameters of the CNN model through a ChOA algorithm to obtain a ChOA optimization-based convolutional neural network ChOA-CNN model, and performing time sequence feature extraction and fault type identification on the processed data by using the ChOA optimization-based convolutional neural network ChOA-CNN model; and c, evaluating the model performance based on the verification set. According to the method, through multi-index evaluation fusing time domain, frequency domain and nonlinear features, weak features and complex modes of partial discharge signals are deeply mined, and local information loss caused by single feature extraction is avoided.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a method for extracting high-frequency partial discharge features and identifying fault types. Background Art

[0002] Partial discharge (PD) is a discharge phenomenon caused by insulation defects in electrical equipment under high voltage. It can be caused by defects such as metal protrusions, freely moving metal particles, and air gaps within the insulation. Recognizing high-frequency PD signals is crucial for preventing power accidents and improving the operational reliability of power systems. Traditional neural networks often struggle with limited recognition rates when processing PD signals due to the large number of time-frequency features and the high number of redundant features. Therefore, improving the recognition efficiency of PD data has become a major challenge in this field.

[0003] Existing technologies have proposed power partial discharge (PD) identification models based on BP neural networks. These models extract PD data features, achieving improved PD identification accuracy compared to statistical methods. A PD identification model based on extreme learning has also been proposed, effectively identifying PD types. Another PD identification model based on wavelet transforms has also been proposed, utilizing the wavelet transform to process the characteristic parameters of current signals. However, these studies use a single model to extract and process high-frequency PD signal features, making it difficult to accurately capture all the features of PD data, resulting in low accuracy.

[0004] The patent with publication number CN119322242A discloses an intelligent power equipment fault diagnosis system and method based on image recognition and partial discharge feature extraction, which includes the following modules: data acquisition module, signal preprocessing and

[0005] Time-frequency image generation module, hybrid deep learning network module, fault classification and diagnosis module; by combining CNN and Transformer networks, it can simultaneously capture the local and global features of partial discharge signals, improve the accuracy and reliability of detection, and the self-attention mechanism of the Transformer network enables the system to accurately identify effective signals even in a high-noise background, reducing the false alarm rate. The system can monitor the operating status of power equipment in real time, provide immediate fault warnings and diagnostic reports, and is suitable for online monitoring and remote diagnosis scenarios; however, in terms of partial discharge data feature extraction, it adopts a combination of Transformer and CNN, mainly capturing the global feature dependency in the image, improving the recognition ability of complex partial discharge patterns, and mainly focusing on improving the overall recognition and perception capabilities of the model, and still has deficiencies in the extraction of detailed features. Summary of the Invention

[0006] The present invention proposes a method for extracting high-frequency partial discharge features and identifying fault types, which solves the problem in the prior art of using a single model to extract and process high-frequency partial discharge signal features, making it difficult to accurately mine all the features of partial discharge data, resulting in low accuracy.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A method for extracting high-frequency partial discharge features and identifying fault types comprises the following steps:

[0009] a. Data processing of partial discharge signals, including feature extraction based on Hankel matrix singular value decomposition, feature parameter optimization using particle swarm optimization, and normalization processing;

[0010] b. Dynamically optimize the weights, biases, and hyperparameters of the CNN model using the ChOA algorithm to obtain the ChOA-CNN model, a convolutional neural network optimized based on ChOA. This model is used to extract time series features and identify fault types from the processed data.

[0011] c. Evaluate model performance based on the validation set.

[0012] Furthermore, the singular value decomposition based on the Hankel matrix in step a specifically includes:

[0013] a1. Convert the one-dimensional partial discharge signal sequence into a Hankel matrix;

[0014] a2. Perform singular value decomposition on the Hankel matrix, select the first n largest singular values and reconstruct the characteristic matrix;

[0015] a3. Calculate the energy moment based on phase space reconstruction and generate the fault judgment feature vector.

[0016] Furthermore, the fitness function of the particle swarm algorithm for optimizing characteristic parameters in step a is:

[0017]

[0018] Among them, FP + and FP - They are calculated by symbol entropy and peak detection function respectively.

[0019] Furthermore, the normalization process uses the following formula:

[0020]

[0021] Among them, T is the original data, T min and T max are the minimum and maximum values of the data, respectively.

[0022] Furthermore, the optimization of the CNN model by the ChOA algorithm in step b includes:

[0023] b1. Initialize the positions of the chimpanzee population, including the roles of chaser, surrounder, pursuer, and attacker;

[0024] b2. Update the position of each character according to the prey position and adjust the search radius through the chaos map vector;

[0025] b3. Iteratively update the CNN weights, biases, and learning rate parameters with the goal of minimizing the prediction error.

[0026] Furthermore, the ChOA-CNN model includes the following structure:

[0027] The number of input layer nodes is 100;

[0028] Two hidden layers with 60 and 20 nodes respectively;

[0029] The number of nodes in the output layer is 6, corresponding to the fault type classification;

[0030] A joint structure consisting of 1D convolutional layers and 2D convolutional layers is used to extract multi-scale spatiotemporal features.

[0031] Furthermore, the convolutional layer uses a ReLU activation function and a 2×2 maximum pooling operation, and the pooling formula is:

[0032]

[0033] Among them, c is the pooling output; k is the number of pooling layers.

[0034] Furthermore, the output layer uses a Softmax activation function and a regression output layer, and the Softmax formula is:

[0035]

[0036] Among them, z j Represents the output of the j-th neuron, and softmax is the activation function.

[0037] Furthermore, the model performance evaluation indicators in step c include:

[0038] Overall accuracy OA, precision P, recall R and F1 score, where:

[0039]

[0040] Furthermore, the method is applicable to identifying insulation defect types of electrical equipment, including casing suspension, casing spikes, insulator metal impurities, and insulator bubble defects.

[0041] The positive effects of the present invention are: by constructing a multi-dimensional evaluation system and combining it with the feature fusion mechanism of high-frequency partial discharge signals, the present invention effectively solves the problem of low accuracy caused by the difficulty of a single model in comprehensively extracting the characteristics of partial discharge data. By integrating multi-index evaluation (such as precision, recall rate, F1 score) of time domain, frequency domain and nonlinear characteristics, the present invention deeply explores the weak characteristics and complex patterns of partial discharge signals, avoiding the loss of local information caused by single feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a real shot of the partial discharge signal collection site in an embodiment of the present invention;

[0043] Figure 2 This is a waveform diagram of the partial discharge signal before acquisition and processing in an embodiment of the present invention;

[0044] Figure 3 This is a waveform diagram of the partial discharge signal after acquisition and processing in an embodiment of the present invention;

[0045] Figure 4 This is the overall flow chart of the ChOA-CNN model in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the stacking results of two convolutional layers, 1DCNN and 2DCNN, in an embodiment of the present invention;

[0047] Figure 6 Schematic diagram of four typical insulation defects and partial discharge defects in the embodiment of the present invention;

[0048] Figure 7 This is a line graph of the training loss of four models in the embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram comparing the accuracy of four models in an embodiment of the present invention;

[0050] Figure 9 Schematic diagram of confusion matrix of four models in the embodiment of the present invention; DETAILED DESCRIPTION

[0051] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0052] With the development of deep learning, a high-frequency partial discharge (PD) identification model based on a joint neural network has been proposed, combining the characteristics of high-frequency partial discharge (PD) data. First, the Hankel matrix and particle swarm optimization (PSO) algorithm are used to process PD data, improving its usability. Then, a high-frequency partial discharge (PD) identification model based on ChOA-CNN is proposed, using ChOA-CNN to extract the temporal characteristics of PD data.

[0053] Example 1

[0054] A method for extracting high-frequency partial discharge features and identifying fault types comprises the following steps:

[0055] a. Data processing of partial discharge signals, including feature extraction based on Hankel matrix singular value decomposition, feature parameter optimization using particle swarm optimization, and normalization processing;

[0056] b. Dynamically optimize the weights, biases, and hyperparameters of the CNN model using the ChOA algorithm to obtain the ChOA-CNN model, a convolutional neural network optimized based on ChOA. This model is used to extract time series features and identify fault types from the processed data.

[0057] c. Evaluate model performance based on the validation set.

[0058] The singular value decomposition based on the Hankel matrix in step a specifically includes:

[0059] a1. Convert the one-dimensional partial discharge signal sequence into a Hankel matrix;

[0060] a2. Perform singular value decomposition on the Hankel matrix, select the first n largest singular values and reconstruct the characteristic matrix;

[0061] a3. Calculate the energy moment based on phase space reconstruction and generate the fault judgment feature vector.

[0062] The fitness function of the particle swarm algorithm to optimize the characteristic parameters in step a is:

[0063]

[0064] Among them, FP + and FP - They are calculated by symbol entropy and peak detection function respectively.

[0065] The normalization process uses the following formula:

[0066]

[0067] Among them, T is the original data, T min and T max are the minimum and maximum values of the data, respectively.

[0068] The details are as follows:

[0069] Due to the complexity and diversity of partial discharge signals of GIS equipment, a partial discharge feature extraction method based on Hankel matrix singular value decomposition (SVD) is proposed to improve the accuracy and robustness of feature extraction.

[0070] The Hankel matrix is a special matrix whose elements satisfy specific symmetries. It is widely used in data analysis and processing in fields such as signal processing and system identification. For a given partial discharge signal, we first discretely sample it to obtain a series of digital signals.

[0071] These digital signals are then used to construct a Hankel matrix. Next, the Hankel matrix is subjected to singular value decomposition, yielding a singular value matrix and corresponding singular value vectors. The elements of the singular value matrix represent the eigenvalues of the partial discharge signal, while the singular value vectors reflect the distribution of the signal at different frequency components.

[0072] Since there are many elements in the singular value matrix and not all elements contain useful feature information, we need to select an appropriate threshold according to the actual situation, filter the singular values, retain the singular values with larger values, and obtain a filtered singular value matrix.

[0073] The modified Hankel matrix is reconstructed using the selected singular value matrix. This step aims to remove noise and redundant information and improve the accuracy and robustness of feature extraction.

[0074] The definition of singular value decomposition is: for any real matrix A∈R m×n , there must be an orthogonal matrix U=[u1,u2,…,u m ] and the orthogonal matrix V=[v1,v2,…,v n ], so that the following formula holds:

[0075] A=UΛV T

[0076] Where: m and n represent the number of rows and columns of A respectively; Λ∈R m×n , q=min(m,n), when m≤n, Λ=[diag(σ1,σ2,…,σ q ),O], when m>n, Λ=[diag(σ1,σ2,…,σ q ),O] T is a zero matrix; σ1,σ2,…,σ qare the singular values of A.

[0077] First, the one-dimensional signal sequence y(i) is converted into the Hankel matrix H as follows:

[0078]

[0079] Where: Hankel matrix AH∈Rm×n, m and n represent the rows and columns of the matrix AH respectively.

[0080] Phase space reconstruction is a commonly used data dimensionality reduction method that can effectively extract key characteristic information from the signal. By calculating the energy moments of each component of the phase space matrix B, we obtain the eigenvectors for partial discharge fault determination. These eigenvectors contain important characteristic information of the partial discharge signal and can be used for subsequent fault determination and pattern recognition. The singular value decomposition method is shown in the following formula:

[0081]

[0082] Where: matrices U and V are m×m and n×n orthogonal matrices respectively; ui and vi are m- and n-dimensional column vectors respectively.

[0083] Connecting the first row vector Hi,1 of the component matrix Ai with the transpose of the last column vector Li,n, we can get the component signal P of the original signal sequence. i , and finally we get:

[0084]

[0085] The particle swarm optimization algorithm is used to optimize the parameters of the feature extractor, achieving automatic extraction and selection of partial discharge features. The particle swarm optimization algorithm is applied to the parameter optimization of the feature extractor, and the optimal feature extraction parameters are found by iteratively updating the position and velocity of the particles.

[0086] The fitness function is constructed by combining symbolic entropy with peak detection function FPs. The formula is as shown in the formula:

[0087] h V-max =max{h τ -h τ-1 ,h τ -h τ-2 ,…,h τ -h τ-k}

[0088] h V+max =max{h τ -h τ+1 ,h τ -h τ+2 ,…,h τ -h τ+k}

[0089] h V-min =min{h τ -h τ-1 ,h τ -h τ-2 ,…,h τ -h τ-k}

[0090] h V+min =min{h τ -h τ+1 ,h τ -h τ+2 ,…,h τ -h τ+k}

[0091]

[0092]

[0093] Where: τ is the delay parameter; h is the symbol entropy; FP is the detection function.

[0094] The fitness function is a criterion for distinguishing the good and bad positions of particles in a group.

[0095] The fitness function used in this paper is:

[0096]

[0097] In order to improve the recognition accuracy, each data needs to be normalized. The formula is as follows:

[0098]

[0099] Where: T is the original data; T min is the minimum value of the original data; T max is the maximum value of the original data.

[0100] The results after partial discharge signal acquisition and processing are as follows Figure 1-3 shown.

[0101] Example 2

[0102] On the basis of Example 1: in step b, optimizing the CNN model by the ChOA algorithm includes:

[0103] b1. Initialize the positions of the chimpanzee population, including the roles of chaser, surrounder, pursuer, and attacker;

[0104] b2. Update the position of each character according to the prey position and adjust the search radius through the chaos map vector;

[0105] b3. Iteratively update the CNN weights, biases, and learning rate parameters with the goal of minimizing the prediction error.

[0106] The ChOA-CNN model contains the following structure:

[0107] The number of input layer nodes is 100;

[0108] Two hidden layers with 60 and 20 nodes respectively;

[0109] The number of nodes in the output layer is 6, corresponding to the fault type classification;

[0110] A joint structure consisting of 1D convolutional layers and 2D convolutional layers is used to extract multi-scale spatiotemporal features.

[0111] In the early stages of iteration, traditional CNN models require a long iteration step to successfully perform global optimization. However, in the later stages of iteration, a long step will lead to a decrease in local optimization capabilities. To improve this problem, this paper uses the ChOA algorithm to optimize the parameters of the CNN model, reducing the search radius generation by generation, which can effectively solve the problem of being trapped in the local optimal solution. Figure 4 This is the overall flow chart of the ChOA-CNN proposed in this embodiment.

[0112] The details are as follows:

[0113] Initialize weights and biases: Use ChOA to initialize the weights and biases of CNN. These initial parameters will become the starting point of the algorithm.

[0114] Update weights and biases: Use ChOA to search the parameter space of CNN and adjust weights and biases to minimize prediction error. The goal of optimization is to enable CNN to better capture features and patterns in the data.

[0115] Adjusting network hyperparameters: There are many hyperparameters in deep learning models, such as learning rate, batch size, number of hidden layer nodes, etc. Using ChOA, you can search for the optimal hyperparameter combination in the hyperparameter space to achieve better model performance.

[0116] Optimize network structure: By applying ChOA, the number of network layers, the number of nodes in each layer, etc. can be adjusted within a certain range to find a network structure that is more suitable for the problem.

[0117] Training and Validation: The optimized CNN is trained on the training set and validated on the validation set. In each iteration, the loss is calculated based on the predicted and actual results, and the CNN parameters are updated using ChOA.

[0118] Model evaluation: After training and validation, the performance of the optimized CNN is evaluated using the test set. This paper uses metrics such as precision, recall, accuracy, F1 score, and AUC to measure the predictive ability of the model.

[0119] The CNN network structure determined after ChOA optimization consists of a four-layer network: the input layer has 100 nodes; the hidden layers have 2 layers, with 60 and 20 nodes in each hidden layer, respectively; and the output layer has 6 nodes. The learning rate and number of iterations of the DBN network are 0.1 and 60, respectively.

[0120] ChOA is an optimization algorithm based on natural chimpanzee behavior, mimicking the behavioral strategies of chimpanzee groups during hunting and predation. This algorithm is designed to address real-world optimization challenges, particularly when the search location is unknown. ChOA's strengths lie in its simplicity and applicability, allowing it to be applied to a wide range of optimization problems, particularly those with unknown or complex search spaces.

[0121] The basic idea behind ChOA is to apply the different roles of chimpanzees—driver, barrier, chaser, and attacker—to optimization problems, thereby constructing a multi-role search strategy. Each role plays a different role in the problem-solving process, mimicking the different behaviors of chimpanzees. The core idea and basic steps of ChOA are as follows:

[0122] (1) After the roles are defined, the role behavior simulation and initialization are first performed: the chaser role is responsible for observing the target and not chasing it; the surrounder role builds a dam on the tree to hinder the target's progress; the pursuer role quickly chases the target; the attacker role follows the pursuer's direction or moves downward into the lower coverage layer. At this time, the initialization of the chimpanzee population position is shown in the formula:

[0123] G i =random·(upper bound-lower bound)+lower bound

[0124] Where: G i represents the position of the i-th individual; random is a random number between 0 and 1; upper bound and lower bound represent the upper and lower limits of the search space respectively.

[0125] (2) Position update and optimization process: The position of each character is adjusted according to a certain update strategy. During optimization, the algorithm solves the problem through multiple iterations. In each cycle, based on the character's behavior and position update strategy, the algorithm continuously searches the solution space to find the optimal solution. During the exploration and attack process, the mathematical model is expressed by the following formula:

[0126] G classification (t+1)=G EMP (t)-k·d

[0127] d=|B·G EMP (t)-N·G classification (t)|

[0128] Where: G EMP (t) represents the current position vector of the prey; G classification (t) represents the position vector of the chimpanzee when hunting according to the location of the prey; t represents the current number of iterations; N is a chaotic mapping vector with ergodicity and order; it can be changed according to the chaotic mapping situation:

[0129] B=2r1

[0130] Where: B is the impact factor of obstacles on chimpanzees during hunting, and r1 is a random number between 0 and 1.

[0131] The mathematical model for searching for prey is shown in the formula:

[0132]

[0133] Where G(t+1) represents the updated position vector of the chimpanzee; G1, G2, G3, and G4 represent the updated position vectors of the attacker, besieger, chaser, and pursuer, respectively.

[0134] The mathematical model of its attack on prey is shown in the formula:

[0135] G1=G attacker -k1·d attacker

[0136] G2=G barrier -k2·d barrier

[0137] G3=G chaser -k3·d chaser

[0138] G4=G driver -k4·d driver

[0139] Where: G attacker , G barrier , Gchaser , G driver Denote the position vectors of the attacker, besieger, chaser and pursuer respectively; d attacker d barrier d chaser d driver Represents the distance between the attacker, surrounder, chaser and pursuer and the prey respectively.

[0140] The simplified pseudocode of the ChOA algorithm is as follows:

[0141] Step 1: Initialize the populations of drivers, besiegers, pursuers, and attackers, with random distribution of positions;

[0142] Step 2: Initialize parameters such as the maximum number of iterations, convergence threshold, coefficient range, etc.

[0143] Step 3: Within the maximum number of iterations:

[0144] For each driver in the population:

[0145] Calculate the new position of the evictor based on its behavior

[0146] Update the location of the evictor

[0147] Step 3.1: For each surrounder in the population:

[0148] Calculate the new position of the surrounder based on its behavior

[0149] Update the position of the besieger

[0150] Step 3.2: For each pursuer in the population:

[0151] Calculate the pursuer's new position based on its behavior

[0152] Update the pursuer's location

[0153] Step 3.3: For each attacker in the population:

[0154] Calculate the attacker's new position based on his behavior

[0155] Update the attacker's location

[0156] Step 3.4: Evaluate the fitness of each role and select the best solution

[0157] Use the convergence threshold to check for convergence

[0158] If the condition is met, terminate the iteration

[0159] End the loop

[0160] Step 4: Return the best solution found

[0161] The convolution layer uses the ReLU activation function and 2×2 maximum pooling operation. The pooling formula is:

[0162]

[0163] Among them, c is the pooling output; k is the number of pooling layers.

[0164] CNN is a neural network structure inspired by biological vision. Its core idea is to extract image features through convolution and pooling operations. The network structure of CNN usually includes convolutional layers, pooling layers, fully connected layers, and activation functions.

[0165] The convolution operation has the characteristics of local receptive field and weight sharing, which can effectively extract the spatial features of high-frequency partial discharge. In order to achieve effective extraction of spatial features, the paper uses two convolutional layers, 1DCNN and 2DCNN, to stack them. The results are shown in the figure. Figure 5 shown.

[0166] Each convolution layer uses multiple different convolution kernels to extract different features and fuse the feature vectors element-by-element. 1DCNN uses a smaller receptive field to automatically extract shallow features, while 2DCNN uses a larger receptive field to extract deep features, combining the features from 1DCNN and 2DCNN. Both 1DCNN and 2DCNN consist of two groups of convolution and pooling layers, and the activation function is Relu. The pooling layer is a 2×2 maximum pooling to further reduce network parameters. Each convolution kernel performs a convolution operation with a different local window of the input feature. The feature vector obtained by the operation is processed by the nonlinear activation function f to generate the features to be output by this layer. The formula is as follows:

[0167] S t =f(WX+b)

[0168] Where: X is the input vector; W is the convolution kernel; parameter b is the bias; f is the nonlinear activation function.

[0169] The ReLU function is used to speed up the convergence of the loss function. The formula is as follows:

[0170] f=max(0,X)

[0171] Where X represents the output vector of the previous layer. The ReLU function converts negative input values to 0 and leaves positive values unchanged, making the network sparse and mitigating overfitting.

[0172] The pooling layer reduces the size of the feature map by downsampling, further reducing the computational complexity of the network. Common pooling operations include maximum pooling and average pooling. The pooling formula is as follows:

[0173]

[0174] Where: c is the pooling output; k is the number of layers.

[0175] The fully connected layer is used for final classification, and Softmax is used as the activation function of the output layer. The Softmax formula is:

[0176]

[0177] Among them, z j Represents the output of the j-th neuron, softmax is the activation function; finally, a regression output layer is connected as the regression prediction output of the network.

[0178] Example 3

[0179] Based on Example 2: In this embodiment, the experimental environment is carried out in a Python environment on a PC with an Intel i9-14900HX processor and an RTX4090 graphics card with 64GB RAM. The network structure of the experiment is built with the TensorFlow2.10 deep learning framework, and the mean square error (MSE) is used as the loss function to evaluate the network performance. The Adam optimizer is used, and the hyperparameters of various layers are gradually optimized and selected according to the experiment. Using 4 evaluation indicators for evaluating classification performance, the model performance evaluation indicators in step c include:

[0180] Overall accuracy OA, precision P, recall R, F1 score and AUC value, where:

[0181]

[0182] Where: OA is an intuitive indicator of the classification accuracy of each class in the entire sample library, P is the proportion of correct classifications in the model's decision-making, R is a measure of the model's ability to correctly classify, F1 is the harmonic mean of precision and recall; TP, FP, and FN are true positives, false positives, and false negatives.

[0183] The AUC value is introduced to evaluate the stability of the model in a noisy environment (such as electromagnetic interference on the power equipment site), and the integral formula The quantization model is robust to abnormal pulses in high-frequency signals, and the false detection rate is greatly reduced;

[0184] The method is applicable to identifying insulation defect types of electrical equipment, including casing suspension, casing spikes, insulator metal impurities, and insulator bubble defects.

[0185] In this embodiment, four typical insulation defects such as shell suspension, shell spikes, insulator metal, and insulator bubbles are selected. The partial discharge defect is shown as follows. Figure 6As shown in a, b, c, and d.

[0186] In this embodiment, the Hankel matrix and particle swarm algorithm are used to process the partial discharge data, and the first five largest singular values of the partial discharge signal are selected and combined as the defect feature quantity for joint identification.

[0187] Comparison model selection: BP, CNN, LSTM, to verify the recognition effect of the model proposed in this paper. As can be seen from the figure, after 60 iterations of BP, CNN, and LSTM models, all models tend to be stable. In contrast, the model proposed in this paper can be trained with losses of 0.245 and 0.212 respectively after 40 epochs. Figure 7 shown.

[0188] The BP, CNN, and LSTM models are introduced as follows:

[0189] BP: It is a multi-layer feedforward neural network trained according to the error back propagation algorithm, which adjusts the connection weights of neurons in each layer along the way according to the error.

[0190] CNN: Through components such as convolutional layers, pooling layers, and fully connected layers, convolution operations are performed by sliding convolution kernels on the data to mine spatial features between data.

[0191] LSTM: By introducing memory units and gating mechanisms, it can effectively capture long-term dependency information in sequences and solve the problems of gradient disappearance and gradient explosion.

[0192] Model accuracy comparison Figure 8 As shown in the figure, the model mentioned in this embodiment has the highest recognition accuracy. This is because the model proposed in this embodiment uses ChOA-CNN to extract the temporal characteristics and spatial characteristics of the data, which has better recognition performance. The high-frequency partial discharge recognition results are shown in Table 1:

[0193] Table 1 Performance comparison of different models

[0194]

[0195] As shown in Table 1, 2DCNN excels at capturing local detail features, but may have limited ability to capture some global, long-range dependencies. 1DCNN can globally integrate information across a single image dimension (such as width or height) through convolution operations in one dimension, complementing 2DCNN's shortcomings in global feature extraction and effectively improving the accuracy of high-frequency partial discharge recognition.

[0196] Figure 9 It is the confusion matrix of the four networks, which intuitively reflects the classification performance of different classes of the model. Figure 9It can be seen that the model proposed in this paper has the lowest number of false detections and missed detections of high-frequency partial discharges, and the highest number of correct identifications. Obviously, this structure better extracts multi-scale spatiotemporal features, enabling the model to achieve the best discrimination effect.

[0197] The test set is input into the trained neural network, and the recognition results are shown in Table 2.

[0198] Table 2 Identification results of different insulation defects

[0199]

[0200] As shown in Table 2, the recognition rates of the four types of defect discharges in the model proposed in this embodiment are all the highest. This is because the ChOA-CNN is used to extract all the temporal features of the partial discharge data, and the ChOA algorithm is used to optimize the parameters of the CNN model, which reduces the search radius generation by generation. This can effectively solve the problem of falling into the local optimal solution and effectively improve the recognition accuracy.

[0201] This example uses the Hankel matrix and particle swarm optimization algorithm to process partial discharge data, improving its usability. A high-frequency partial discharge identification model based on ChOA-CNN is then proposed, using the ChOA-CNN to extract temporal features from the partial discharge data. Finally, experiments validate the model's effectiveness. Future work will focus on further optimizing the structure and parameter settings of the joint neural network to improve its robustness to complex environments and multi-source interference.

[0202] The above-mentioned embodiments are described in a relatively detailed and specific manner, expressing preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. However, they are not limited to the present invention alone, and the patent scope of the present invention cannot be limited solely by these embodiments. That is, any equivalent changes or modifications made to the spirit disclosed by the present invention, for researchers or technicians in this field, without departing from the structure of the present invention, local improvements within the system and changes and conversions between subsystems, etc., are still within the patent scope of the present invention.

Claims

1. A method for extracting high-frequency partial discharge features and identifying fault types, characterized in that: The following steps are involved: a. Perform data processing on partial discharge signals, including feature extraction based on Hankel matrix singular value decomposition, feature parameter optimization using particle swarm optimization, and normalization processing; b. Dynamically optimize the weights, biases, and hyperparameters of the CNN model using the ChOA algorithm to obtain the ChOA-CNN model, a convolutional neural network optimized based on ChOA. This model is used to extract time series features and identify fault types from the processed data. c. Evaluate model performance based on the validation set.

2. A method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, characterized in that: The singular value decomposition based on the Hankel matrix in step a specifically includes: a1. Convert the one-dimensional partial discharge signal sequence into a Hankel matrix; a2. Perform singular value decomposition on the Hankel matrix, select the first n largest singular values and reconstruct the characteristic matrix; a3. Calculate the energy moment based on phase space reconstruction and generate the fault judgment feature vector.

3. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, wherein: The fitness function of the particle swarm algorithm to optimize the characteristic parameters in step a is: Among them, FP + and FP - They are calculated by symbol entropy and peak detection function respectively.

4. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, wherein: The normalization process uses the following formula: Among them, T is the original data, T min and T max are the minimum and maximum values of the data, respectively.

5. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, wherein: Optimizing the CNN model using the ChOA algorithm in step b includes: b1. Initialize the positions of the chimpanzee population, including the roles of chaser, surrounder, pursuer, and attacker; b2. Update the position of each character according to the prey position and adjust the search radius through the chaos map vector; b3. Iteratively update the CNN weights, biases, and learning rate parameters with the goal of minimizing the prediction error.

6. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, characterized in that: The ChOA-CNN model contains the following structure: The number of input layer nodes is 100; Two hidden layers with 60 and 20 nodes respectively; The number of nodes in the output layer is 6, corresponding to the fault type classification; A joint structure consisting of 1D convolutional layers and 2D convolutional layers is used to extract multi-scale spatiotemporal features.

7. A method for extracting high-frequency partial discharge features and identifying fault types according to claim 6, characterized in that: The convolutional layer uses the ReLU activation function and 2×2 maximum pooling operation. The pooling formula is: Among them, c is the pooling output; k is the number of pooling layers.

8. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 6, characterized in that: The output layer uses the Softmax activation function and regression output layer, and the Softmax formula is: Among them, z j Represents the output of the j-th neuron, and softmax is the activation function.

9. The method for extracting high-frequency partial discharge features and identifying fault types according to claim 1, characterized in that: The model performance evaluation indicators in step c include: Overall accuracy OA, precision P, recall R and F1 score, where:

10. A method for extracting high-frequency partial discharge features and identifying fault types according to any one of claims 1 to 9, characterized in that: The method is applicable to identifying insulation defect types of electrical equipment, including casing suspension, casing spikes, insulator metal impurities, and insulator bubble defects.

Images