A neural network partial discharge identification method fusing polarized particle swarm optimization

Abstract

The application belongs to the technical field of partial discharge detection of power equipment, and specifically discloses a neural network partial discharge identification method fusing a polarization particle swarm optimization. The application adopts a structure parameter particle swarm and a threshold parameter particle swarm to group and cooperatively optimize scale parameters, translation parameters and unified soft thresholds of a wavelet kernel function, and takes a no-reference fitness function composed of pulse energy concentration and pulse envelope form retention as a target to realize the balance between the partial discharge pulse feature retention and the noise suppression under the condition of lacking a noise-free reference signal, and improves the image construction quality. Meanwhile, the application designs a residual neural network fusing a multi-path feature enhancement, a learnable local texture attention module and a light-weight selection kernel attention module, which can simultaneously model the local texture, edge transition, multi-scale structure and global distribution of the image, improve the discrimination ability and generalization performance of the network on the partial discharge image, and guarantee the accuracy and robustness of the identification.

Description

Technical Field

[0001] This invention belongs to the field of partial discharge detection and intelligent signal processing technology for power equipment, specifically relating to a neural network partial discharge identification method that integrates polarized particle swarm optimization. Background Technology

[0002] Partial discharge (PD) is a common insulation defect in high-voltage electrical equipment. Its long-term development may lead to aging or even breakdown of insulation materials, thereby endangering the safe operation of the equipment.

[0003] Therefore, the effective detection and identification of partial discharge signals has always been an important research direction in the condition monitoring of power equipment.

[0004] Existing methods typically involve wavelet denoising and image processing of the original signal, followed by pattern recognition using classification algorithms. Among these methods, wavelet transform has become the mainstream signal preprocessing technique due to its excellent time-frequency localization capabilities.

[0005] However, traditional wavelet thresholding methods often rely on fixed or empirically set thresholds, which often make it difficult to balance noise suppression and feature preservation when faced with strong noise and non-stationary partial discharge signals.

[0006] Meanwhile, the lack of a global optimization mechanism for key parameters such as wavelet basis functions and thresholds leads to unstable processing results.

[0007] On the other hand, most existing recognition models use single-path or simple convolutional networks, which are insufficient for extracting multi-scale details and global patterns from partial discharge signal images, making it difficult to meet the feature expression needs under complex working conditions.

[0008] Therefore, there is an urgent need to propose a new method that combines adaptive optimization and feature enhancement residual networks to achieve efficient denoising, feature enhancement and accurate identification of partial discharge signals. There is no corresponding solution in the existing technology.

[0009] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0010] To address the aforementioned problems in existing technologies, this invention proposes a neural network partial discharge identification method that integrates polarization particle swarm optimization, in order to achieve efficient denoising, feature enhancement, and accurate identification of partial discharge signals.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A neural network partial discharge identification method incorporating polarization particle swarm optimization includes the following steps:

[0013] Step 1. Automatically parse the hexadecimal files output by the original partial discharge monitoring equipment; each original signal file contains a complete instantaneous partial discharge waveform signal, and data cleaning is performed;

[0014] Step 2. A method combining wavelet filtering and grouped cooperative polarization particle swarm optimization is proposed to denoise and reconstruct the instantaneous waveform signal of partial discharge after data cleaning, and convert it into a two-dimensional image;

[0015] Step 3. Construct a partial discharge identification model, which includes an input layer, a backbone feature extraction network, a global average pooling layer, and an output layer; the backbone feature extraction network includes four consecutive residual feature extraction stages;

[0016] Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism;

[0017] Building upon this foundation, a learnable local texture attention module and a lightweight selection kernel attention module are further introduced to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features.

[0018] Step 4. Based on the two-dimensional image data after denoising signal reconstruction and image conversion processing in Step 2, train the partial discharge recognition model built in Step 3, and use the trained model to identify partial discharge.

[0019] Furthermore, based on the aforementioned neural network partial discharge identification method optimized by fusion polarization particle swarm optimization, this invention also proposes a corresponding neural network partial discharge identification system optimized by fusion polarization particle swarm optimization, the scheme of which is as follows:

[0020] A neural network partial discharge identification system incorporating polarization particle swarm optimization includes the following steps:

[0021] The data acquisition module is used to automatically parse the hexadecimal files output by the original partial discharge monitoring equipment; each original signal file contains a complete instantaneous waveform signal of partial discharge, and data cleaning is performed.

[0022] The denoising signal reconstruction and image conversion module is used to propose a method that combines wavelet filtering and grouped cooperative polarization particle swarm optimization to perform denoising signal reconstruction and image conversion on the obtained instantaneous waveform signal of partial discharge.

[0023] And a partial discharge identification module, used to build a partial discharge identification model, which includes an input layer, a backbone feature extraction network, a global average pooling and an output layer; wherein the backbone feature extraction network includes four consecutive residual feature extraction stages;

[0024] Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; wherein, each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism;

[0025] Building upon this foundation, a learnable local texture attention module and a lightweight selection kernel attention module are further introduced to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features.

[0026] Specifically, based on the image data after denoising signal reconstruction and image conversion processing, the constructed partial discharge recognition model is trained, and the trained model is used to identify partial discharge.

[0027] The present invention has the following advantages:

[0028] As described above, this invention proposes a neural network partial discharge identification method that integrates polarization particle swarm optimization (PPSO). This method, targeting the obtained instantaneous waveform signal of partial discharge, proposes a method combining wavelet filtering and grouped collaborative polarization particle swarm optimization to achieve signal denoising, reconstruction, and image conversion. Specifically, this invention uses structural parameter particle swarm optimization and threshold parameter particle swarm optimization to perform grouped collaborative optimization of the scale parameter, translation parameter, and unified soft threshold of the wavelet kernel function. Using a referenceless fitness function composed of pulse energy concentration and pulse envelope morphology preservation as the objective, it achieves a balance between preserving partial discharge pulse features and suppressing noise under the condition of lacking a noiseless reference signal, thereby improving the image reconstruction quality and possessing good global optimization and boundary control capabilities. Meanwhile, addressing the issues of edge weakening, detail blurring, texture non-uniformity, and local noise interference in partial discharge images, this invention designs a residual neural network that integrates multi-path feature enhancement, a learnable local texture attention module (LTA), and a lightweight selection kernel attention module (LSK). The LTA module enhances local texture details and impulse edge responses, while the LSK module enables adaptive selection and fusion of different receptive field features. This allows for simultaneous modeling of local texture, edge transitions, multi-scale structures, and global distribution of the image, significantly improving the network's discrimination ability and generalization performance for partial discharge images, further ensuring the accuracy and robustness of recognition. This invention achieves accurate partial discharge recognition by performing parametric wavelet transform, PPSO adaptive threshold denoising, and image conversion on the original hexadecimal discharge signal, followed by deep neural network classification training. Attached Figure Description

[0029] Figure 1 This is a flowchart of the neural network partial discharge recognition method with fusion polarization particle swarm optimization in Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart of the polarized particle swarm optimization (PPSO) algorithm in Embodiment 1 of the present invention;

[0031] Figure 3 This is a flowchart illustrating the working principle of the adaptive wavelet denoising module (AWD) in Embodiment 1 of the present invention.

[0032] Figure 4 This is a flowchart of partial discharge signal preprocessing and image conversion in Embodiment 1 of the present invention;

[0033] Figure 5 This is a schematic diagram of the overall architecture of the partial discharge identification module in Embodiment 1 of the present invention;

[0034] Figure 6This is a schematic diagram of the internal structure of the multi-path feature fusion residual unit in Embodiment 1 of the present invention;

[0035] Figure 7 This is a schematic diagram of the Learnable Local Texture Attention Module (LTA) in Embodiment 1 of the present invention;

[0036] Figure 8 This is a schematic diagram of the Lightweight Selective Kernel Attention (LSK) module in Embodiment 1 of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0038] Example 1

[0039] This embodiment describes a neural network partial discharge identification method that integrates polarization particle swarm optimization. By performing parametric wavelet transform, PPSO adaptive threshold denoising, and image conversion on the original hexadecimal discharge signal, followed by deep neural network classification training, accurate identification of partial discharge patterns is achieved. Compared to existing technical solutions that rely on fixed parameters or static network structures, this invention significantly improves signal reconstruction, feature extraction capability, and classification accuracy.

[0040] like Figure 1 As shown, a neural network partial discharge identification method integrating polarization particle swarm optimization includes the following steps:

[0041] Step 1. Automatically parse the hexadecimal files (.txt format) output by the original partial discharge monitoring equipment. Each original signal file contains a complete instantaneous partial discharge waveform signal. Data cleaning is performed as follows:

[0042] Step 1.1. Read the file content as the string .hex_data, retain the valid data segment in the middle, remove the device header and footer identifiers, and only keep the hexadecimal characters 0-9 and the letters A-F.

[0043] Step 1.2. Divide the hexadecimal string cleaned in Step 1.1 into groups of two characters each, and convert them into decimal integers between 0 and 255 to form a one-dimensional signal sequence.

[0044] Step 1.3. Denote the one-dimensional signal sequence as... .

[0045] in For signal length, For the first The signal amplitude at each sampling point; further, the one-dimensional signal sequence is converted into a three-dimensional tensor structure required for the input convolution operation: .

[0046] in, The original one-dimensional partial discharge signal sequence The tensor quantization representation has the following dimensions: the first dimension is the batch size; the second dimension is the number of channels (i.e., a single channel); the third dimension... This is the signal length.

[0047] Step 2. To achieve adaptive parameter control of partial discharge (PD) signals during wavelet denoising, a method combining wavelet filtering and grouped collaborative polarization particle swarm optimization is proposed to denoise and reconstruct the instantaneous waveform signal of partial discharge after data cleaning, and convert it into a two-dimensional image.

[0048] This method jointly optimizes the scale parameter, translation parameter, and unified soft threshold of the wavelet kernel function. The polarized particle swarm optimization (PPSO) algorithm uses angle perturbation for high-dimensional parameter search, such as... Figure 2 As shown.

[0049] Unlike existing technologies that directly use the mean square error between the noiseless reference signal and the reconstructed signal as the optimization target, this invention addresses the characteristic that partial discharge engineering signals typically lack a noiseless reference signal, and proposes the following design:

[0050] An adaptive no-reference fitness function is constructed, and the parameters to be optimized are divided into structural parameter group and threshold parameter group. The parameter search is completed through group collaboration, thereby improving the noise suppression capability while ensuring the preservation of partial discharge pulse characteristics.

[0051] In summary, this invention first defines a 17-dimensional optimization parameter vector, dividing it into a structural parameter group and a threshold parameter group, which are then searched collaboratively by the structural parameter particle swarm and the threshold parameter particle swarm, respectively. Each parameter is constrained within a preset parameter range through an "angle → actual parameter" mapping, thereby achieving bounded search and improving search stability. Each iteration process is as follows:

[0052] First, the angle variables of the two sets of particles are mapped to actual parameters, and a complete candidate parameter combination is constructed by pairing them. Then, the candidate parameter combination is input into the Adaptive Wavelet Denoise Module (AWD) to complete wavelet kernel convolution, soft thresholding denoising and signal reconstruction.

[0053] Subsequently, based on the pulse energy concentration evaluation term and the pulse envelope shape preservation evaluation term, the adaptive no-reference fitness value corresponding to the candidate parameter combination is calculated, and the individual optimal state of each particle in the structure parameter particle swarm and the threshold parameter particle swarm, as well as the population optimal state of each of the two particle swarms, are updated respectively.

[0054] Finally, based on the dynamic inertia weight, random perturbation, the individual optimal state, and the group optimal state, the angle variables of each particle in the structural parameter particle swarm and the threshold parameter particle swarm are updated.

[0055] Repeat the above process until the number of iterations reaches the preset maximum number of iterations, then output the final optimal combination of parameters and complete the denoised signal reconstruction and image conversion.

[0056] The following section provides a detailed explanation of the method combining wavelet filtering and grouped cooperative polarization particle swarm optimization.

[0057] Step 2.1. Optimize parameter definitions, grouping, and initial particle swarm configuration.

[0058] Use a parameterized wavelet kernel with a quantity of n=8. (For its specific function form, see step II of step 2.4).

[0059] Define the scale parameter vector ,scope Translation parameter vector ,scope Unified soft threshold .

[0060] The overall optimization parameter vector is constructed as follows: .

[0061] To improve parameter search efficiency, the 17-dimensional parameters are divided into two parameter groups: .in For structural parameter groups, This is a set of threshold parameters.

[0062] During the initialization phase, the total size of the particle swarm is set to... And set the structure parameter particle swarm size to respectively. Threshold parameter: particle swarm size ,in In this context, each particle in the structure parameter particle swarm corresponds to a set of candidate solutions for structure parameters, and each particle in the threshold parameter particle swarm corresponds to a set of candidate solutions for threshold parameters.

[0063] Step 2.2. Angle encoding, parameter mapping and particle swarm state symbol definition.

[0064] To enhance boundary control capabilities and improve search stability, this invention uses angle encoding to represent the parameters to be optimized, and converts angle variables into actual parameter values ​​through a monotonic mapping of "angle → actual parameter".

[0065] set up Index for parameter dimensions, ;No. The lower and upper bounds of the dimension parameter are respectively and Its corresponding angle variable is Then we have: .

[0066] The mapping relationship between parameters and angle variables is as follows: .

[0067] in, Indicates the first Dimensional actual parameter values.

[0068] For the structure parameter particle swarm, the first The particle, its first The angle vector of the wheel is denoted as For the threshold parameter particle swarm, the th The particle, its first The angle vector of the wheel is denoted as .

[0069] As of the In each iteration, the historical optimal angle vectors of each particle in the structure parameter particle swarm and the threshold parameter particle swarm are denoted as follows: The two groups of particles were in the first... The optimal angle vectors of the group in each iteration are denoted as follows: .

[0070] Step 2.3. Definition of original signal, reconstructed signal and adaptive no-reference fitness function.

[0071] After completing the optimization parameter definition, grouping and initial configuration of the particle swarm in step 2.1, and the angle encoding and parameter mapping rule definition in step 2.2, the original noisy partial discharge signal, the reconstructed signal corresponding to the candidate parameter combination and the adaptive no-reference fitness function are further defined as the evaluation basis for subsequent grouped cooperative polarization particle swarm optimization.

[0072] Let the original partial discharge noisy signal sequence obtained in step 1.3 be... ,in, Let be the total number of signal sampling points; its corresponding tensor quantization representation is denoted as . This is used for convolution operations in the subsequent adaptive wavelet denoising module.

[0073] For any candidate parameter combination The reconstructed signal is obtained after processing by the adaptive wavelet denoising module:

[0074] .

[0075] To characterize the performance of candidate parameter combinations in terms of both noise suppression and partial discharge pulse feature preservation in the absence of a noiseless reference signal, this invention constructs an adaptive no-reference fitness function:

[0076] .

[0077] in, This is an evaluation item for pulse energy concentration. For the evaluation item of pulse envelope morphology preservation, and To combine with candidate parameters The corresponding adaptive weights satisfy:

[0078] .

[0079] This invention aims to minimize the adaptive no-reference fitness function as the cooperative optimization objective, namely: .

[0080] Step 2.4. After completing the preliminary configurations described in Steps 2.1 and 2.2, and based on the collaborative optimization objective defined in Step 2.3, perform grouped collaborative polarization particle swarm optimization iterations on the structural parameter particle swarm and the threshold parameter particle swarm.

[0081] Iterate to minimize the adaptive no-reference fitness function With the goal of updating the angle variables, cooperative fitness, individual optimal state and group optimal state of the two groups of particles through multiple rounds of iteration, until the number of iterations reaches the preset maximum number of iterations.

[0082] A single iteration specifically includes the following steps:

[0083] Step I. Constructing candidate parameter combinations.

[0084] The angle vectors of each particle in the structure parameter particle swarm and the threshold parameter particle swarm are converted into actual parameter values ​​through the mapping relationship in step 2.2, respectively, to obtain the first... Candidate solutions for structural parameters under round iteration and threshold parameter candidate solutions .

[0085] in, Indicates the first In the particle swarm of wheel structure parameters, the first Candidate solutions for the actual structure parameters of each particle. Indicates the first In the first round of threshold parameter particle swarm optimization Candidate solutions for the actual threshold parameters of each particle.

[0086] In evaluating the structure parameters of particle swarm optimization, the first When there are 1 particle, compare it with the current threshold parameter particle swarm's optimal actual threshold parameter. Pairing to form candidate parameter combinations: .

[0087] In evaluating the threshold parameter of particle swarm optimization, the first When there are 1 particle, compare it with the swarm optimal actual structure parameter of the particle swarm with the current structure parameter. Pairing to form candidate parameter combinations: .

[0088] and The first The current population optimal actual parameters of the particle swarm with wheel structure parameters and threshold parameters.

[0089] Step II. Adaptive wavelet denoising and signal reconstruction.

[0090] The candidate parameter combination constructed in step I is input into the adaptive wavelet denoising module for signal processing, such as... Figure 3 As shown.

[0091] The wavelet scaling parameters, translation parameters, and unified soft threshold involved in step II are all actual parameters obtained from the angle variable mapping in step I. To perform multi-scale, multi-location time-frequency analysis of the original signal, this invention employs a convolution kernel group based on the parameterized Ricker wavelet function for feature extraction.

[0092] Let the time discrete grid be ( (This refers to the filter length).

[0093] Then the first The small parameterized wavelet kernel is defined as follows: .in, and They represent the first The scaling and translation parameters of each parameterized wavelet kernel have their ranges as shown in step 2.1.

[0094] Perform on each wavelet kernel Norm normalization yields: .

[0095] in, For the index variable in the normalization summation process, Represents the first in the discrete-time grid One sampling point; and They belong to the same discrete-time grid, only their index variables are different.

[0096] By normalized wavelet kernel With the original signal tensor Performing discrete one-dimensional convolution operation yields the first... Wavelet coefficient components: , .

[0097] in , This represents the discrete one-dimensional convolution operation. This represents the transposed convolution reconstruction operation corresponding to the wavelet kernel. The components are then stacked, as shown in the following formula: .

[0098] in, This indicates that the wavelet coefficient components are stacked along the channel dimension.

[0099] Then, a unified soft threshold operator is introduced for noise reduction: .

[0100] in: Indicated by threshold A unified soft thresholding operator with parameters is used to perform amplitude compression on the coefficients obtained from wavelet decomposition; This represents a single coefficient obtained from wavelet decomposition; Represents a symbolic function.

[0101] The denoised coefficient tensor is obtained after soft thresholding: .remember for The Each channel is connected to the normalized wavelet kernel. The corresponding transpose convolution reconstruction operator yields the reconstructed signal: .

[0102] in, This represents the transpose convolution reconstruction operation based on the normalized wavelet kernel.

[0103] Step III. Calculation of adaptive no-reference fitness value.

[0104] The reconstructed signal obtained in step II is evaluated based on the adaptive no-reference fitness function defined in step 2.3. For any candidate parameter combination... Let the original partial discharge noise signal be... With reconstructed signal .

[0105] in, , This represents the total number of signal sampling points. First, an analytic signal is constructed based on the Hilbert transform, defining the original signal. With reconstructed signal The envelopes are as follows:

[0106] , ;in, Represents the Hilbert transform. Represents the imaginary unit. Represents the magnitude of the analytic signal; Represents the original envelope. This indicates the reconstruction of the envelope.

[0107] , ;in Represents the Hilbert transform. Represents the imaginary unit. Represents the magnitude of the analytic signal; Represents the original envelope. Indicates the reconstruction of the envelope;

[0108] Further calculations were made of the mean and standard deviation of the envelope:

[0109] ;

[0110] ;

[0111] ;

[0112] .

[0113] in , These are the mean values ​​of the original envelope and the reconstructed envelope, respectively. , Let be the standard deviations of the original envelope and the reconstructed envelope, respectively; define the original impulse candidate support set and the reconstructed impulse candidate support set as follows:

[0114] , .

[0115] in, The candidate pulse extraction coefficient is denoted as .

[0116] To avoid the support set being empty in extreme cases, if Then let .

[0117] like Then let .

[0118] Define the pulse energy percentage of the reconstructed signal. for: .

[0119] in, To prevent extremely small positive numbers with a denominator of zero.

[0120] Therefore, the first evaluation item is defined. for:

[0121] .

[0122] Furthermore, in the original pulse candidate support set The correlation coefficient between the original envelope and the reconstructed envelope is defined above. for:

[0123] .

[0124] in . and They represent the support sets respectively. The mean of the original envelope and the reconstructed envelope; Represents a set The number of elements in the middle.

[0125] Define the second evaluation item for:

[0126] .

[0127] To adaptively adjust the weights based on the two evaluation results, an adaptive weight is defined. and for:

[0128] , .

[0129] in, To prevent extremely small positive numbers with a denominator of zero; and These represent the evaluation items. and The corresponding adaptive weights satisfy: .

[0130] Thus, the candidate parameter combinations are obtained. The corresponding adaptive no-reference fitness function:

[0131] .

[0132] in, Indicates candidate parameter combinations The smaller the value, the better the overall performance of the candidate parameter combination in terms of noise suppression capability and partial discharge pulse characteristic preservation capability.

[0133] For the structure parameter particle swarm, the first When evaluating the cooperative fitness of a particle, it is paired with the swarm's optimal actual threshold parameter to form a candidate parameter combination: Its corresponding cooperative fitness Defined as: ;in Indicates the first In the particle swarm of wheel structure parameters, the first Candidate solutions for the actual structure parameters of each particle. Indicates the first The optimal actual threshold parameter for the particle swarm.

[0134] For the threshold parameter particle swarm, the th When evaluating the cooperative fitness of a particle, it is paired with the swarm's optimal actual structural parameters to form candidate parameter combinations: Its corresponding cooperative fitness Defined as: ; Indicates the first In the first round of threshold parameter particle swarm optimization Candidate solutions for the actual threshold parameters of each particle. Indicates the first The swarm optimal actual structural parameters of the particle swarm for the wheel structure parameters.

[0135] Step IV. Individual and group optimal updates.

[0136] After obtaining the cooperative fitness value calculated in step III, individual optimal state and group optimal state updates are performed on the particle swarm with structural parameters and the particle swarm with threshold parameters, respectively.

[0137] in, This represents the particle index in the particle swarm, which is a structural parameter. This represents the particle index in the particle swarm that represents the threshold parameter. and These represent the number of particles in the structure parameter particle swarm and the threshold parameter particle swarm, respectively.

[0138] set up Indicates as of the date In the first iteration of the structure parameter particle swarm, the first... The historical optimal fitness value of each particle. Indicates as of the date During the first iteration, the threshold parameter in the particle swarm is the first... The historical optimal fitness value of each particle; and These represent the corresponding individual optimal angle vectors.

[0139] For the structure parameter particle swarm, the first A particle, if its current cooperative fitness satisfies Then, update its individual optimal angle vector and individual optimal fitness as follows: , .

[0140] Otherwise, the individual's optimal state from the previous round remains unchanged.

[0141] For the threshold parameter particle swarm, the th A particle, if its current cooperative fitness satisfies Then, update its individual optimal angle vector and individual optimal fitness as follows: , .

[0142] Otherwise, the individual's optimal state from the previous round remains unchanged.

[0143] After updating the individual optimal states of the two groups of particles, the optimal state of each particle swarm is further determined in the th... The optimal state of the swarm in round iteration. Let the index of the current optimal particle in the structure parameter particle swarm be: .

[0144] Then the structure parameter particle swarm is in the th The optimal angle vector for the population is updated in each iteration as follows: .

[0145] Let the index of the current best particle in the threshold parameter particle swarm be: Then the threshold parameter particle swarm is in the th The optimal angle vector for the population is updated in each iteration as follows: .

[0146] The updated group optimal angle vector and Do not convert the "angle → actual parameter" mapping relationship described in step 2.2 into actual parameter values ​​to obtain the structural parameter particle swarm and threshold parameter particle swarm in the th step. The population-optimal actual parameters in each iteration are denoted as follows: and .

[0147] Therefore, the first is defined The optimal combination of parameters for wheel coordination is: ;

[0148] Its corresponding cooperative optimal fitness is: .

[0149] Step V. Dynamic update of angle variables and reactivation of stagnation.

[0150] After determining and recording individual and particle swarm optimalities, the angle variable of the particle swarm parameter is dynamically updated to ensure a balance between exploration and development in the algorithm. Indicates the category of parametric particle swarms, corresponding to structural parameter particle swarms and threshold parameter particle swarms respectively; Indicates particle index, Indicates the current iteration round.

[0151] First, define the first The dynamic inertia weight for each iteration is: .in, The maximum number of iterations, and These represent the maximum and minimum values ​​of the dynamic inertia weight, respectively.

[0152] Subsequently, random perturbation vectors are generated respectively. , : .in, Indicates that it is defined in the interval A uniform distribution on the surface. Let... Represents the parameterized particle swarm The Middle The particle in the first Angle increment in round iteration, This represents its current angle vector. Indicates that it is in the first The optimal angle vector for each individual in the round of iteration. Represents the parameterized particle swarm In the The optimal angle vector for the population in each iteration. Then the angle increment update for this particle is:

[0153] .

[0154] in, and These represent the individual learning acceleration coefficient and the group learning acceleration coefficient, respectively. This indicates element-wise multiplication.

[0155] Therefore, the angle variable is updated as follows: .

[0156] Since the angle variable needs to remain within the valid range defined in step 2.2 Therefore, boundary clipping is performed on the updated angle variable: .

[0157] in, Indicates the vector Each component is trimmed to the interval element by element. Inside.

[0158] To avoid the algorithm getting trapped in local optima, when the cooperative optimal fitness is... In continuous If no improvement occurs within a round of iterations, then the worst fit is selected. A restricted phase perturbation is applied by proportional particles. Let... Indicates the application of parameters to the particle swarm. The Middle The phase perturbation vector of each particle is then: .

[0159] in, Indicates that it is defined in the interval Uniform distribution on Indicates the amplitude of the phase disturbance. Indicates the threshold for determining stagnation. This indicates the proportion of particles involved in reactivation.

[0160] For particles that meet the reactivation condition, their angle variable is updated as follows: .

[0161] After completing the dynamic update of the angle variables and the stagnation reactivation, the structural parameter particle swarm and the threshold parameter particle swarm enter the next round of iteration and repeat steps I to V until the number of iterations reaches the preset maximum number of iterations.

[0162] Step 2.5. After completing the dynamic update of the angle variable and the stagnation reactivation, the structural parameter particle swarm and the threshold parameter particle swarm enter the next iteration, and repeat steps I to V in step 2.4 in sequence until the number of iterations reaches the preset maximum number of iterations. .

[0163] After the iteration terminates, the particle swarm optimization algorithm stops searching and outputs the final optimal angle vector for the swarm. and Subsequently, through the angle-parameter mapping relationship described in step 2.2, the... and Converted into the final population optimal actual parameters respectively and This yields the optimal combination of parameters for collaboration, denoted as: .

[0164] The cooperative optimal parameter combination Meanwhile, the final configuration of the adaptive wavelet denoising module was determined.

[0165] After the iteration terminates, the adaptive wavelet denoising module described in step II of step 2.4 is used, and... The original noisy partial discharge signal is finally denoised and reconstructed using the parameters to obtain the optimal denoised signal: .

[0166] Step 2.6. To adapt to the input of the deep neural network, the signal reconstructed by the AWD module needs to be converted into a two-dimensional image format, such as... Figure 4 As shown. The specific steps are as follows:

[0167] (1) The optimal denoised signal sequence A one-dimensional vector is formed according to the sampling order.

[0168] (2) Normalize it to the interval [0, 255] and round it to the nearest integer.

[0169] (3) Calculate the length and width of the image.

[0170] To minimize the irregular shape of the matrix, if the actual number of points is insufficient, padding with 0 is used.

[0171] (4) Rearrange the signal data into a two-dimensional matrix.

[0172] (5) Use the cv2.resize() function in the OpenCV library to scale the image size to 224×224, and use bilinear interpolation (INTER_LINEAR) as the interpolation method.

[0173] (6) Save as a PNG grayscale image, and name the image the same as the original .txt file.

[0174] The final image obtained through the above processing can be directly used as input data for deep learning models.

[0175] Step 3. Build a partial discharge identification model, the network structure of which is as follows: Figure 5 As shown, it is mainly divided into three parts: the input layer, the backbone feature extraction network, and the global average pooling and classification output part.

[0176] The backbone feature extraction network consists of four consecutive residual feature extraction stages.

[0177] Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; wherein each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism.

[0178] Building upon this foundation, a learnable local texture attention module (LTA) and a lightweight selection kernel attention module (LSK) are further introduced to enhance local texture details, model global scale information, and aggregate and enhance multi-level semantic features.

[0179] This invention achieves local texture detail enhancement and multi-scale feature adaptive selection through the LTA module and LSK module, respectively, thereby realizing efficient modeling, fusion and multi-level semantic information aggregation of multi-scale features within the global framework.

[0180] The input layer primarily receives the partial discharge image generated after polarization particle swarm optimization wavelet thresholding denoising. The input image first undergoes a convolution operation with a 7×7 kernel, a stride of 2, and padding of 3, transforming the number of channels from 1 to 64, which is used to extract low-level edge contours and detailed textures. The convolution output is then processed by a batch normalization layer to standardize the feature distribution, eliminating statistical differences between batches and improving training stability and convergence efficiency.

[0181] Subsequently, the ReLU function is introduced to enhance the nonlinear expressive power of the model, and a max pooling layer with a kernel size of 3×3, a stride of 2, and a padding of 1 is used to spatially compress the image, thereby providing suitable receptive field and computational resources for subsequent deep feature extraction. The final output is the feature map obtained from the initial convolution and pooling.

[0182] To improve the accuracy and robustness of partial discharge image recognition, this invention designs a residual neural network structure that integrates multi-path feature enhancement, learnable local texture attention, and lightweight selection kernel attention.

[0183] This residual neural network architecture constructs multi-scale, multi-directional parallel enhancement paths and introduces a feature fusion mechanism into the residual connections. Building upon this, the residual neural network architecture further incorporates LTA and LSK modules to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features. This improves the feature representation ability, classification accuracy, and recognition robustness of partial discharge images, enabling high-precision classification of partial discharge images.

[0184] Each residual feature extraction stage contains 3, 4, 6, and 3 multi-path feature fusion residual units, respectively, with output channel numbers of 64, 128, 256, and 512. The multi-path feature fusion residual unit, as shown... Figure 6 As shown. Among them. Figure 6 In this context, Conv represents convolution, DConv represents dilated convolution, BN represents batch normalization, ReLU represents non-linear activation function, LTA represents local texture attention, LSK represents kernel selection attention, and Concat represents channel-level concatenation.

[0185] The multi-path feature fusion residual unit includes a main path and three enhancement branches. In the structure of this multi-path feature fusion residual unit, the main path convolutional branch undertakes the basic feature extraction task.

[0186] The feature mapping of the main path first achieves local spatial modeling through a 3×3 convolution. After batch normalization and ReLU nonlinear transformation, it further extracts features through a second 3×3 convolution, and finally outputs the result after batch normalization. .

[0187] Meanwhile, feature mapping is developed through three enhanced branches to address different modeling needs.

[0188] The first enhancement branch uses a 1×1 convolution to compress the input features through channels, while simultaneously suppressing the propagation of background noise energy. This is followed by a 3×3 convolution for feature extraction and weighting with Learnable Local Texture Attention (LTA). The output is denoted as... .

[0189] The first enhancement branch captures fine-scale textures and edge amplitudes at a relatively low parameter cost, which can significantly improve the response to short-duration high-frequency pulses and local edges.

[0190] The second enhancement branch first performs 1×1 compression, then combines 1×3 and 3×1 convolutions to capture horizontal texture information, and finally applies a 3×3 dilated convolution with a dilation rate of 5 and LTA weighting. The output is denoted as... .

[0191] The second enhancement branch is used to model lateral texture variations and capture structural correlations over longer distances.

[0192] The third enhancement branch uses 1×1 convolution for dimensionality reduction, followed by a combination of 3×1 and 1×3 convolutions to capture texture information in the vertical direction. After further processing with a 3×3 dilated convolution with a dilation rate of 5 and LTA weighting, the output is denoted as... .

[0193] The third enhancement branch is used to compensate for the shortcomings of the second enhancement branch in modeling the vertical structure.

[0194] in, , , Indicates the height and width of the feature map. This indicates the number of channels. The features extracted by the three enhancement branches are consistent in both spatial and channel dimensions.

[0195] First of all , , By concatenating along the channel dimension, a tensor is obtained. .

[0196] The concatenated features are then remapped back to their original state using a 1×1 convolution. The channels ultimately yield the fused features, expressed by the following formula: .

[0197] Based on this, Lightweight selection kernel attention (LSK) is used to perform multi-scale selection kernel weighting to obtain... This enables adaptive recalibration of different receptive field features.

[0198] In addition to the main path and enhancement branches, residual branches are also introduced. Their core function is to ensure that input features can be directly transmitted to the output without going through complex convolutions through a shortcut path, thereby achieving faithful transmission of cross-layer features.

[0199] The residual branch adjusts the number of channels in the input features through a 1×1 convolution, while batch normalization stabilizes the feature distribution, and the output is... This ensures consistency with the output dimension of the main path.

[0200] In the stage of merging enhanced branches and residual branches, a learnable scaling factor is introduced. This is used to control the contribution of enhanced features to the final output; the fusion formula is: .

[0201] in These are the features after fusion.

[0202] Finally, the main path output With fusion results Element-wise addition, followed by ReLU activation, yields the final output of the residual enhancement unit: ;

[0203] in This represents the final feature representation that integrates main path convolution, three-branch enhancement, and residual branches.

[0204] To further enhance the network's ability to learn fine-grained texture features and local energy distribution in partial discharge images, this invention introduces a learnable local texture attention module (LTA) into each enhancement branch.

[0205] like Figure 7 As shown, the processing flow of Learnable Local Texture Attention (LTA) is as follows:

[0206] in, Figure 7 In the diagram, DW Conv win×win represents channel-wise window convolution, avgpool represents global average pooling, MLP represents channel mapping structure, and reconstruction represents the inverse splicing reconstruction of local weighted features.

[0207] The LTA module first dynamically determines the local window size based on the layer depth of the current feature extraction stage. The window size is defined as: ,in, , Index the stages (i.e., the four stages mentioned above).

[0208] This dynamic windowing mechanism allows shallow features to use smaller windows to highlight high-frequency pulse textures, while deep features use larger windows to take into account a wider range of energy distribution information.

[0209] After completing the local window division, channel-wise window convolution is used to independently extract the region-level local response of each local window, thus obtaining the local texture response features. Its expression is: .

[0210] in Indicates input features, This indicates a channel-wise convolution operation.

[0211] Used to characterize the fine-grained texture response and local structural changes of input features within each local window.

[0212] Then on Perform global average pooling to obtain the channel description vector. :

[0213] Then The input consists of an MLP-style channel mapping structure composed of two 1×1 convolutional layers, and is activated by the ReLU function. With Sigmoid activation function Generate local energy weights Its expression is:

[0214] ;in, This indicates a global average pooling operation. and These represent the 1×1 convolution mappings of the first and second layers, respectively.

[0215] Through the above operations, the energy differences in different local regions along the channel dimension can be adaptively modeled.

[0216] Local energy weight After being broadcast to the spatial dimension, it is compared with local texture response features. Element-wise multiplication yields weighted local features. Then, the corresponding local windows Perform inverse stitching to reconstruct a global-local weighted feature map. .

[0217] When the reconstructed global-local weighted feature map Size and Input Features When the dimensions are inconsistent, bilinear interpolation is used to weight the global and local feature maps. Adjust to Same space dimensions.

[0218] Finally, the LTA module uses element-wise multiplication to achieve dynamic fusion, and its output is denoted as... The expression is: .

[0219] in, This indicates element-wise multiplication.

[0220] Through the above processing, the LTA module can highlight the key texture areas of the discharge pulse against a complex noise background, enhance the response of local edges and high-frequency details, and suppress irrelevant background components.

[0221] In the first, second, and third enhancement branches, the features refined by convolution are all input into the LTA module for local texture weighting. The enhanced features output by the LTA module are used as the final outputs of the corresponding enhancement branches to achieve adaptive enhancement of texture details and local energy distribution in partial discharge images.

[0222] To achieve adaptive selection of different receptive field features and multi-scale global modeling, this invention introduces a lightweight selection kernel module (LSK) after the fused features obtained by channel splicing and convolution mapping of the three enhancement branches.

[0223] The network structure of the LSK module is as follows: Figure 8 As shown, Figure 8 In this context, Conv represents standard convolution, DConv represents dilated convolution, GAP represents global average pooling, and Softmax represents branch dimension normalization. The processing flow is as follows:

[0224] Let the fusion feature input to the LSK module be: and order ,in, .

[0225] The LSK module first processes the input features. Two parallel convolutional branches are constructed to extract response features under different receptive fields.

[0226] The first convolutional branch uses a standard 3×3 convolution to model the input features at a local scale, resulting in the first branch features. ;

[0227] The second convolutional branch uses a dilated 3×3 convolution to model the input features with a larger receptive field, simulating the feature extraction effect of a 5×5 receptive field, thus obtaining the second branch features. The formula is expressed as follows: , ;

[0228] in This represents a standard 3×3 convolution operation. This indicates a 3×3 dilation convolution operation.

[0229] The outputs of the two parallel convolutional branches are then summed element-wise to obtain the global descriptive features. and to Perform global average pooling to obtain channel-level description vectors. Its expression is: , .

[0230] in, This indicates a global average pooling operation. It is used to characterize the overall response strength of the current input feature in each channel and to provide a global statistical basis for the adaptive allocation of subsequent branch weights.

[0231] Obtaining the channel-level description vector Then, it is input into a bottleneck mapping structure consisting of two 1×1 convolutional layers to generate response weights for two parallel convolutional branches on each channel. The weights are then normalized along the branch dimension using the Softmax function to obtain the weights of the first branch. With the weight of the second branch .

[0232] For any channel position All satisfy the condition that the sum of the weights of the corresponding branches is 1, that is: .

[0233] Through the above normalization mechanism, the LSK module can adaptively allocate the response intensity of the two parallel convolution branches on each channel according to the input feature content, thereby achieving selective enhancement and fusion of different receptive field features.

[0234] The normalized weights of the two branches are applied to the corresponding branch features, and then element-wise weighted summation is performed to obtain the multi-scale selection output. Its expression is: .

[0235] in This represents element-wise multiplication; This represents the multi-scale fusion feature after adaptive kernel selection and weighting.

[0236] At this point, the LSK module has completed the adaptive selection and weighted fusion of features from different receptive field branches, and its output... As the final output of the module, it is used for subsequent feature representation and classification recognition.

[0237] The LSK module achieves adaptive selection, weighted fusion, and stable output of different receptive field features without significantly increasing the number of parameters, thereby improving the network's ability to model global scale information and multi-scale structural differences in partial discharge images. The outputs of the three enhancement branches are concatenated and convolutionally mapped, and the fused features are input into the LSK module for multi-scale selection kernel weighting. The enhanced features output by the LSK module serve as the final output after the enhancement branches are fused, achieving adaptive selection of different receptive field features in partial discharge images, multi-scale information fusion, and global semantic representation enhancement.

[0238] After four stages of residual enhancement feature extraction, a size of [size missing] is obtained. The high-dimensional feature map is generated. To reduce redundant parameters while preserving the global structural semantics, a Global Average Pooling (GAP) strategy is adopted. This strategy compresses the two-dimensional feature map of each channel into a single scalar through averaging, thereby generating a high-dimensional feature map of length [length missing]. The global semantic vector is then input into the classification output layer, and the partial discharge category prediction result is obtained through fully connected mapping and Softmax probability normalization.

[0239] Partial discharge is classified into five types: tip discharge, particle discharge, air gap discharge, suspension discharge, and surface discharge.

[0240] Step 4. Based on the image data after denoising signal reconstruction and image conversion processing in Step 2, train the partial discharge recognition model built in Step 3, and use the trained model to identify partial discharge.

[0241] During training, follow these steps:

[0242] Step I. Data loading is completed using image folders with automatic labeling by category. Then, the dataset is randomly divided into training and validation sets in a ratio of 8:2 to ensure the balance of the model during training and generalization.

[0243] All images are subjected to tensor quantization and size normalization before being input into the network, and the batch size is set to 16 to ensure training stability and computational efficiency.

[0244] Step II. Record the processed image as... First, the input layer operation in step 3.1 is performed, and the final output is the feature map obtained from the initial convolution and pooling, denoted as... .

[0245] in Describe the specific operations of the input layer, that is .

[0246] Preliminary feature mapping The input is the backbone feature extraction network, which completes the extraction and fusion of multi-scale and multi-directional features through four consecutive residual feature extraction stages:

[0247] The first stage contains three multi-path feature fusion residual units, with a total of 64 output channels. Feature maps are obtained after feature extraction and fusion. ,Right now ,in This indicates the feature extraction operation in the first stage.

[0248] The second stage contains four multi-path feature fusion residual units, with 128 output channels. Feature maps are obtained after feature extraction and fusion. ,Right now ,in This indicates the feature extraction operation in the second stage.

[0249] The third stage contains 6 multi-path feature fusion residual units, with a total of 256 output channels. Feature maps are obtained after feature extraction and fusion. ,Right now ,in This indicates the feature extraction operation in the third stage.

[0250] The fourth stage contains three multi-path feature fusion residual units, with a total of 512 output channels. Feature maps are obtained after feature extraction and fusion. ,Right now ,in This indicates the feature extraction operation in the fourth stage.

[0251] Step III. Obtain the high-dimensional feature map output by the backbone feature extraction network in Step II. Next, the classification results are generated and decisions are made, including steps such as fully connected mapping, probability normalization, and loss function construction, as detailed below:

[0252] First, the high-dimensional feature map is... The input is fed into a fully connected layer for feature mapping and dimensionality compression to obtain a classification score vector. Its mathematical definition is: .

[0253] in This represents the fully connected layer mapping operator. , The total number of categories involved in the classification task.

[0254] Secondly, regarding the classification score vector Perform softmax normalization to transform it into a probability distribution vector. :

[0255] .

[0256] in, This indicates that the input sample belongs to the first... The predicted probability of a class. Furthermore, based on the probability distribution... Compared with the true category label of the sample Construct the cross-entropy loss function Its definition is as follows:

[0257] ;in Indicates the actual label category The corresponding predicted probability, This serves as the optimization objective to guide the iterative updates of network parameters. Based on the loss function, the gradient information of each network parameter is calculated through backpropagation, thereby quantifying the contribution of each parameter to the loss.

[0258] By utilizing gradient information and combining iterative updates of all trainable parameters in the network with an optimizer (Adam in this example), the parameter values ​​are gradually adjusted, so that the network continuously reduces the loss function and improves the classification accuracy during the training iteration process.

[0259] Repeat the above process until the set number of training rounds is completed (50 rounds in this example).

[0260] Step 3.3. After completing the training in Step 3.2, the partial discharge image to be tested is input into the trained residual neural network that integrates multi-path feature enhancement, learnable local texture attention, and lightweight selection kernel attention for classification and recognition.

[0261] During testing, each input image is first subjected to tensor quantization and size normalization operations according to the preprocessing method used in the training phase to ensure the consistency of the network input features, denoted as . Record the processed image as... First, the input layer operation in step 3.1 is performed, and the final output is the feature map obtained from the initial convolution and pooling, denoted as... .

[0262] Then, The trained network model is input and sequentially processed through a four-stage multi-path feature enhancement residual network for feature extraction and fusion to obtain a high-dimensional feature map. This leads to the classification and decision-making stage.

[0263] In the classification decision stage, the high-dimensional feature map is... Input fully connected mapping operator Generate classification score vectors : .in This indicates the total number of categories involved in the classification task.

[0264] Then on Perform softmax normalization to obtain the probability distribution vector. :

[0265] .

[0266] in, This indicates that the input test image belongs to the first... The predicted probability of a class. Based on the probability distribution vector, the model selects the predicted class using the maximum index: .

[0267] in This represents the index of the network's predicted class for the test image. The predicted class will then be... With test data true labels Comparison, via indicator function The number of correctly predicted samples is counted, and then the overall classification accuracy is calculated. :

[0268] .

[0269] in, The total number of test samples, and The value is 1 if the prediction is correct, and 0 otherwise. Through the above testing process, the network's output on an independent test set can be obtained, and various metrics can be calculated.

[0270] Table 1 Objective Evaluation

[0271]

[0272] Experimental results show that the self-designed residual neural network with multi-path feature enhancement proposed in this invention can achieve high accuracy and robustness in partial discharge image classification tasks, providing reliable support for partial discharge status diagnosis.

[0273] Compared with existing technical solutions that rely solely on fixed parameters or static network structures, the neural network partial discharge identification method proposed in this invention, which integrates polarization particle swarm optimization, has higher signal reconstruction, feature extraction capabilities, and classification accuracy.

[0274] Example 2

[0275] This embodiment 2 describes a neural network partial discharge identification system with fusion polarization particle swarm optimization, which is based on the same concept as the neural network partial discharge identification method with fusion polarization particle swarm optimization in embodiment 1 above.

[0276] The neural network partial discharge identification system in this embodiment, which integrates polarization particle swarm optimization, includes the following modules:

[0277] The data acquisition module is used to automatically parse the hexadecimal files output by the original partial discharge monitoring equipment; each original signal file contains a complete instantaneous waveform signal of partial discharge, and data cleaning is performed.

[0278] The denoising signal reconstruction and image conversion module is used to propose a method that combines wavelet filtering and grouped cooperative polarization particle swarm optimization to perform denoising signal reconstruction and image conversion on the obtained instantaneous waveform signal of partial discharge.

[0279] And a partial discharge identification module, used to build a partial discharge identification model, which includes an input layer, a backbone feature extraction network, a global average pooling and an output layer; wherein the backbone feature extraction network includes four consecutive residual feature extraction stages;

[0280] Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; wherein, each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism;

[0281] Building upon this foundation, a learnable local texture attention module and a lightweight selection kernel attention module are further introduced to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features.

[0282] Specifically, based on the image data after denoising signal reconstruction and image conversion processing, the constructed partial discharge recognition model is trained, and the trained model is used to identify partial discharge.

[0283] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0284] Example 3

[0285] This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the neural network partial discharge identification method with fused polarization particle swarm optimization described in embodiment 1 above.

[0286] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0287] Example 4

[0288] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the neural network partial discharge identification method with fused polarization particle swarm optimization in embodiment 1 above.

[0289] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0290] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A neural network partial discharge identification method integrating polarization particle swarm optimization, characterized in that, Includes the following steps: Step 1. Automatically parse the hexadecimal files output by the original partial discharge monitoring equipment; each original signal file contains a complete instantaneous partial discharge waveform signal, and data cleaning is performed; Step 2. A method combining wavelet filtering and grouped cooperative polarization particle swarm optimization is proposed to denoise and reconstruct the instantaneous waveform signal of partial discharge after data cleaning, and convert it into a two-dimensional image; Step 3. Construct a partial discharge identification model, which includes an input layer, a backbone feature extraction network, a global average pooling layer, and an output layer; the backbone feature extraction network includes four consecutive residual feature extraction stages; Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism; Building upon this foundation, a learnable local texture attention module and a lightweight selection kernel attention module are further introduced to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features. Step 4. Based on the two-dimensional image data after denoising signal reconstruction and image conversion processing in Step 2, train the partial discharge recognition model built in Step 3, and use the trained model to identify partial discharge.

2. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 1, characterized in that, In step 1, the data cleaning process is as follows: Step 1.

1. Read the file content as the string .hex_data, and keep only the hexadecimal characters in the string; Step 1.

2. Group each pair of characters and convert them into decimal integers between 0 and 255 to form a one-dimensional signal sequence; Step 1.

3. Convert the sequence to a PyTorch tensor And adjust it to the three-dimensional structure required for the input convolution operation: ； The first dimension is the batch size; the second dimension is the number of channels, i.e., a single channel; the third dimension... This is the signal length.

3. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 1, characterized in that, Step 2 specifically involves: Define a 17-dimensional optimization parameter vector, and divide it into a structure parameter group and a threshold parameter group, which are searched collaboratively by the structure parameter particle swarm and the threshold parameter particle swarm, respectively; the iteration process is as follows: First, the angle variables of the two sets of particles are mapped to actual parameters, and a complete candidate parameter combination is constructed by pairing them. Then, the candidate parameter combination is input into the adaptive wavelet denoising module to complete wavelet kernel convolution, soft thresholding denoising and signal reconstruction. Subsequently, based on the pulse energy concentration evaluation term and the pulse envelope shape preservation evaluation term, the adaptive no-reference fitness value corresponding to the candidate parameter combination is calculated, and the individual optimal state of each particle in the structure parameter particle swarm and the threshold parameter particle swarm, as well as the population optimal state of each of the two particle swarms, are updated respectively. Finally, based on the dynamic inertia weight, random disturbance, individual optimal state, and group optimal state, the angle variables of each particle in the structural parameter particle swarm and the threshold parameter particle swarm are updated. Repeat the above iterative process until the preset maximum number of iterations is reached, output the final optimal combination of parameters, and complete the denoised signal reconstruction and image conversion based on the final optimal combination of parameters.

4. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 3, characterized in that, In step 2, the process of optimizing parameter definition, grouping, and initial particle swarm configuration is as follows: Use a parameterized wavelet kernel with a quantity of n=8. ; Define the scale parameter vector ,scope Translation parameter vector ,scope Unified soft threshold ; The overall optimization parameter vector is constructed as follows: To improve parameter search efficiency, the 17-dimensional parameters are divided into two parameter groups: ; in For structural parameter groups, For the threshold parameter set; during the initialization phase, the total size of the particle swarm is set to... And set the structure parameter particle swarm size to respectively. Threshold parameter: particle swarm size ,in ; In this context, each particle in the structure parameter particle swarm corresponds to a set of candidate solutions for the structure parameters, and each particle in the threshold parameter particle swarm corresponds to a set of candidate solutions for the threshold parameters.

5. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 3, characterized in that, In step 2, the process of angle encoding, parameter mapping, and particle swarm state symbol definition is as follows: Angle encoding is introduced, and a monotonic mapping method of "angle → actual parameter" is used to convert angle variables into actual parameter values. The processing procedure is as follows: Let the upper and lower bounds of any parameter be respectively... and Its corresponding angle variable is Then we have: ； The mapping relationship between parameters and angle variables is as follows: ； in, Indicates the first Actual parameter values; For the structure parameter particle swarm, the first The particle, its first The angle vector of the wheel is denoted as For the threshold parameter particle swarm, the th The particle, its first The angle vector of the wheel is denoted as The optimal angle vectors for each individual particle in the structure parameter particle swarm and the threshold parameter particle swarm are respectively denoted as... , ; The two groups of particles in the first The optimal angle vectors of the group in each iteration are denoted as follows: , .

6. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 3, characterized in that, In step 2, the process of constructing the no-reference fitness function is as follows: First, define the original signal. With reconstructed signal The envelopes are as follows: , ;in Represents the Hilbert transform. Represents the imaginary unit. Represents the magnitude of the analytic signal; Represents the original envelope. Indicates the reconstruction of the envelope; The mean and standard deviation of the envelope are further calculated using the following formulas: ； ； ； ； in , These are the mean values ​​of the original envelope and the reconstructed envelope, respectively. , These are the standard deviations of the original envelope and the reconstructed envelope, respectively. The signal length; Define the original pulse candidate support set and the reconstructed pulse candidate support set as follows: ； ； in The candidate pulse extraction coefficients; To avoid the support set being empty in extreme cases, if Then let ; like Then let ; Define the pulse energy percentage of the reconstructed signal. for: ； in, To prevent extremely small positive numbers with a denominator of zero; Define the first evaluation item for: ； In the original pulse candidate support set The correlation coefficient between the original envelope and the reconstructed envelope is defined above. for: ； in: ; and Support sets The mean of the original envelope and the reconstructed envelope; Represents a set The number of elements in the middle; Define the second evaluation item for: ； To adaptively adjust the weights based on the two evaluation results, an adaptive weight is defined. and for: ， ； in, To prevent extremely small positive numbers with a denominator of zero; and These represent the evaluation items. and The corresponding adaptive weights satisfy: ; Thus, the candidate parameter combinations are obtained. The corresponding adaptive no-reference fitness function: ； in Indicates candidate parameter combinations The overall evaluation value; For the structure parameter particle swarm, the first When evaluating the cooperative fitness of a particle, it is paired with the swarm's optimal actual threshold parameter to form a candidate parameter combination: Its corresponding cooperative fitness Defined as: ;in Indicates the first In the particle swarm of wheel structure parameters, the first Candidate solutions for the actual structure parameters of each particle. Indicates the first The optimal actual threshold parameter for the particle swarm. For the threshold parameter particle swarm, the th When evaluating the cooperative fitness of a particle, it is paired with the swarm's optimal actual structural parameters to form candidate parameter combinations: Its corresponding cooperative fitness Defined as: ; Indicates the first In the first round of threshold parameter particle swarm optimization Candidate solutions for the actual threshold parameters of each particle. Indicates the first The swarm optimal actual structural parameters of the particle swarm for the wheel structure parameters.

7. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 1, characterized in that, In step 3, the multi-path feature fusion residual unit includes one main path and three enhancement branches; The processing flow of the multi-path feature fusion residual unit is as follows: The feature mapping of the main path first achieves local spatial modeling through a 3×3 convolution. After batch normalization and ReLU nonlinear transformation, it further extracts features through a second 3×3 convolution, and finally outputs the result after batch normalization. ; at the same time The feature mapping is carried out through three enhancement branches, each addressing different modeling needs. The first enhancement branch uses a 1×1 convolution to compress the input features through channels, while simultaneously suppressing the propagation of background noise energy. This is followed by a 3×3 convolution for feature extraction and weighting with Learnable Local Texture Attention (LTA). The output is denoted as... ; The second enhancement branch first performs 1×1 compression, then combines 1×3 and 3×1 convolutions to capture horizontal texture information, and finally applies a 3×3 dilated convolution with a dilation rate of 5 and LTA weighting. The output is denoted as... ; The third enhancement branch uses 1×1 convolution for dimensionality reduction, followed by a combination of 3×1 and 1×3 convolutions to capture texture information in the vertical direction. After further processing with a 3×3 dilated convolution with a dilation rate of 5 and LTA weighting, the output is denoted as... ; in, , , Indicates the height and width of the feature map. Indicates the number of channels; First, , , By concatenating along the channel dimension, a tensor is obtained. ; The concatenated features are then remapped back to their original state using a 1×1 convolution. The channels ultimately yield fused features, expressed as follows: ； Based on this, Lightweight selection kernel attention (LSK) is used to perform multi-scale selection kernel weighting to obtain... ; In addition to the main path and enhancement branches, a residual branch is introduced. The residual branch adjusts the number of channels in the input features through a 1×1 convolution, while batch normalization stabilizes the feature distribution. The output is... ; In the stage of merging enhanced branches and residual branches, a learnable scaling factor is introduced. This is used to control the contribution of enhanced features to the final output; the fusion formula is: ; Features after fusion; Finally, the main path output... With fusion results Element-wise addition, followed by ReLU activation, yields the final output of the residual enhancement unit: ; in This represents the final feature representation that integrates main path convolution, three-branch enhancement, and residual branches.

8. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 7, characterized in that, The processing flow of the learnable local texture attention (LTA) is as follows: The LTA module first dynamically determines the local window size based on the layer depth of the current feature extraction stage. The window size is defined as: ,in, , For stage index; After completing the local window division, channel-wise window convolution is used to independently extract the region-level local response of each local window, thus obtaining the local texture response features. Its expression is: ; in Indicates input features, This represents a channel-wise convolution operation; Then on Perform global average pooling to obtain the channel description vector. ; Then The input consists of an MLP-style channel mapping structure composed of two 1×1 convolutional layers, and is activated by the ReLU function. With Sigmoid activation function Generate local energy weights Its expression is: ;in, This indicates a global average pooling operation. and These represent the 1×1 convolution mappings of the first and second layers, respectively. Local energy weight After being broadcast to the spatial dimension, it is compared with local texture response features. Element-wise multiplication yields weighted local features. Then, the corresponding local windows Perform inverse stitching to reconstruct a global-local weighted feature map. ; When the reconstructed global-local weighted feature map Size and Input Features When the dimensions are inconsistent, bilinear interpolation is used to weight the global and local feature maps. Adjust to Same space dimensions; Finally, the LTA module uses element-wise multiplication to achieve dynamic fusion, and its output is denoted as... The expression is: in, This indicates element-wise multiplication.

9. The neural network partial discharge identification method based on fusion polarization particle swarm optimization according to claim 7, characterized in that, The processing flow of the Lightweight Selective Kernel Attention (LSK) module is as follows: Let the fusion feature input to the LSK module be: and order ,in The LSK module first processes the input features. Two parallel convolutional branches are constructed to extract response features under different receptive fields; The first convolutional branch uses a standard 3×3 convolution to model the input features at a local scale, resulting in the first branch features. ; The second convolutional branch uses a dilated 3×3 convolution to model the input features with a larger receptive field, simulating the feature extraction effect of a 5×5 receptive field, thus obtaining the second branch features. The formula is expressed as follows: ； in This represents a standard 3×3 convolution operation. This represents a 3×3 dilation convolution operation; The outputs of the two parallel convolutional branches are then summed element-wise to obtain the global descriptive features. and to Perform global average pooling to obtain channel-level description vectors. Its expression is: , ; in, This represents the global average pooling operation; it obtains the channel-level description vector. Then, it is input into a bottleneck mapping structure consisting of two 1×1 convolutions to generate response weights for two parallel convolution branches on each channel. Then, the weights of the first branch are obtained by normalizing the weights along the branch dimension using the Softmax function. Weights of the second branch For any channel position All satisfy the condition that the sum of the weights of the corresponding branches is 1, i.e. ; The normalized weights of the two branches are applied to the corresponding branch features, and then element-wise weighted summation is performed to obtain the multi-scale selection output. Its expression is: ; in This represents element-wise multiplication; This represents the multi-scale fusion features after adaptive kernel selection and weighting; At this point, the LSK module has completed the adaptive selection and weighted fusion of features from different receptive field branches, and its output... As the final output of the module, it is used for subsequent feature representation and classification recognition.

10. A neural network partial discharge identification system integrating polarization particle swarm optimization, characterized in that, Includes the following steps: The data acquisition module is used to automatically parse the hexadecimal files output by the original partial discharge monitoring equipment; each original signal file contains a complete instantaneous waveform signal of partial discharge, and data cleaning is performed. The denoising signal reconstruction and image conversion module is used to propose a method that combines wavelet filtering and grouped cooperative polarization particle swarm optimization to perform denoising signal reconstruction and image conversion on the obtained instantaneous waveform signal of partial discharge. And a partial discharge identification module, used to build a partial discharge identification model, which includes an input layer, a backbone feature extraction network, a global average pooling and an output layer; wherein the backbone feature extraction network includes four consecutive residual feature extraction stages; Each residual feature extraction stage is designed with a multi-path feature fusion residual unit; wherein, each of the multi-path feature fusion residual units constructs a multi-scale, multi-directional parallel enhancement path and introduces a feature fusion mechanism; Building upon this foundation, a learnable local texture attention module and a lightweight selection kernel attention module are further introduced to achieve local texture detail enhancement, global scale information modeling, and aggregation and enhancement of multi-level semantic features. Specifically, based on the image data after denoising signal reconstruction and image conversion processing, the constructed partial discharge recognition model is trained, and the trained model is used to identify partial discharge.

Images