Sea wind power GIS equipment partial discharge pattern recognition method based on improved SAE network

Abstract

The application provides an offshore wind power GIS equipment partial discharge mode recognition method based on an improved SAE network, and belongs to the technical field of electric digital data processing.The application constructs a partial discharge experiment platform conforming to the IEC60270 standard, designs four typical defect models to collect PRPD spectrum diagrams, removes interference elements by using HSV color space threshold segmentation, and combines a partial discharge physical mechanism equation to enhance image features.The improved SAE network fuses an encoder with three layers of convolution and an up-sampling decoder, introduces an ECA attention mechanism to optimize feature channel weights, combines a pre-trained PDCN model to enhance feature extraction capability, introduces a physical parameter constraint in the training process, and realizes high-precision recognition of four defect types, namely, sharp discharge, surface discharge, air gap discharge and suspended metal discharge, after global pooling dimension reduction and full connection layer processing, so that the technical problem of low recognition accuracy and poor generalization capability of offshore wind power GIS equipment partial discharge mode recognition is solved.

Description

Technical Field

[0001] This invention belongs to the field of electrical digital data processing technology, and more specifically, relates to a method for partial discharge pattern recognition of offshore wind power GIS equipment based on an improved SAE network. Background Technology

[0002] Gas-insulated switchgear (GIS) for offshore wind power is widely used in power systems due to its high reliability and miniaturization. Partial discharge within GIS equipment is a crucial indicator for assessing insulation status, and its real-time monitoring and accurate identification are essential for ensuring the safe operation of offshore wind power systems. Traditional partial discharge detection methods primarily rely on manual analysis of phase-resolved partial discharge (PRPD) spectra, judging different defect types based on expert experience. This method is highly subjective and inefficient. In recent years, machine learning methods have been introduced into the field of GIS partial discharge pattern recognition, with commonly used methods including support vector machines, random forests, and convolutional neural networks. While these methods have achieved some success on specific datasets, they suffer from insufficient feature extraction and limited model generalization ability in practical applications. Especially in the complex environment of offshore wind power, equipment is affected by factors such as temperature, humidity, and salt spray, leading to significant changes in discharge characteristics and further increasing the difficulty of identification. Most existing deep learning models ignore the physical mechanisms of partial discharge, relying solely on data-driven feature learning. This lack of understanding and constraints on the physical processes of discharge results in a significant decrease in recognition accuracy and insufficient generalization ability when faced with unseen discharge patterns or discharge signals in complex environments. This technical problem severely restricts the practicality and reliability of the intelligent monitoring system for offshore wind power GIS equipment. Summary of the Invention

[0003] In view of this, the present invention provides a method for partial discharge pattern recognition of offshore wind power GIS equipment based on an improved SAE network, which can solve the technical problems of low accuracy and poor generalization ability of partial discharge pattern recognition of offshore wind power GIS equipment in the prior art.

[0004] This invention is implemented as follows: This invention provides a method for partial discharge pattern recognition of offshore wind power GIS equipment based on an improved SAE network, comprising: constructing a GIS partial discharge experimental platform; designing four defect models based on the GIS partial discharge experimental platform and acquiring PRPD spectra; preprocessing the acquired PRPD spectra data; constructing an improved SAE network model, fusing an ECA attention mechanism between the encoder output and decoder input, and using a pre-trained... The PDCN model assists in network feature extraction; a training dataset is established based on the preprocessed PRPD spectral data to train the improved SAE network model; the GIS device under test and the processed PRPD spectral data are input into the trained improved SAE network model to identify defect types and obtain identification results.

[0005] Specifically, the steps for constructing the GIS partial discharge experimental platform are carried out according to the IEC 60270 standard. The experimental circuit consists of an AC power supply, a transformer, a protective resistor, a coupling capacitor, a detection impedance, a defect model, and a general partial discharge measurement and analysis system. The protective resistor is set to 5kΩ, the coupling capacitor to 800pF, and the GIS cavity is filled with air at a pressure of 0.1MPa. The gas used is the MPD 800 general partial discharge measurement and analysis system, with the MPD 800 center frequency set to 250kHz.

[0006] The four defect models are specifically pointed discharge, surface discharge, air gap discharge, and suspended metal discharge. The test voltage for pointed discharge is 4.9 kV, the test voltage for surface discharge is 3.5 kV, the test voltage for air gap discharge is 6.9 kV, and the test voltage for suspended metal discharge is 7.5 kV. 500 PRPD spectra are collected for each defect type at 2-minute intervals.

[0007] Specifically, the step of preprocessing the acquired PRPD spectral data involves using the color thresholding method of the HSV color space to remove the coordinate axes and grid lines from the original image, and then uniformly adjusting all images to a size of 256×256 and performing normalization processing.

[0008] The HSV color thresholding method converts the image from the BGR color space to the HSV color space, defines an HSV range suitable for gray and black grids, extracts pixels in the image that fit the range through color thresholding, and replaces the grid lines and coordinate axes with the background color.

[0009] In the preprocessing step, the PRPD spectral features are further enhanced based on the partial discharge physical mechanism equation, which describes... The physical process of partial discharge in a gaseous medium and its relationship with the surrounding electric field distribution.

[0010] The inputs to the partial discharge physical mechanism equation include electric field strength, gas pressure, temperature, air gap distance, and relative permittivity of the medium, and the outputs are discharge initiation voltage and discharge amplitude. The discharge amplitude is used for discharge amplitude calibration in the PRPD spectrum feature enhancement process.

[0011] The encoder part of the improved SAE network model includes three convolutional layers that use 64, 128, and 256 filters respectively for feature extraction, while the decoder part restores the spatial dimension of the image through an upsampling layer.

[0012] The ECA attention mechanism is an efficient channel attention mechanism that optimizes the weight distribution of feature channels by adaptively selecting the convolution kernel size. First, it obtains the description of each channel through global average pooling, and then uses one-dimensional convolution to perform information interaction between channels.

[0013] Among them, the The PDCN model is a type of model used for A pre-trained deep convolutional neural network model for partial discharge identification in a gaseous environment is constructed using a deep convolutional neural network based on multi-scale residual blocks.

[0014] Among them, the description The PDCN model comprises an initial convolutional layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. Each multi-scale residual block contains three parallel convolutional branches with kernel sizes of 3×3, 5×5, and 7×7. The steps for establishing the training dataset for the PDCN model specifically include collecting 10,000 discharge spectra of different types from a publicly available GIS partial discharge database, extracting physical feature vectors from each spectra, and simultaneously collecting data under different operating conditions. The gas insulation characteristic parameters were used to obtain the training dataset.

[0015] The step of dividing the data into training and testing sets involves dividing 2000 PRPD spectral images into training and testing sets. The training set comprises 70% (1400 images), and the testing set comprises 30% (600 images). The improved SAE network model is trained using the training set. The step of inputting the testing set data into the trained improved SAE network model for defect type identification involves global pooling dimensionality reduction and fully connected layers to process features. Finally, a Softmax classifier is used to output the identification results for four defect types. During the training process, partial discharge physical parameters are used as constraints, including discharge initiation voltage constraints, discharge phase distribution constraints, and discharge amplitude distribution constraints.

[0016] This invention proposes to acquire high-quality data by constructing an experimental platform, enhance PRPD spectral features by combining the physical mechanism equation of partial discharge, and integrate ECA attention mechanism and pre-trained data. The PDCN model achieves highly accurate identification of four common defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. This method combines the feature extraction capabilities of an autoencoder with adaptive feature weight allocation through an attention mechanism in its network structure design. Furthermore, by introducing constraints on partial discharge physical parameters, it integrates physical laws into the deep learning process, effectively addressing the problems of insufficient feature extraction and limited model generalization ability in traditional identification methods. Compared to traditional methods, this invention not only improves identification accuracy but also ensures the stability and reliability of the model under different environmental conditions. Through the organic combination of physical mechanisms and deep learning, it solves the technical problems of low accuracy and poor generalization ability in partial discharge pattern recognition for offshore wind power GIS equipment. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 This is a schematic diagram of the overall structure of the GIS partial discharge experimental platform in Example 2.

[0019] Figure 3 This is a schematic diagram of the partial discharge defect model in Example 2.

[0020] Figure 4 The partial discharge PRPD spectra (after preprocessing) for different defect types in Example 2 are shown.

[0021] Figure 5 This is an actual diagram of the GIS partial discharge experimental platform in Example 2. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0023] like Figure 1 The diagram shows a flowchart of a partial discharge pattern recognition method for offshore wind power GIS equipment based on an improved SAE network, provided by this invention. The method includes the following steps: S01. Constructing a GIS partial discharge experimental platform: The experimental circuit consists of an AC power supply, transformer, protective resistor, coupling capacitor, detection impedance, defect model, and MPD 800. The protective resistor is set to 5kΩ, the coupling capacitor to 800pF, and the GIS cavity is filled with gas at a pressure of 0.1MPa. For the gas, the center frequency of the MPD 800 is set to 250kHz, and the ambient noise is controlled at around 200fC. S02. Design defects: Based on the GIS partial discharge experimental platform, four defect models were designed: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. The test voltage for tip discharge was 4.9kV, the test voltage for surface discharge was 3.5kV, the test voltage for air gap discharge was 6.9kV, and the test voltage for suspended metal discharge was 7.5kV. 500 PRPD spectra were collected for each defect type at 2-minute intervals. S03. Preprocess the collected PRPD spectral data: Use the color threshold segmentation method of HSV color space to remove the coordinate axes and grid lines in the original image. Then, adjust all images to a size of 256×256 and normalize them. At the same time, enhance the PRPD spectral features based on the partial discharge physical mechanism equation. S04. Constructing an improved SAE network model: The encoder part consists of three convolutional layers using 64, 128, and 256 filters for feature extraction, respectively. The decoder part restores the image spatial dimension through an upsampling layer. An ECA attention mechanism is fused between the encoder output and the decoder input, and a pre-trained... PDCN model assists in network feature extraction; S05. Model Training: The 2000 PRPD spectral images are divided into a training set (70%, 1400 images) and a test set (30%, 600 images). The improved SAE network model is trained using the training set, with partial discharge physical parameters used as constraints. The test set data is then input into the trained improved SAE network model. Features are processed through global pooling dimensionality reduction and fully connected layers. Finally, a Softmax classifier is used to output the identification results for four defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. The identification results are verified by constraining the partial discharge physical parameters. S06. Model Application: Input the GIS equipment under test and the processed PRPD spectral data into the trained improved SAE network model to identify the defect type and obtain the identification results.

[0024] The GIS partial discharge experimental platform was built in accordance with the IEC60270 standard to ensure the quality of data acquisition.

[0025] The color thresholding method in the HSV color space converts the image from the BGR color space to the HSV color space, defines the HSV range suitable for gray and black grids, extracts pixels in the image that meet the range through color thresholding, replaces the grid lines and coordinate axes with the background color, and preserves the original image content of other areas.

[0026] Among them, the PRPD spectrum is a phase-resolved partial discharge spectrum, which shows the distribution characteristics of the partial discharge signal within the AC voltage cycle. The horizontal axis represents the phase angle from 0 to 360 degrees, the vertical axis represents the discharge amplitude, and the density of points represents the discharge frequency. Different defect types exhibit unique discharge patterns on the PRPD spectrum.

[0027] Among them, the physical mechanism equation of partial discharge is used to describe The physical process of partial discharge in a gaseous medium and its relationship with the surrounding electric field distribution are investigated. Inputs include electric field strength E, gas pressure p, temperature T, gas gap distance d, and the relative permittivity ε of the medium. The electric field strength E is calculated from the test voltage and electrode spacing, and the gas pressure p is derived from the gas filling the GIS cavity. The gas pressure value, the temperature T derived from laboratory ambient temperature measurements, the air gap distance d derived from the geometric dimensions of the defect model, and the relative permittivity ε of the medium derived from... The gas physical properties database outputs a discharge initiation voltage Ui and a discharge amplitude q. The discharge initiation voltage Ui is used to determine the discharge characteristic thresholds for different defect types, and the discharge amplitude q is used for discharge amplitude calibration during the PRPD spectrum feature enhancement process.

[0028] Among them, the ECA attention mechanism is an efficient channel attention mechanism that optimizes the weight allocation of feature channels by adaptively selecting the convolution kernel size. First, it obtains the description of each channel through global average pooling, then uses one-dimensional convolution to perform information interaction between channels, and finally generates the attention weight of each channel through the Sigmoid activation function.

[0029] Among them, the SAE network is a deep learning model composed of multiple autoencoder layers stacked together. The autoencoder compresses the input data into a low-dimensional space through the encoder and then reconstructs the input data through the decoder. The encoder consists of an input layer and a hidden layer, which extracts features from the input vector through nonlinear transformation. The decoder consists of a hidden layer and an output layer, which forms an output value that approximates the input vector based on the extracted feature mapping.

[0030] The improved SAE network uses global pooling for feature dimensionality reduction, processes features through a fully connected layer containing 128 neurons, and finally uses a four-neuron Softmax output layer to identify and classify four defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge.

[0031] in, The PDCN model is a type of model used for A pre-trained deep convolutional neural network model for partial discharge identification in a gaseous environment, the The specific structure of the PDCN model is a deep convolutional neural network based on multi-scale residual blocks, including an initial convolutional layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. The initial convolutional layer uses 64 7×7 convolutional kernels to extract features from the input PRPD spectrum. Each multi-scale residual block contains three parallel convolutional branches with kernel sizes of 3×3, 5×5, and 7×7, respectively, capturing discharge features at different scales. The number of output channels for each multi-scale residual block is 64, 128, 256, and 512, respectively. Residual connections are used to ensure the stable transmission of gradient information during deep network training. The feature fusion layer integrates multi-level features using a channel-weighted approach, with the weight coefficients adaptively adjusted according to three parameters: phase distribution density, discharge amplitude concentration, and phase span of the PRPD spectrum. The fully connected classification layer consists of a 256-node hidden layer and a four-node output layer.

[0032] in, The steps for establishing the training dataset for the PDCN model specifically include collecting 10,000 discharge spectra of different types from a publicly available GIS partial discharge database, extracting physical feature vectors from each spectra, and including parameters such as discharge initiation phase, discharge termination phase, discharge amplitude distribution, and discharge frequency distribution. Simultaneously, data under different operating conditions are also collected. Gas insulation characteristic parameters were used to establish a mapping relationship between physical parameters and spectral features. Data augmentation techniques were used to expand the dataset to 30,000 images. These data augmentation techniques included operations such as rotation, translation, scaling, and adding Gaussian noise. The training dataset, validation dataset, and test dataset were divided in a 7:2:1 ratio to ensure that the model has sufficient generalization ability.

[0033] in, The PDCN model incorporates partial discharge physical constraints into the loss function during training. These constraints are constructed based on discharge characteristic parameters calculated from the partial discharge physical mechanism equation, ensuring that the model's predictions conform to physical laws. An early stopping strategy is employed to avoid overfitting; this strategy involves stopping training when the accuracy on the validation dataset shows no significant improvement for five consecutive cycles. Finally, the model's performance is evaluated on an independent test dataset. Performance metrics include recognition accuracy, F1 score, and confusion matrix, ensuring a recognition accuracy of over 95%.

[0034] Among them, the physical parameter constraints of partial discharge are the relationship between theoretical parameter values ​​calculated by the physical mechanism equation of partial discharge and actual measured values. This relationship guides and constrains the learning process of the neural network. The physical parameter constraints of partial discharge include discharge initiation voltage constraints, discharge phase distribution constraints, and discharge amplitude distribution constraints. By integrating physical laws into the deep learning training process, the generalization ability and interpretability of the model are improved.

[0035] It should be noted that integrating the ECA attention mechanism between the encoder and decoder of the improved SAE network, and combining it with the pre-trained SF6PDCN model to assist feature extraction, forms a highly efficient hybrid feature enhancement architecture. The ECA attention mechanism effectively highlights key channel information by adaptively selecting the convolution kernel size and finely allocating weights to feature channels. This mechanism first captures the feature profile of each channel through global average pooling, then uses one-dimensional convolution to achieve information interaction between channels, and finally generates importance weights for each channel using the Sigmoid function, enabling the network to focus on salient features related to discharge patterns. Simultaneously, the pre-trained SF6PDCN model provides domain expertise for feature extraction; its multi-scale residual structure can capture discharge features from different scales and maintain information integrity through residual connections. The combination of these two mechanisms enables top-down feature enhancement and bottom-up knowledge transfer, which has three major advantages over traditional methods: First, it improves the ability to identify small sample discharge patterns and reduces the dependence on large-scale labeled data; second, it enhances the robustness of the model to noise and interference, especially in complex power system environments; and third, by integrating physical constraints and deep learning, it improves the interpretability and technical reliability of the model, making the identification results more consistent with the physical laws of partial discharge.

[0036] It should be noted that partial discharge physical parameter constraints are a method that integrates theoretical physical models with deep learning. Its core lies in using the relationship between theoretical parameter values ​​calculated from the partial discharge physical mechanism equations and actual measured values ​​to guide and constrain the neural network learning process. This physical constraint mechanism is mainly achieved through three aspects: discharge initiation voltage constraint, discharge phase distribution constraint, and discharge amplitude distribution constraint. When the recognition result output by the neural network deviates from the physical laws, the physical parameter constraints apply an additional penalty term to adjust the network parameters, ensuring that the model learning process follows the physical laws of partial discharge in SF6 gas medium. Compared to traditional purely data-driven methods, this physical constraint mechanism has significant advantages: First, it greatly improves the model's generalization ability, enabling the network to correctly handle discharge patterns not present in the training data; second, it improves the model's robustness in environments with sparse data or noise interference, reducing the risk of overfitting; third, it endows the neural network with interpretability through physical mechanisms, making the black-box model's decision-making process more transparent and reliable; finally, physical parameter constraints also reduce the dependence on large-scale labeled data, achieving efficient complementarity between physical knowledge and data learning, enabling the model not only to accurately identify discharge types but also to predict discharge behavior under different operating conditions, providing a more reliable technical foundation for equipment condition assessment and predictive maintenance.

[0037] The specific implementation methods of the above steps are described in detail below.

[0038] The specific implementation of step S01 involves constructing a GIS partial discharge experimental platform conforming to the IEC 60270 standard. First, an AC power supply with a rated capacity of 100kVA and a maximum output voltage of 50kV is selected as the voltage source, connected to a step-up transformer with a transformation ratio of 220V:50kV to provide sufficient experimental voltage. A high-power 5kΩ resistor is used as the protection resistor, primarily to limit the discharge current amplitude and prevent equipment damage. An 800pF high-voltage ceramic capacitor is selected as the coupling capacitor, providing a transmission path for the discharge signal and used for measurement system calibration. A wideband impedance with a frequency response range of 10kHz to 10MHz is selected for the detection impedance, used to convert the high-frequency discharge pulse signal into a measurable voltage signal. The GIS chamber is filled with a pressure of 0.1MPa. The gas, the pressure value is selected based on The gas exhibits good insulating properties at this pressure, closely resembling the operating conditions of real-world equipment. The MPD 800 partial discharge detection system's center frequency was set to 250kHz, with a bandwidth of 100kHz to 400kHz. This frequency range was chosen to effectively capture discharge signals and suppress external interference. Ambient noise was controlled to approximately 200fC through shielding, signal filtering, and grounding optimization to ensure the signal-to-noise ratio of the measured signal met experimental requirements. The purpose of this step was to establish a standardized experimental platform, providing a reliable foundation for subsequent defect model discharge data acquisition.

[0039] The specific implementation of step S02 involves designing and fabricating four typical GIS defect models. The tip discharge model uses a copper needle tip to a flat plate electrode structure with a needle tip curvature radius of 0.05 mm and a distance of 3 mm between the needle tip and the plate. The test voltage is set to 4.9 kV. The surface discharge model is made of epoxy resin insulating material, with a 15 mm long triangular surface discharge channel on its surface. The test voltage is set to 3.5 kV. The air gap discharge model consists of two parallel metal electrodes with a 0.5 mm thick insulating pad inserted between them to form an air gap. The test voltage is set to 6.9 kV. The suspended metal discharge model suspends a 2 mm diameter metal ball near the high-voltage electrode at a distance of 4 mm from the electrode. The test voltage is set to 7.5 kV. The test voltage value for each defect type is determined based on 1.2 times the discharge initiation voltage calculated from the partial discharge physical mechanism equation to ensure stable discharge. During data acquisition, 500 PRPD spectra for each defect type are continuously acquired at 2-minute intervals. The acquisition interval is selected based on the partial discharge stability analysis results to ensure the uniformity and representativeness of the time distribution of the acquired data. The purpose of this step is to build a standardized defect model and collect sufficient data samples to provide a diverse data foundation for subsequent model training and identification.

[0040] The specific implementation of step S03 involves preprocessing the acquired PRPD spectral data. First, the image is converted from the BGR color space to the HSV color space, defining the HSV threshold range for gray and black grids as: H value 0 to 180, S value 0 to 30, and V value 100 to 220. Based on this threshold range, a binarization method is used to extract the coordinate axes and grid line regions in the image, generating a mask image. Then, the pixels in the identified grid line regions are replaced with the background color, while keeping the discharge signal regions unchanged. Next, the processed image is uniformly adjusted to a size of 256×256 pixels, and a bicubic interpolation algorithm is used for image scaling, which can achieve smooth scaling while preserving image details. Normalization processing uses the Min-Max normalization method to map pixel values ​​to the 0-1 range. PRPD spectral feature enhancement is performed based on the partial discharge physical mechanism equation. Specifically, an adjustment coefficient α is calculated using the discharge initiation voltage Ui and the discharge amplitude q. The adjustment coefficient typically ranges from 1.05 to 1.3. Subsequently, nonlinear contrast enhancement processing is performed on the image based on the α value to enhance the visual salience of the concentrated discharge region. The purpose of this step is to remove irrelevant information from the image, retain key discharge features, and standardize the data into a format that is easy for neural networks to process.

[0041] The specific implementation of step S04 involves constructing an improved SAE network model. The network encoder consists of three convolutional layers: the first layer uses 64 3×3 filters with a stride of 1 and padding of 1; the second layer uses 128 3×3 filters with a stride of 2 and padding of 1; and the third layer uses 256 3×3 filters with a stride of 2 and padding of 1. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, with the momentum parameter of the batch normalization layer set to 0.9. The decoder uses three transposed convolutional layers for upsampling: the first layer uses 128 3×3 filters with a stride of 2 and padding of 1; the second layer uses 64 3×3 filters with a stride of 2 and padding of 1; and the third layer uses three 3×3 filters with a stride of 1 and padding of 1. An ECA attention mechanism is fused between the encoder output and decoder input. First, channel descriptors are obtained through global average pooling. Then, a one-dimensional convolution with kernel size k is used to achieve inter-channel information exchange, where the value of k is adaptively determined by the number of channels C: k = Typically, for 256 channels, k is set to 9. Finally, weight coefficients are generated using the Sigmoid activation function to weight the feature channels. The network also utilizes pre-trained... PDCN model-assisted feature extraction, The output features of the second and third multi-scale residual blocks of the PDCN model are fused with the corresponding layer features of the encoder of the improved SAE network, with a fusion weight ratio of 0.3:0.7. The purpose of this step is to construct a deep learning model with efficient feature extraction and reconstruction capabilities, and to enhance the ability to identify partial discharge features by combining attention mechanisms and pre-trained models.

[0042] The specific implementation of step S05 involves model training. First, the 2000 PRPD spectral images in the dataset are randomly divided into a training set of 1400 images and a test set of 600 images at a 7:3 ratio. The training set is then further divided into actual training samples (1260 images) and validation samples (140 images) at a 9:1 ratio. Mini-batch gradient descent with a batch size of 32 is used for model training, with an initial learning rate of 0.001. The Adam optimizer is used, and its parameters are... Set it to 0.9. Set it to 0.999. Set as The training process employs a cosine annealing learning rate adjustment strategy, with the minimum learning rate set to 0.01 times the initial learning rate, and the number of training epochs set to 100. The loss function is designed as a weighted sum of the reconstruction loss and the classification loss. ,in Mean squared error loss, used to evaluate reconstruction quality; Cross-entropy loss is used to evaluate classification accuracy. To account for the constraint loss of partial discharge physical parameters, the deviation between the predicted parameters and the theoretical parameters is calculated, and the weighting coefficients are used. =0.4, =0.5, =0.1. Physical parameter constraints include discharge initiation phase constraints (threshold ±15°), discharge amplitude distribution constraints (threshold ±20%), and discharge frequency distribution constraints (threshold ±15%). An early stopping strategy is employed during training; training stops when the validation set loss shows no significant decrease for five consecutive rounds. The purpose of this step is to optimize network parameters so that the model can learn both the visual features of PRPD spectra and conform to the physical laws of partial discharge. Next, model testing and discharge type identification are performed. First, 600 PRPD spectra from the test set are input into the trained improved SAE network model. Feature processing includes: feature extraction through the encoder, with the final convolutional layer outputting a feature map of size 64×64×256; global average pooling dimensionality reduction is applied to compress the feature map into a 1×1×256 feature vector; features are processed through a fully connected layer containing 128 neurons, with ReLU activation function; finally, a Softmax output layer with 4 neurons is used to calculate the probability distribution of the four defect types. During prediction, a probability threshold of 0.75 is set; samples below this threshold are marked as uncertain and require further analysis. Simultaneously, the identification results are verified by constraining the physical parameters of partial discharge, and the consistency index (CI) between the predicted discharge parameters and the theoretical parameters is calculated: ,in To predict parameter values, These are theoretical parameter values. Tolerance for deviation. When CI is below 0.85, a warning is triggered and manual review is recommended. The recognition results are evaluated using metrics such as accuracy, precision, recall, and F1 score, requiring an overall recognition accuracy of no less than 95% and an F1 score of no less than 0.92 for each defect type. The purpose of this step is to verify the model's recognition performance and ensure the reliability and accuracy of the recognition results through physical parameter constraints.

[0043] The specific implementation of step S06 is model application. Specifically, the GIS device under test and the processed PRPD spectral data are input into the trained improved SAE network model to identify the defect type and obtain the identification result.

[0044] It should be noted that, The detailed structure of the PDCN model is as follows: This model is based on a deep convolutional neural network architecture with multi-scale residual blocks, comprising an initial convolutional layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. The initial convolutional layer uses 64 7×7 convolutional kernels with a stride of 2 and padding of 3 to extract features from an input 256×256×3 PRPD spectral map, with an output size of 128×128×64. Each multi-scale residual block contains three parallel convolutional branches using convolutional kernels of different sizes: 3×3, 5×5, and 7×7, capturing discharge features at different scales. The 5×5 and 7×7 convolutions are implemented by stacking multiple 3×3 convolutions to reduce the number of parameters. The first multi-scale residual block outputs 64 channels, the second 128 channels, the third 256 channels, and the fourth 512 channels. Each residual block also contains a shortcut connection of 1×1 convolutions to ensure consistency between the number of input and output channels. The feature fusion layer integrates multi-level features using a channel-weighted approach, with weight coefficients... , , The weights, corresponding to shallow, mid, and deep features respectively, are initially set to 0.2, 0.3, and 0.5, and are adaptively adjusted during network training. The fully connected classification layer consists of a 256-node hidden layer and a 4-node output layer. The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function. The model has approximately 8.5 million parameters. Its design philosophy is to capture fine-grained features of partial discharge through multi-scale feature extraction, while ensuring stable gradient information transmission during deep network training through residual connections.

[0045] The specific implementation method for establishing the training dataset of the PDCN model is as follows: First, 10,000 discharge spectra of different types are collected from three publicly available GIS partial discharge databases, including the EPRI partial discharge database, the IEC partial discharge standard atlas, and the power equipment intelligent diagnostic open data platform, covering different voltage levels and different... Discharge samples under gas pressure conditions. Physical feature vectors, including the discharge initiation phase, are extracted from each spectrum. Discharge end phase Discharge amplitude distribution function Discharge frequency distribution function Parameters such as discharge start and end phases were determined using a threshold method based on the phase histogram, with the threshold set at 5% of the maximum frequency. Discharge amplitude distribution was obtained through amplitude statistics within a 95% confidence interval. Discharge frequency distribution was derived through statistical analysis of the number of discharges per 1° within a 360° phase interval. Simultaneously, data from different operating conditions were collected. Gas insulation characteristic parameters, including gas pressures ranging from 0.1 to 0.6 MPa, preferably 0.1 to 0.4 MPa, are defined as breakdown field strength, ionization coefficient, adhesion coefficient, and diffusion coefficient per 0.1 MPa. A mapping relationship between physical parameters and spectral characteristics is established. A polynomial fitting method is used to construct the functional relationship between gas pressure, temperature, electrode structure, and discharge characteristic parameters. The fitting accuracy requirement is... Greater than 0.95. Data augmentation techniques were used to expand the dataset to 30,000 images. These augmentation operations included: rotation ±10°, translation ±5%, scaling from 0.95 to 1.05 times, adding Gaussian noise with a standard deviation 10% of the original image's standard deviation, adjusting brightness ±10%, and adjusting contrast ±15%. The dataset was divided into a 7:2:1 ratio: 21,000 images for training, 6,000 images for validation, and 3,000 images for testing, ensuring a balanced distribution of defect samples across each subset. The percentages of each defect type were: tip discharge 25%, surface discharge 25%, air gap discharge 25%, and suspended metal discharge 25%.

[0046] It should be noted that the physical mechanism equation of partial discharge plays a core theoretical supporting role in the entire identification method, revealing at the microscopic level... The fundamental laws governing discharge processes in gaseous media are revealed by this equation. This equation establishes a rigorous connection between macroscopically measurable external conditions such as electric field strength, gas pressure, and temperature, and the microscopic electron collision ionization process. By analyzing the motion behavior and energy accumulation process of gas particles under the influence of an electric field, it predicts the discharge initiation voltage threshold and the evolution trend of discharge amplitude. The significant role of this mechanistic equation lies in introducing prior physical constraints to neural networks. This ensures that model training not only relies on statistical data but also incorporates the fundamental physical laws of gas discharge, thus maintaining high accuracy and strong generalization ability even with small sample sizes. Particularly when the parameters of defective models deviate from standard operating conditions, the theoretical predictions provided by the physical mechanism equation can effectively correct the neural network output, reducing the misjudgment rate under abnormal conditions.

[0047] It's important to note that the SAE network, as a deep autoencoder structure, primarily excels in feature extraction and dimensionality reduction. Unlike traditional convolutional neural networks that directly perform classification learning, SAE first learns the intrinsic representation of the data in an unsupervised manner, compressing the high-dimensional PRPD spectrum into a low-dimensional latent feature space. Then, it verifies the fidelity of the extracted features through a decoding and reconstruction process. This self-supervised learning mechanism allows SAE to learn more essential feature representations from limited samples, avoiding the overfitting trap that direct classification models often fall into. SAE's multi-layered stacked design allows the network to extract hierarchical features from low-level texture to high-level semantics, forming a deep understanding of the intrinsic structure of the PRPD spectrum. Improved SAE networks, by introducing convolutional operations and pooling layers, further enhance their ability to extract spatial local features, making them particularly suitable for handling complex situations where local morphological features and global distribution patterns coexist in discharge spectra.

[0048] It's important to note that the ECA attention mechanism plays a crucial role in this method, adaptively weighting channel features. This mechanism first obtains the global response value for each channel through global average pooling, then cleverly utilizes one-dimensional convolution operations to capture the dependencies between channels, avoiding the parameter redundancy problem of fully connected layers in traditional attention mechanisms. The design philosophy of ECA is to efficiently extract channel importance weights through a lightweight structure, enabling the network to automatically focus on feature channels related to specific discharge types while suppressing the influence of irrelevant or noisy channels. In partial discharge identification tasks, the discharge features of different types of defects often concentrate in specific frequency or phase ranges. ECA attention capitalizes on this characteristic, enhancing the model's ability to perceive key regions and improving the discriminative power of feature representation through adaptive weight allocation.

[0049] It should be noted that, The PDCN model, as a pre-trained deep convolutional network, serves as a feature extractor for the domain. Pre-trained on large-scale GIS discharge data, it has learned general feature representations of discharge spectra, enabling it to identify key texture and morphological features of different defect types. The multi-scale residual block design allows the network to simultaneously capture fine-grained local features and large-scale global patterns in the discharge spectra, while residual connections effectively solve the gradient vanishing problem in deep network training, ensuring effective learning of deep features. The feature fusion layer, through adaptive channel weighting, dynamically adjusts feature importance based on the characteristics of the input spectrum, further enhancing the model's adaptability to different discharge patterns. The synergistic effect of the PDCN model and the improved SAE network enables the complementary fusion of general features and task-specific features, significantly improving the robustness and accuracy of the overall recognition system.

[0050] It should be noted that the physical parameter constraints for partial discharge serve as a bridge connecting physical mechanisms and deep learning, playing a role in both guiding the training process and verifying inference results. During the training phase, physical parameter constraints participate in network optimization through a loss function, guiding the model to learn feature representations that conform to physical laws and avoiding potential physical inconsistencies that may arise from relying solely on data-driven approaches. During the inference phase, physical parameter constraints act as a post-processing mechanism, verifying the physical rationality of the neural network's output and filtering out erroneous judgments that clearly violate the physical laws of discharge. This constraint mechanism is particularly suitable for application scenarios such as offshore wind power, which involve complex environments and variable operating conditions, effectively improving the model's generalization ability when facing non-standard operating conditions and novel defects in actual operation.

[0051] Specifically, the principle of this invention is: The core technical principle of this invention lies in the organic integration of physical mechanism knowledge with deep learning models, achieving efficient identification of partial discharge modes through an improved SAE network structure. Its principle is mainly reflected in the following aspects: First, this invention establishes a mapping relationship between the physical mechanism of partial discharge and the PRPD spectral features. By parameterizing the physical process of partial discharge in SF6 gas, a physical mechanism equation is established that includes input variables such as electric field strength, gas pressure, temperature, gas gap distance, and relative permittivity of the medium, outputting discharge initiation voltage and discharge amplitude parameters. These physical parameters are directly used for feature enhancement of the PRPD spectrum, integrating physical mechanism information into the data preprocessing stage and providing more physically meaningful input features for subsequent deep learning models.

[0052] Secondly, the improved SAE network structure of this invention is optimized for partial discharge recognition tasks. Traditional autoencoders often suffer from information loss when processing high-dimensional image data. This invention extracts features at different scales using an encoder composed of three convolutional layers, and uses an upsampling layer in the decoder to restore the image spatial dimension, preserving key features of the discharge pattern. Simultaneously, the introduction of an ECA attention mechanism enables adaptive weight allocation for feature channels, allowing the network to focus on more discriminative features and improving the model's feature learning ability.

[0053] Furthermore, this invention will pre-train... The PDCN model incorporates an improved SAE network, utilizing transfer learning mechanisms to enhance feature extraction capabilities. The PDCN model captures discharge features at different scales through multi-scale residual blocks, and incorporates partial discharge physical constraints during training, making the extracted features more consistent with the physical laws of discharge. This pre-trained model not only improves the efficiency of feature extraction but also enhances the model's adaptability to discharge features under different environmental conditions.

[0054] Most importantly, this invention introduces constraints on the physical parameters of partial discharge during deep learning training, incorporating discharge initiation voltage constraints, discharge phase distribution constraints, and discharge amplitude distribution constraints into the loss function. This approach enables the network to learn not only statistical characteristics of the data but also feature representations that conform to physical laws, thus maintaining high recognition capabilities even when faced with unseen discharge patterns or complex environmental conditions. The introduction of physical constraints essentially expands the effective learning space of the model, enabling it to focus not only on "what" but also on understanding "why," thereby improving the model's generalization ability and interpretability.

[0055] In summary, this invention effectively solves the problems of insufficient feature extraction and limited generalization ability in traditional methods by integrating physical mechanisms with deep learning, providing a theoretical basis and implementation path for the accurate identification of partial discharge modes in offshore wind power GIS equipment.

[0056] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0057] The specific implementation of step S01 involves constructing a GIS partial discharge experimental platform conforming to the IEC 60270 standard. First, an AC power supply with a rated capacity of 100kVA and a maximum output voltage of 50kV is selected as the voltage source, connected to a step-up transformer with a transformation ratio of 220V:50kV to provide sufficient experimental voltage. A high-power 5kΩ resistor is used as the protection resistor, primarily to limit the discharge current amplitude and prevent equipment damage. An 800pF high-voltage ceramic capacitor is selected as the coupling capacitor, providing a transmission path for the discharge signal and used for measurement system calibration. A wideband impedance with a frequency response range of 10kHz to 10MHz is selected for the detection impedance, used to convert the high-frequency discharge pulse signal into a measurable voltage signal. The GIS chamber is filled with a pressure of 0.1MPa. The gas, the pressure value is selected based on The gas exhibits good insulating properties at this pressure, closely resembling the operating conditions of real-world equipment. The MPD 800 partial discharge detection system's center frequency was set to 250kHz, with a bandwidth of 100kHz to 400kHz. This frequency range was chosen to effectively capture discharge signals and suppress external interference. Ambient noise was controlled to approximately 200fC through shielding, signal filtering, and grounding optimization to ensure the signal-to-noise ratio of the measured signal met experimental requirements. The purpose of this step was to establish a standardized experimental platform, providing a reliable foundation for subsequent defect model discharge data acquisition.

[0058] The specific implementation of step S02 involves designing and fabricating four typical GIS defect models. The tip discharge model uses a copper needle tip to plate electrode structure with a needle tip curvature radius of 0.05 mm and a distance of 3 mm between the needle tip and the plate. The test voltage is set to 4.9 kV. The surface discharge model is made of epoxy resin insulating material, with a 15 mm long triangular surface discharge channel on its surface. The test voltage is set to 3.5 kV. The air gap discharge model consists of two parallel metal electrodes, with a 0.5 mm thick insulating pad inserted between the electrodes to form an air gap. The test voltage is set to 6.9 kV. The suspended metal discharge model suspends a 2 mm diameter metal ball near the high-voltage electrode, at a distance of 4 mm from the electrode. The test voltage is set to 7.5 kV. The test voltage value for each defect type is determined based on 1.2 times the discharge initiation voltage calculated from the partial discharge physical mechanism equation to ensure stable discharge. According to the Paschen condition, the discharge initiation voltage... The calculation formula is: ; In the formula, The gas constant is... The gas value is 27.0 V·Pa. ·m ; This refers to gas pressure, measured in Pa. The distance between the electrodes is in meters (m). The ionization coefficient, The gas value is 0.025 Pa. ·m ; The secondary electron emission coefficient ranges from 0.01 to 0.1. During data acquisition, 500 PRPD spectra for each defect type were continuously acquired at 2-minute intervals. The acquisition interval was selected based on the partial discharge stability analysis results to ensure the uniformity and representativeness of the temporal distribution of the acquired data. The purpose of this step is to construct a standardized defect model and collect sufficient data samples to provide a diverse data foundation for subsequent model training and identification.

[0059] Step S03 specifically involves preprocessing the acquired PRPD spectral data. First, the image is converted from the BGR color space to the HSV color space, defining the HSV threshold ranges for the gray and black grids as: H value 0 to 180, S value 0 to 30, and V value 100 to 220. Based on these threshold ranges, a binarization method is used to extract the coordinate axes and grid line regions from the image, generating a mask image. The binarization process can be represented as: ; In the formula, For the mask image in coordinates The value at; , , The original image is located at coordinates The hue, saturation, and brightness values ​​at that location; , , , , , These represent the lower and upper limits of the HSV threshold range, respectively. Then, the pixels in the identified grid line areas are replaced with the background color, while keeping the discharge signal area unchanged. Next, the processed image is uniformly resized to 256×256 pixels using a bicubic interpolation algorithm for image scaling, which can achieve smooth scaling while preserving image details. Normalization is performed using the Min-Max normalization method, mapping pixel values ​​to the 0-1 range. ; In the formula, To normalize the image in coordinates Pixel value at; The original image in coordinates Pixel value at; and These represent the minimum and maximum pixel values ​​of the original image, respectively. PRPD spectral feature enhancement is performed based on the partial discharge physical mechanism equation, utilizing the discharge initiation voltage. and discharge amplitude Calculate the adjustment factor : ; In the formula, This represents the measured discharge amplitude. This is the theoretical reference discharge amplitude. This is the test voltage; This is the discharge initiation voltage. The adjustment factor typically ranges from 1.05 to 1.3, and then... Non-linear contrast enhancement processing is applied to the image: ; In the formula, To enhance the image in coordinates Pixel value at; To normalize the image in coordinates The pixel values ​​at that location. The purpose of this step is to remove irrelevant information from the image, preserve key discharge features, and standardize the data into a format that is easy for neural networks to process.

[0060] The specific implementation of step S04 is to construct an improved SAE network model. The network encoder part contains three convolutional layers. The first layer uses 64 3×3 filters with a stride of 1 and padding of 1; the second layer uses 128 3×3 filters with a stride of 2 and padding of 1; and the third layer uses 256 3×3 filters with a stride of 2 and padding of 1. Each convolutional operation can be represented as: ; In the formula, For the first Layer feature map; For the first Layer convolution kernel weights; For the first Layer feature map; For the first Layer bias term; Indicates the convolution operation; The activation function is represented by the ReLU function: Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The formula for calculating the batch normalization layer is as follows: ; ; In the formula, Input features; This represents the mean of a small batch of data. The variance of the small batch of data; To prevent small constants from being divided by zero, the value is taken as... ; and These are learnable parameters; This is the output of the batch normalization layer. The momentum parameter of the batch normalization layer is set to 0.9. The decoder uses three transposed convolutional layers for upsampling. The first layer uses 128 3×3 filters with a stride of 2 and padding of 1; the second layer uses 64 3×3 filters with a stride of 2 and padding of 1; and the third layer uses three 3×3 filters with a stride of 1 and padding of 1. An ECA attention mechanism is fused between the encoder output and the decoder input. First, channel descriptors are obtained through global average pooling. ; In the formula, For the first Descriptors for each channel; For the feature map Each channel is located in The value at; and These represent the height and width of the feature map, respectively. Then, a convolution kernel with a size of [size missing] is applied. One-dimensional convolution enables information exchange between channels, where The value is determined by the number of channels. Adaptive determination: ; When the number of channels hour, The value is 9. A one-dimensional convolution operation is represented as: ; In the formula, For the first The weighting coefficients of each channel; These are the convolution kernel parameters; For channel descriptors; For the Sigmoid activation function: Finally, weight coefficients are generated using the Sigmoid activation function to weight the feature channels. ; In the formula, For the weighted feature map, the first... One channel; These are the weighting coefficients; For the original feature map, the first One channel. The network simultaneously uses pre-trained... PDCN model-assisted feature extraction, The output features of the second and third multi-scale residual blocks of the PDCN model are fused with the corresponding layer features of the encoder of the improved SAE network, with a fusion weight ratio of 0.3:0.7: ; In the formula, Features after fusion; for PDCN model output features; To improve the output features of the SAE network encoder, this step aims to build a deep learning model with efficient feature extraction and reconstruction capabilities, combining attention mechanisms and pre-trained models to enhance the ability to identify partial discharge features.

[0061] The specific implementation of step S05 involves model training. First, the 2000 PRPD spectral images in the dataset are randomly divided into a training set of 1400 images and a test set of 600 images at a 7:3 ratio. The training set is then further divided into actual training samples (1260 images) and validation samples (140 images) at a 9:1 ratio. Mini-batch gradient descent with a batch size of 32 is used for model training, with an initial learning rate of 0.001. The Adam optimizer is used, and its parameters are... Set it to 0.9. Set it to 0.999. Set as The training process employs a cosine annealing learning rate adjustment strategy. The learning rate is calculated using the following formula: ; In the formula, For the first The learning rate of each training round; The minimum learning rate is set to 0.01 times the initial learning rate. The maximum learning rate is set to an initial learning rate of 0.001. This is the current training round number; Let 100 be the total number of training rounds. The loss function is designed as a weighted sum of the reconstruction loss and the classification loss: ; In the formula, Mean squared error loss, used to evaluate reconstruction quality: ; Cross-entropy loss is used to evaluate classification accuracy. ; To account for the constraint loss of partial discharge physical parameters, the deviation between the predicted parameters and the theoretical parameters is calculated: ; In the formula, The number of samples; For the first The true value of each sample; For the first Predicted values ​​for each sample; This represents the number of categories, which is 4 in this example. For the first The sample belongs to the first The true probability of a class; For the first The sample belongs to the first The predicted probability of a class; The number of physical parameters; For the first Weighting coefficients for each physical parameter; For the first Predicted values ​​of each physical parameter; For the first The theoretical values ​​of each physical parameter; For the first The allowable deviation of each physical parameter. Weighting coefficient. , , The physical parameter constraints include discharge initiation phase constraints (threshold ±15°), discharge amplitude distribution constraints (threshold ±20%), and discharge frequency distribution constraints (threshold ±15%). An early stopping strategy is employed during training; training is halted when the validation set loss shows no significant decrease for five consecutive rounds. The purpose of this step is to optimize the network parameters so that the model can learn both the visual features of the PRPD spectrum and conform to the physical laws of partial discharge.

[0062] Next, model testing and discharge type identification are performed. First, 600 PRPD spectra from the test set are input into the trained improved SAE network model. Feature processing includes: feature extraction via the encoder, with the final convolutional layer outputting a feature map of size 64×64×256; and global average pooling for dimensionality reduction, compressing the feature map into a 1×1×256 feature vector. ; In the formula, For the first Global characteristics of each channel; For the feature map Each channel is located in The value at; and These represent the height and width of the feature map, respectively. Features are processed through a fully connected layer with 128 neurons, using ReLU activation. ; In the formula, Output for hidden layer; This is the weight matrix; The input feature vector; It is the bias vector; This represents the ReLU activation function. Finally, a softmax output layer with four neurons is used to calculate the probability distributions for the four defect types: ; In the formula, For the sample to belong to the first The probability of a class; For the output layer The input of each neuron; For the output layer The input to each neuron. During the prediction process, a probability threshold of 0.75 is set; samples below this threshold are marked as uncertain and require further analysis. Simultaneously, the identification results are validated using constraints from partial discharge physical parameters, and the consistency index (CI) between the predicted and theoretical discharge parameters is calculated. ; In the formula, It is a consistency index; For the first Predicted values ​​of each physical parameter; For the first The theoretical values ​​of each physical parameter; For the first The allowable deviation of each physical parameter; This represents the number of physical parameters. When the CI is below 0.85, a warning is triggered and manual review is recommended. The identification results are evaluated using metrics such as accuracy, precision, recall, and F1 score. The accuracy calculation formula is: ; The formula for calculating accuracy is: ; The formula for calculating recall rate is: ; The formula for calculating the F1 score is: ; In the formula, The number of true cases; The number of true negative examples; The number of false positives; This represents the number of false negatives. The overall recognition accuracy must be no less than 95%, and the F1 score for each type of defect must be no less than 0.92. The purpose of this step is to verify the model's recognition performance and ensure the reliability and accuracy of the recognition results through physical parameter constraints.

[0063] The detailed structure of the PDCN model is as follows: This model is based on a deep convolutional neural network architecture with multi-scale residual blocks, containing an initial convolutional layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. The initial convolutional layer uses 64 7×7 convolutional kernels with a stride of 2 and padding of 3 to extract features from an input 256×256×3 PRPD spectral map, with an output size of 128×128×64. Each multi-scale residual block contains three parallel convolutional branches using convolutional kernels of different sizes: 3×3, 5×5, and 7×7, capturing discharge features at different scales. The 5×5 and 7×7 convolutions are implemented by stacking multiple 3×3 convolutions to reduce the number of parameters. The computation of the residual block can be represented as: ; In the formula, Output for residual blocks; Input for the residual block; For the residual function, Here are the parameters for the convolutional layer. The feature extraction process for multi-scale residual blocks can be represented as: ; In the formula, The results are from multi-scale feature extraction. , , The features extracted using 3×3, 5×5, and 7×7 convolutional kernels are respectively; This represents a feature concatenation operation along the channel dimension. The first multi-scale residual block outputs 64 channels, the second 128 channels, the third 256 channels, and the fourth 512 channels. Each residual block also contains a 1×1 convolutional shortcut connection to ensure the number of input and output channels is consistent. The feature fusion layer integrates multi-level features using a channel-weighted approach, with weight coefficients... , , The weights correspond to the shallow, middle, and deep features, respectively, with initial values ​​of 0.2, 0.3, and 0.5, which are adaptively adjusted during network training. ; In the formula, Features after fusion; , , These are characteristics of the shallow, intermediate, and deep layers, respectively. , , Let be the weight coefficient, and satisfy... The weighting coefficients are based on the phase distribution density of the PRPD spectrum. Discharge amplitude concentration Phase span Three parameters are adaptively adjusted: ; In the formula, This is the weight matrix; It is the bias vector; The normalization function is used. The fully connected classification layer consists of a 256-node hidden layer and a 4-node output layer. The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function. The model has approximately 8.5 million parameters. Its design philosophy is to capture fine-grained features of partial discharge through multi-scale feature extraction, while ensuring stable gradient information transmission during deep network training through residual connections.

[0064] The specific implementation of step S06 is model application, which involves inputting the GIS device under test and the processed PRPD spectral data into the trained improved SAE network model to identify the defect type and obtain the identification result. To better understand and implement this invention, a specific application scenario is provided below as Example 2: A research team applied the method of this invention to a 35kV GIS device in an offshore wind farm to improve the fault diagnosis capability of the GIS device in an offshore environment. The researchers first constructed a partial discharge experimental platform according to the IEC60270 standard, as shown in the schematic diagram below. Figure 2 As shown, the actual experimental deployment is as follows: Figure 5 As shown, a 100kVA AC power supply was used, with the experimental voltage provided through a 220V:50kV step-up transformer. The platform was equipped with a 5kΩ protection resistor, an 800pF coupling capacitor, and a detection impedance with a frequency response range of 10kHz to 10MHz. The GIS chamber was filled with a pressure of 0.1MPa. For the gas, the center frequency of the MPD800 partial discharge detection system is set to 250kHz, and the ambient noise is controlled at around 185fC.

[0065] Researchers designed four typical GIS defect models, such as Figure 3 As shown, the following models were tested: the tip discharge model used a copper needle with a radius of curvature of 0.05 mm on a flat plate, with a 3 mm gap between the needle tip and the plate, and a test voltage of 4.9 kV; the surface discharge model was made of epoxy resin and featured a 15 mm long triangular surface discharge channel, with a test voltage of 3.5 kV; the air gap discharge model consisted of two parallel metal electrodes with a 0.5 mm thick insulating pad inserted between them, with a test voltage of 6.9 kV; the suspended metal discharge model had a 2 mm diameter metal ball suspended near the high-voltage electrode at a distance of 4 mm from the electrode, with a test voltage of 7.5 kV. The statistical data samples collected by the researchers for each defect type are shown in Table 1. Table 1. Statistics on Partial Discharge Defect Types

[0066] The acquired PRPD spectra were preprocessed using color thresholding in the HSV color space. The defined HSV threshold ranges were: H value 0 to 180, S value 0 to 30, and V value 100 to 220. The processed images were then uniformly resized to 256×256 pixels and normalized. Figure 4 As shown. Adjustment coefficients calculated based on the physical mechanism equation of partial discharge. The value distribution is shown in Table 2: Table 2. Statistics of Feature Enhancement Adjustment Coefficients for Different Defect Types

[0067] The specific structural parameters of the improved SAE network model are shown in Table 3: Table 3 Structural parameters of the improved SAE network model

[0068] The 2000 PRPD spectral images were divided into a training set of 1400 images and a test set of 600 images in a 7:3 ratio. The training process employed mini-batch gradient descent with a batch size of 32, an initial learning rate of 0.001, and loss function weights set to... , , The changes in loss values ​​at each stage of the training process are shown in Table 4: Table 4. Changes in loss value during model training.

[0069] The evaluation metrics for the recognition results of the 600 PRPD spectra in the test set are shown in Table 5: Table 5 Performance evaluation results for different discharge types

[0070] Traditional methods for identifying partial discharge in GIS equipment primarily rely on manually extracting feature parameters from phase maps, such as the ratio of positive to negative half-cycle discharge amplitudes, the maximum amplitude of the discharge pulse, and the discharge phase distribution. These parameters are then classified using machine learning methods such as support vector machines, random forests, and backpropagation neural networks. This approach suffers from several problems: first, the feature extraction process is highly subjective and struggles to fully represent all characteristics of the discharge pattern; second, traditional machine learning methods have limited ability to fit complex nonlinear relationships; and third, they lack consideration of the physical mechanisms of the discharge, resulting in weak generalization ability.

[0071] The improved SAE network method used in this embodiment has significant advantages: First, it automatically extracts high-level features of the PRPD spectrum through a deep convolutional network, avoiding the subjectivity and bias of manual feature extraction; second, it integrates the ECA attention mechanism, enabling the network to adaptively focus on key feature regions of different discharge modes; third, it introduces constraints from the physical mechanism of partial discharge, allowing the model to learn not only statistical laws from data but also conform to physical laws, significantly improving the model's generalization ability and interpretability; fourth, it utilizes pre-trained... The PDCN model-assisted feature extraction effectively solves the overfitting problem under small sample conditions. Experimental results show that compared with traditional methods, this method improves the recognition accuracy by 5.7%, the F1 score by 0.078, and the consistency index by 0.113. It also demonstrates better stability and anti-interference ability, especially in complex background noise environments.

[0072] It should be noted that the variables involved in this invention are explained in detail in Tables 6 and 7 below.

[0073] Table 6. Variable Explanation Table (Part 1)

[0074] Table 7. Variable Explanation Table (Part Two)

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for partial discharge pattern recognition of offshore wind power GIS equipment based on an improved SAE network, characterized in that, include: Construct a GIS partial discharge experimental platform; design four defect models based on the GIS partial discharge experimental platform and collect PRPD spectra; The acquired PRPD spectral data were preprocessed; an improved SAE network model was constructed, which fused an ECA attention mechanism between the encoder output and the decoder input, and a pre-trained... PDCN model assists in network feature extraction; A training dataset is built based on the preprocessed PRPD spectral data to train the improved SAE network model. The PRPD spectral data of the GIS device under test and the processed PRPD spectral data are input into the trained improved SAE network model for defect type identification to obtain the identification results. In the preprocessing step, the PRPD spectral features are also enhanced based on the partial discharge physical mechanism equation, which describes... The physical process of partial discharge in gaseous media and its relationship with the surrounding electric field distribution; The inputs to the partial discharge physical mechanism equation include electric field strength, gas pressure, temperature, air gap distance, and relative permittivity of the medium. The outputs are the discharge initiation voltage and discharge amplitude. The discharge amplitude is used for discharge amplitude calibration in the PRPD spectrum feature enhancement process. The encoder part of the improved SAE network model includes three convolutional layers that use 64, 128, and 256 filters respectively for feature extraction, and the decoder part restores the image spatial dimension through an upsampling layer. The ECA attention mechanism is an efficient channel attention mechanism that optimizes the weight distribution of feature channels by adaptively selecting the convolution kernel size. First, it obtains the description of each channel through global average pooling, and then uses one-dimensional convolution to perform information interaction between channels. The The PDCN model is a type of model used for A pre-trained deep convolutional neural network model for partial discharge identification in a gaseous environment is constructed using a deep convolutional neural network based on multi-scale residual blocks.

2. The method for partial discharge pattern recognition of offshore wind power GIS equipment based on improved SAE network according to claim 1, characterized in that, The steps for constructing the GIS partial discharge experimental platform are specifically based on the IEC 60270 standard. The experimental circuit consists of an AC power supply, a transformer, a protective resistor, a coupling capacitor, a detection impedance, a defect model, and a general partial discharge measurement and analysis system. The protective resistor is set to 5kΩ, the coupling capacitor to 800pF, and the GIS cavity is filled with gas at a pressure of 0.1MPa. The gas, the center frequency of the general partial discharge measurement and analysis system is set to 250kHz.

3. The method for partial discharge pattern recognition of offshore wind power GIS equipment based on improved SAE network according to claim 2, characterized in that, The four defect models are specifically pointed discharge, surface discharge, air gap discharge, and suspended metal discharge. The test voltage for pointed discharge is 4.9 kV, the test voltage for surface discharge is 3.5 kV, the test voltage for air gap discharge is 6.9 kV, and the test voltage for suspended metal discharge is 7.5 kV. 500 PRPD spectra are collected for each defect type at 2-minute intervals.

4. The method for partial discharge pattern recognition of offshore wind power GIS equipment based on improved SAE network according to claim 3, characterized in that, The preprocessing step for the acquired PRPD spectral data specifically involves using the color thresholding method of the HSV color space to remove the coordinate axes and grid lines from the original image, and then uniformly adjusting all images to a size of 256×256 and performing normalization processing.

5. The method for partial discharge pattern recognition of offshore wind power GIS equipment based on improved SAE network according to claim 4, characterized in that, The color thresholding method of the HSV color space converts the image from a color space to the HSV color space, defines an HSV range suitable for gray and black grids, extracts pixels in the image that meet the range through color thresholding, and replaces the grid lines and coordinate axes with the background color.

Images