Offshore wind power GIS equipment partial discharge mode identification method based on improved SAE network

Abstract

The invention provides an offshore wind power GIS equipment partial discharge mode identification method based on an improved SAE network, and belongs to the technical field of electrical digital data processing.The method includes the steps that a partial discharge experiment platform meeting the IEC60270 standard is constructed, four typical defect models are designed, and PRPD spectrograms are collected; hSV color space threshold segmentation is utilized to remove interference elements, and image features are enhanced in combination with a partial discharge physical mechanism equation. The improved SAE network integrates a three-layer convolution encoder and an up-sampling decoder, introduces an ECA attention mechanism to optimize a feature channel weight, and combines a pre-trained # imgabs0 # PDCN model to enhance the feature extraction capability. Physical parameter constraints are introduced in the training process, and after global pooling dimensionality reduction and full connection layer processing, high-precision identification of four defect types of point discharge, creeping discharge, air gap discharge and suspended metal discharge is realized, and the technical problems of low accuracy and poor generalization ability of offshore wind power GIS equipment partial discharge mode identification are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrical digital data processing, and in particular relates to a method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network. Background Art

[0002] Offshore wind power gas-insulated switchgear (GIS) is widely used in power systems due to its high reliability and miniaturization. Partial discharge (PD) within GIS equipment is a key indicator for assessing insulation condition. Real-time monitoring and accurate identification are crucial for ensuring the safe operation of offshore wind power systems. Traditional PD detection methods rely primarily on manual analysis of phase-resolved partial discharge (PRPD) spectra, using expert experience to determine different defect types. This approach is highly subjective and inefficient. In recent years, machine learning methods have been introduced to the field of GIS PD pattern recognition. Commonly used methods include 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 in practical applications. This is particularly true in the complex environment of offshore wind power plants, where equipment is affected by factors such as temperature, humidity, and salt spray, resulting in significant variations in discharge characteristics, further complicating identification. Existing deep learning models mostly ignore the physical mechanisms of PD and rely solely on data-driven feature learning. This lacks understanding and constraints of the physical processes involved in PD. This results in significantly reduced recognition accuracy and insufficient generalization when faced with unseen discharge patterns or discharge signals in complex environments. This technical problem seriously restricts the practicality and reliability of the intelligent monitoring system of offshore wind power GIS equipment. Summary of the Invention

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

[0004] The present invention is implemented as follows: The present invention provides a method for identifying partial discharge patterns 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 collecting PRPD spectra; pre-processing the collected PRPD spectra data; constructing an improved SAE network model, integrating the ECA attention mechanism between the encoder output and the decoder input, and using pre-trained The PDCN model assists in network feature extraction; a training data set is established based on the preprocessed PRPD spectrum data to train the improved SAE network model; the processed PRPD spectrum data of the GIS equipment to be tested is input into the trained improved SAE network model to identify the defect type and obtain the recognition result.

[0005] The steps of constructing the GIS partial discharge experimental platform are specifically constructed in accordance with the IEC60270 standard. The experimental circuit consists of an AC power supply, a transformer, a protection resistor, a coupling capacitor, a detection impedance, a defect model and a universal partial discharge measurement and analysis system. The protection resistor is set to 5kΩ, the coupling capacitor is 800pF, and the GIS cavity is filled with an air pressure of 0.1MPa. The general partial discharge measurement and analysis system adopts MPD 800, and the center frequency of MPD 800 is set to 250kHz.

[0006] Among them, the four defect models specifically refer to tip discharge, surface discharge, air gap discharge, and suspended metal discharge. The tip discharge test voltage is 4.9kV, the surface discharge test voltage is 3.5kV, the air gap discharge test voltage is 6.9kV, and the suspended metal discharge test voltage is 7.5kV. 500 PRPD spectra are collected for each defect type at a time interval of 2 minutes.

[0007] Among them, the step of preprocessing the collected PRPD spectrum data specifically adopts the color threshold segmentation method of the HSV color space to remove the coordinate axis and grid lines in the original image, and then all images are uniformly adjusted to a size of 256×256 and normalized.

[0008] Among them, the color threshold segmentation method of the HSV color space is to convert the image from the BGR color space to the HSV color space, define the HSV range suitable for the gray and black grids, extract the pixels in the image that meet the range through color threshold segmentation, and replace the grid lines and coordinate axes with the background color.

[0009] Among them, in the pre-processing step, the PRPD spectrum characteristics are also enhanced based on the partial discharge physical mechanism equation, which is used to describe the partial discharge physical mechanism equation. The physical process of partial discharge in gaseous media and its relationship with the surrounding electric field distribution.

[0010] The inputs of the partial discharge physical mechanism equation include electric field intensity, gas pressure, temperature, air gap distance and dielectric relative permittivity, and the outputs are discharge starting voltage and discharge amplitude. The discharge amplitude is used for discharge amplitude calibration during PRPD spectrum feature enhancement.

[0011] Among them, the encoder part of the improved SAE network model contains three convolutional layers that use 64, 128, and 256 filters for feature extraction respectively, and the decoder part restores the image space dimension through the upsampling layer.

[0012] Among them, 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, the description of each channel is obtained through global average pooling, and then one-dimensional convolution is used to interact information between channels.

[0013] Among them, the The PDCN model is a A pre-trained deep convolutional neural network model for partial discharge recognition in gas environment, whose structure is a deep convolutional neural network based on multi-scale residual blocks.

[0014] It states The PDCN model includes an initial convolution layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. The multi-scale residual block contains three parallel convolution branches with convolution kernel sizes of 3×3, 5×5, and 7×7, respectively. The steps for establishing the training dataset of the PDCN model include collecting 10,000 different types of discharge spectra from the public GIS partial discharge database, extracting the physical feature vectors for each spectrum, and collecting the data under different working conditions. Gas insulation characteristic parameters are obtained to obtain the training data set.

[0015] The step of dividing the training set and the test set specifically involves dividing the 2,000 PRPD spectrogram data into a training set and a test set, with the training set accounting for 70% (1,400 images) and the test set accounting for 30% (600 images), respectively, and using the training set to train the improved SAE network model. The step of inputting the test set data into the trained improved SAE network model for defect type identification specifically involves processing features through global pooling dimensionality reduction and a fully connected layer, and ultimately using a Softmax classifier to output identification results for four defect types. The training process incorporates partial discharge physical parameters as constraints, and the partial discharge physical parameter constraints include discharge starting voltage constraints, discharge phase distribution constraints, and discharge amplitude distribution constraints.

[0016] The present invention obtains high-quality data by constructing an experimental platform, enhances the PRPD spectrum features by combining the partial discharge physical mechanism equation, and integrates the ECA attention mechanism and pre-trained The PDCN model achieves highly accurate recognition of four common defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. In terms of network structure design, this method combines the feature extraction capability of the autoencoder with the adaptive feature weight allocation of the attention mechanism. At the same time, by introducing the physical parameter constraints of partial discharge, it integrates physical laws into the deep learning process, effectively solving the problems of insufficient feature extraction and limited model generalization ability in traditional recognition methods. Compared with traditional methods, the present invention not only improves the recognition 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, the technical problems of low accuracy and poor generalization ability of partial discharge pattern recognition in offshore wind power GIS equipment are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart 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 Schematic diagram of the partial discharge defect model in Example 2.

[0020] Figure 4 PRPD spectra of partial discharge of different defect types in Example 2 (after pretreatment).

[0021] Figure 5 This is an actual picture of the GIS partial discharge experimental platform in Example 2. DETAILED DESCRIPTION

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

[0023] like Figure 1 FIG. 1 is a flow chart of a method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network provided by the present invention. The method includes the following steps: S01. Construct GIS partial discharge experimental platform: The experimental circuit consists of AC power supply, transformer, protection resistor, coupling capacitor, detection impedance, defect model and MPD 800. The protection resistor is set to 5kΩ, the coupling capacitor is 800pF, and the GIS cavity is filled with air at a pressure of 0.1MPa. For gas, the center frequency of MPD 800 is set to 250kHz, and the ambient noise is controlled at around 200fC; S02. Design defects: Four defect models were designed based on the GIS partial discharge experimental platform: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. The test voltages for tip discharge were 4.9 kV, surface discharge 3.5 kV, air gap discharge 6.9 kV, and suspended metal discharge 7.5 kV. 500 PRPD spectra were collected at 2-minute intervals for each defect type. S03. Preprocessing the collected PRPD spectrum data: Using the color threshold segmentation method in the HSV color space to remove the coordinate axes and grid lines in the original image, all images are then resized to 256×256 and normalized. At the same time, the PRPD spectrum features are enhanced based on the partial discharge physical mechanism equation. S04. Build an improved SAE network model: The encoder part contains three convolutional layers that use 64, 128, and 256 filters for feature extraction. The decoder part restores the image space dimension through the upsampling layer, integrates the ECA attention mechanism between the encoder output and the decoder input, and uses pre-trained PDCN model assists network feature extraction; S05. Model training: 2000 PRPD spectra data were divided into training set and test set, with the training set accounting for 70% (1400 images) and the test set accounting for 30% (600 images). The improved SAE network model was trained using the training set, and the physical parameters of partial discharge were used as constraints during the training process. The test set data was input into the trained improved SAE network model, and the features were processed through global pooling dimensionality reduction and fully connected layers. Finally, the Softmax classifier was used to output the recognition results of four defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge. The recognition results were then verified using the physical parameters of partial discharge constraints. S06. Model application: Input the processed PRPD spectrum data of the GIS equipment to be tested into the trained improved SAE network model to identify the defect type and obtain the identification result.

[0024] Among them, the GIS partial discharge experimental platform is built in accordance with the IEC60270 standard to ensure the quality of data collection.

[0025] Among them, the color threshold segmentation method of the HSV color space is to convert the image from the BGR color space to the HSV color space, define the HSV range suitable for the gray and black grids, extract the pixels in the image that meet the range through color threshold segmentation, replace the grid lines and coordinate axes with the background color, and keep the original image content in 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 the points represents the discharge frequency. Different defect types present unique discharge patterns on the PRPD spectrum.

[0027] Among them, the partial discharge physical mechanism equation is used to describe The physical process of partial discharge in gas medium and its relationship with the surrounding electric field distribution. The input includes electric field strength E, gas pressure p, temperature T, air gap distance d and dielectric relative dielectric constant ε. The electric field strength E is derived from the test voltage and the electrode spacing. The gas pressure p is derived from the gas filled in the GIS cavity. The gas pressure value, the temperature T is derived from the laboratory ambient temperature measurement value, the air gap distance d is derived from the geometric size of the defect model, and the dielectric constant ε is derived from The gas physical properties database outputs the discharge initiation voltage Ui and the discharge amplitude q. The discharge initiation voltage Ui is used to determine the discharge characteristic threshold of 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 distribution 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 interact with information 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 stacked autoencoder layers. 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, and extracts features from the input vector through nonlinear transformation. The decoder consists of a hidden layer and an output layer, and forms an output value approximate to the input vector based on the proposed feature map.

[0030] Among them, 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 Softmax output layer with four neurons to realize the recognition and classification of four defect types: tip discharge, surface discharge, air gap discharge, and suspended metal discharge.

[0031] in, The PDCN model is a Pre-trained deep convolutional neural network model for partial discharge recognition in gas environment, The specific structure of the PDCN model is a deep convolutional neural network based on multi-scale residual blocks, which includes an initial convolution layer, four multi-scale residual blocks, a feature fusion layer and a fully connected classification layer. The initial convolution layer uses 64 7×7 convolution kernels to extract features from the input PRPD spectrum. The multi-scale residual block contains three parallel convolution branches, and the convolution kernel sizes are 3×3, 5×5 and 7×7, respectively, which capture discharge features of different scales. The number of output channels of each multi-scale residual block is 64, 128, 256 and 512, respectively. The residual connection is used to ensure the stable transmission of gradient information during deep network training. The feature fusion layer uses channel weighting to integrate multi-level features. The weight coefficient is adaptively adjusted according to the three parameters of the phase distribution density, discharge amplitude concentration and phase span of the PRPD spectrum. The fully connected classification layer consists of a hidden layer of 256 nodes and an output layer of four nodes.

[0032] in, The steps of establishing the training data set of the PDCN model specifically include collecting 10,000 different types of discharge spectra from the public GIS partial discharge database, extracting the physical feature vector for each spectrum, which includes parameters such as discharge start phase, discharge end phase, discharge amplitude distribution, and discharge frequency distribution, and collecting the data under different working conditions. Gas insulation characteristic parameters, establish a mapping relationship between physical parameters and spectral features, and use data enhancement technology to expand the data set size to 30,000 images. The data enhancement technology includes operations such as rotation, translation, scaling, and adding Gaussian noise. The training data set, validation data set, and test data set are divided in a ratio of 7:2:1 to ensure that the model has sufficient generalization ability.

[0033] in, During the training process, the PDCN model introduces a partial discharge physical constraint term into the loss function. The partial discharge physical constraint term is constructed based on the discharge characteristic parameters calculated based on the partial discharge physical mechanism equation to ensure that the model prediction results conform to physical laws. At the same time, an early stopping strategy is adopted to avoid overfitting. The early stopping strategy means stopping training when the accuracy of the verification data set has not been significantly improved for five consecutive cycles. Finally, the model performance is evaluated on an independent test data set. The model performance indicators include recognition accuracy, F1 score, confusion matrix, etc., to ensure that the recognition accuracy reaches more than 95%.

[0034] Among them, the local discharge physical parameter constraints are the relationship between the theoretical parameter values ​​calculated by the local discharge physical mechanism equation and the actual measured values, which guide and constrain the neural network learning process. The local discharge physical parameter constraints include discharge starting 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 the ECA attention mechanism is integrated between the encoder and decoder of the improved SAE network, and combined with the pre-trained SF6PDCN model to assist in feature extraction, forming an efficient hybrid feature enhancement architecture. The ECA attention mechanism adaptively selects the convolution kernel size and finely weights the feature channels, effectively highlighting key channel information. 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 the importance weight of each channel through the Sigmoid function, allowing the network to focus on significant features related to the discharge pattern. At the same time, 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 achieves top-down feature enhancement and bottom-up knowledge transfer, and has three major advantages over traditional methods: first, it improves the recognition ability of small sample discharge patterns and reduces the dependence on large-scale labeled data; second, it enhances the model's robustness to noise and interference, especially in complex power system environments; third, by integrating physical constraints and deep learning, it improves the interpretability and technical reliability of the model, making the recognition results more in line with the physical laws of partial discharge.

[0036] It should be noted that partial discharge physical parameter constraints are a method that integrates theoretical physics models with deep learning. Its core is to guide and constrain the neural network learning process by leveraging the relationship between theoretical parameter values ​​calculated from the partial discharge physical mechanism equations and actual measured values. This physical constraint mechanism is primarily implemented through three aspects: discharge onset voltage constraints, discharge phase distribution constraints, and discharge amplitude distribution constraints. When the recognition results output by the neural network deviate from physical laws, the physical parameter constraints impose additional penalty terms to adjust the network parameters, ensuring that the model learning process adheres to the physical laws of partial discharge in the SF6 gas medium. Compared with 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 that have never appeared in the training data; second, it improves the model's robustness in data-sparse or noise-interference environments, reducing the risk of overfitting; third, it gives the neural network interpretability through physical mechanisms, making the black box model decision-making process more transparent and credible; finally, physical parameter constraints also reduce dependence on large-scale labeled data, achieving efficient complementarity between physical knowledge and data learning, so that the model can not only accurately identify discharge types, but also predict discharge behavior under different working conditions, providing a more reliable technical foundation for equipment status assessment and predictive maintenance.

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

[0038] The specific implementation method of step S01 is to construct a GIS partial discharge experimental platform that complies with the IEC60270 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, and a step-up transformer with a transformation ratio of 220V:50kV is connected to provide sufficient experimental voltage. The protective resistor adopts a high-power resistor with a resistance of 5kΩ, which is mainly used to limit the amplitude of the discharge current and prevent equipment damage. An 800pF high-voltage ceramic capacitor is selected as a coupling capacitor, which provides a transmission channel for the discharge signal and is used for measurement system calibration. The detection impedance selects a wide-band impedance with a frequency response range of 10kHz to 10MHz, which is used to convert the high-frequency discharge pulse signal into a measurable voltage signal. The GIS cavity is filled with a pressure of 0.1MPa. Gas, the pressure value is selected based on Gas at this pressure exhibits excellent insulation properties and is close to actual equipment operating conditions. The MPD 800 partial discharge detection system's center frequency is set at 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 kept to approximately 200fC through shielding, signal filtering, and grounding optimization, ensuring that the signal-to-noise ratio of the measured signal met experimental requirements. This step aims 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 utilizes a copper needle tip against a flat plate electrode structure, with a 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 at 4.9 kV. The surface discharge model is constructed using epoxy resin insulation material, with a 15 mm long triangular surface discharge channel formed on its surface. The test voltage is set at 3.5 kV. The air gap discharge model consists of two parallel metal electrodes with a 0.5 mm thick insulating spacer inserted between them to form an air gap. The test voltage is set at 6.9 kV. The suspended metal discharge model suspends a 2 mm diameter metal sphere near a high-voltage electrode at a distance of 4 mm from the electrode. The test voltage is set at 7.5 kV. The test voltage for each defect type is determined by 1.2 times the discharge inception voltage calculated based on the partial discharge physical mechanism equation to ensure stable discharge. During data acquisition, 500 PRPD spectra are continuously collected at 2-minute intervals for each defect type. The acquisition interval is selected based on the results of the partial discharge stability analysis to ensure the temporal uniformity and representativeness of the collected 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 method of step S03 is to preprocess the collected PRPD spectrum data. First, the image is converted from the BGR color space to the HSV color space, and the HSV threshold range of the gray and black grids is defined as: H value 0 to 180, S value 0 to 30, and V value 100 to 220. Based on this threshold range, the coordinate axis and grid line area in the image are extracted using a binarization method to generate a mask image. The pixels in the identified grid line area are then replaced with the background color while keeping the discharge signal area unchanged. The processed image is then uniformly adjusted to a size of 256×256 pixels and image scaling is performed using a bicubic interpolation algorithm, which can achieve smooth scaling while maintaining image details. Normalization processing uses the Min-Max normalization method to map pixel values ​​to the range of 0 to 1. Based on the partial discharge physical mechanism equation, the PRPD spectrum feature is enhanced. The specific method is to calculate the adjustment coefficient α using the discharge initiation voltage Ui and the discharge amplitude q. The adjustment coefficient usually ranges from 1.05 to 1.3. The image is then subjected to nonlinear contrast enhancement processing based on the α value to enhance the visual significance of the discharge concentrated area. The purpose of this step is to remove irrelevant information from the image, retain key discharge features, and normalize the data into a format that is easy for the neural network to process.

[0041] 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 a padding of 1; the second layer uses 128 3×3 filters with a stride of 2 and a padding of 1; and the third layer uses 256 3×3 filters with a stride of 2 and a padding of 1. Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The momentum parameter of the batch normalization layer is set to 0.9. The decoder part uses three transposed convolutional layers to achieve upsampling. The first layer uses 128 3×3 filters with a stride of 2 and a padding of 1; the second layer uses 64 3×3 filters with a stride of 2 and a padding of 1; and the third layer uses 3 3×3 filters with a stride of 1 and a padding of 1. The ECA attention mechanism is integrated between the encoder output and the decoder input. First, the channel descriptor is obtained by global average pooling, and then the one-dimensional convolution with a convolution kernel size of k is used to realize the information interaction between channels, where the k value is adaptively determined by the number of channels C: k = , usually for 256 channels, the k value is 9. Finally, the weight coefficient is generated by the Sigmoid activation function to weight the feature channels. The network also uses pre-trained PDCN model assists feature extraction, The output features of the second and third multi-scale residual blocks of the PDCN model are fused with the corresponding encoder layer features of the improved SAE network, with a fusion weight ratio of 0.3:0.7. The purpose of this step is to build a deep learning model with efficient feature extraction and reconstruction capabilities, combining the attention mechanism and pre-trained model to enhance the recognition of partial discharge characteristics.

[0042] The specific implementation of step S05 is to perform model training. First, the 2000 PRPD spectrogram dataset is randomly divided into a training set of 1400 and a test set of 600 at a ratio of 7:3. The training set is then divided into 1260 actual training samples and 140 validation samples at a ratio of 9:1. The model is trained using a mini-batch gradient descent method with a batch size of 32. The initial learning rate is set to 0.001. The Adam optimizer is used with parameters Set to 0.9, Set to 0.999, Set as The training process adopts the cosine annealing learning rate adjustment strategy, the minimum learning rate is set to 0.01 times the initial learning rate, and the number of training rounds is set to 100 rounds. The loss function is designed to be the weighted sum of reconstruction loss and classification loss: ,in is the mean square error loss, used to evaluate the reconstruction quality; is the cross entropy loss, used to evaluate classification accuracy; To constrain the loss of physical parameters of partial discharge, calculate the deviation between the predicted parameters and the theoretical parameters, and the weight coefficient =0.4, =0.5, =0.1. Physical parameter constraints include a discharge onset phase constraint with a threshold of ±15°; a discharge amplitude distribution constraint with a threshold of ±20%; and a discharge frequency distribution constraint with a threshold of ±15%. An early stopping strategy was used during training, stopping training when the validation set loss showed no significant decrease for five consecutive rounds. This step aims to optimize network parameters so that the model can both learn the visual features of PRPD spectra and conform to the physical laws of partial discharge. Next, the model is tested and discharge type identification is performed. First, the 600 PRPD spectra from the test set are input into the trained improved SAE network model. Feature processing involves extracting features through the encoder, with the final convolutional layer outputting a feature map of size 64×64×256. Dimensionality reduction is achieved by applying global average pooling to compress the feature map into a 1×1×256 feature vector. Features are processed through a fully connected layer consisting of 128 neurons with a Reluctant Unit (ReLU) activation function. Finally, a 4-neuron Softmax output layer is used to calculate the probability distribution of the four defect types. During the prediction process, a probability threshold of 0.75 is set; samples below this threshold are marked as uncertain and require further analysis. At the same time, the identification results are verified by the constraints of the physical parameters of partial discharge, and the consistency index CI between the predicted discharge parameters and the theoretical parameters is calculated: ,in is the predicted parameter value, is the theoretical parameter value, The CI is the tolerance for deviation. When the CI falls below 0.85, a warning is triggered and manual review is recommended. Recognition results are evaluated using metrics such as accuracy, precision, recall, and F1 score. The overall recognition accuracy must be no less than 95%, and the F1 score for each defect type must be no less than 0.92. This step verifies the model's recognition performance and ensures the reliability and accuracy of the recognition results through physical parameter constraints.

[0043] The specific implementation of step S06 is model application. Specifically, the processed PRPD spectrum data of the GIS device to be tested is input into the trained improved SAE network model to perform defect type identification to obtain an 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, consisting of 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 convolution kernels with a stride of 2 and padding of 3 to extract features from the 256×256×3 input PRPD spectrogram, resulting in an output of size 128×128×64. Each multi-scale residual block contains three parallel convolution branches, using kernels of different sizes: 3×3, 5×5, and 7×7, to capture 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 1×1 convolutional shortcut connection to align the number of input and output channels. The feature fusion layer uses channel weighting to integrate multi-level features, and the weight coefficient 、 、 The weights for 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 the fine-grained characteristics of partial discharge through multi-scale feature extraction, while ensuring stable gradient information transfer during deep network training through residual connections.

[0045] The specific implementation of the PDCN model training data set is as follows: First, 10,000 different types of discharge spectra are collected from three public GIS partial discharge databases, including the EPRI partial discharge database, the IEC partial discharge standard atlas, and the power equipment intelligent diagnosis open data platform, covering different voltage levels and different Discharge samples under gas pressure conditions. Physical feature vectors are extracted for each spectrum, including the discharge starting phase , discharge end phase , discharge amplitude distribution function , discharge frequency distribution function The discharge start and end phases are determined by the threshold method of the phase histogram, and the threshold is set to 5% of the maximum frequency. The discharge amplitude distribution is obtained by the amplitude statistics within the 95% confidence interval. The discharge frequency distribution is obtained by the number of discharges per 1° within the 360° phase interval. At the same time, the data under different working conditions are collected. Gas insulation characteristic parameters, including gas pressure in the range of 0.1 to 0.6 MPa, preferably 0.1 to 0.4 MPa, with breakdown field strength, ionization coefficient, adhesion coefficient, diffusion coefficient and other parameter values ​​for every 0.1 MPa. Establish a mapping relationship between physical parameters and spectral characteristics, and use polynomial fitting method to construct the functional relationship between gas pressure, temperature, electrode structure and discharge characteristic parameters. The fitting accuracy requirement is The data set was expanded to 30,000 images using data augmentation techniques. These augmentations included rotations of ±10°, translations of ±5%, scaling by 0.95 to 1.05 times, the addition of Gaussian noise with a standard deviation of 10% of the original image, brightness adjustments of ±10%, and contrast adjustments of ±15%. The data set was partitioned into a training set of 21,000 images, a validation set of 6,000 images, and a test set of 3,000 images in a 7:2:1 ratio, ensuring a balanced distribution of defect types within each data subset. The proportions of each defect type were as follows: 25% for tip discharge, 25% for surface discharge, 25% for air gap discharge, and 25% for suspended metal discharge.

[0046] It should be noted that the physical mechanism equation of partial discharge plays a core theoretical supporting role in the entire identification method. It reveals the The essential laws of the discharge process in gas media. This equation establishes a strict connection between macroscopically measurable external conditions such as electric field strength, gas pressure, and temperature and the microscopic electron impact ionization process. By analyzing the motion behavior and energy accumulation process of gas particles under the action of the electric field, it predicts the discharge starting voltage threshold and the evolution trend of the discharge amplitude. The important role of this mechanism equation is reflected in the introduction of prior physical constraints to the neural network, so that model training not only relies on data statistical laws, but also incorporates the basic physical laws of gas discharge, thereby maintaining high accuracy and strong generalization capabilities even in small sample conditions. In particular, when the defect model parameters deviate from the standard operating conditions, the theoretical prediction values ​​provided by the physical mechanism equation can effectively correct the neural network output and reduce the misjudgment rate under abnormal conditions.

[0047] It should be noted that as a deep autoencoding structure, the effect of the SAE network is mainly reflected in feature extraction and dimensionality reduction. Unlike traditional convolutional neural networks that directly perform classification learning, SAE first learns the intrinsic expression of the data in an unsupervised manner, compresses the high-dimensional PRPD spectrum into a low-dimensional potential feature space, and then verifies the fidelity of the extracted features through a decoding and reconstruction process. This self-supervised learning mechanism enables SAE to learn more essential feature expressions from limited samples, avoiding the overfitting trap that direct classification models are prone to fall into. The multi-layer stacking design of SAE allows the network to extract hierarchical features from low-level textures to high-level semantics layer by layer, forming a deep understanding of the intrinsic structure of the PRPD spectrum. The improved SAE network further enhances the ability to extract local spatial features by introducing convolution operations and pooling layers, which is particularly suitable for dealing with complex situations where local morphological features and global distribution patterns coexist in discharge spectra.

[0048] It should be noted that the ECA attention mechanism plays a key role in adaptively weighting channel features in this method. This mechanism first obtains the global response value of each channel through global average pooling, and then cleverly uses one-dimensional convolution operations to capture the dependencies between channels, avoiding the problem of parameter redundancy in the fully connected layer in the traditional attention mechanism. The design concept of ECA is to efficiently extract channel importance weights through a lightweight structure, so that the network can automatically focus on feature channels related to specific discharge types, while suppressing the influence of irrelevant or noisy channels. In the task of local discharge identification, the discharge characteristics of different types of defects are often concentrated in specific frequency or phase intervals. ECA attention seizes this feature and enhances the model's perception of key areas through adaptive weight distribution, thereby improving the discriminability of feature expression.

[0049] It should be noted that As a pre-trained deep convolutional network, the PDCN model provides a domain-specific feature extractor. Pre-trained on large-scale GIS discharge data, the model has learned universal feature representations of discharge spectra and can identify key texture and morphological features of different defect types. The multi-scale residual block design enables the network to simultaneously capture fine-grained local features and large-scale global patterns in discharge spectra. The residual connection effectively solves the vanishing gradient problem in deep network training, ensuring effective learning of deep features. The feature fusion layer dynamically adjusts feature importance based on the characteristics of the input spectra through adaptive channel weighting, further enhancing the model's adaptability to different types of discharge patterns. The synergy between the PDCN model and the improved SAE network achieves 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 of partial discharge serve as a bridge connecting physical mechanisms and deep learning, and their role is reflected in two aspects: guiding the training process and verifying the inference results. During the training phase, the physical parameter constraints participate in network optimization in the form of a loss function, guiding the model to learn feature expressions that conform to physical laws, avoiding physical inconsistencies that may arise from relying entirely on data-driven methods. During the inference phase, the physical parameter constraints serve as a post-processing mechanism to verify the physical rationality of the output results of the neural network and filter 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 have complex environments and changeable operating conditions. It can effectively improve the generalization ability of the model in the face of non-standard operating conditions and new defects in actual operation.

[0051] Specifically, the principle of the present invention is: The core technical principle of this invention lies in the organic integration of physical mechanism knowledge with deep learning models, and the efficient recognition of partial discharge patterns through an improved SAE network structure. Its principle is mainly reflected in the following aspects: First, the present invention establishes a mapping relationship between the physical mechanism of partial discharge and the characteristics of PRPD spectra. 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 intensity, gas pressure, temperature, air gap distance, and relative dielectric constant of the medium, and outputs discharge inception voltage and discharge amplitude parameters. These physical parameters are directly used to enhance the features of the PRPD spectra, incorporating physical mechanism information into the data preprocessing stage, providing more physically meaningful input features for subsequent deep learning models.

[0052] Secondly, the improved SAE network structure of the present invention is optimized for the task of partial discharge identification. Traditional autoencoders often suffer from information loss when processing high-dimensional image data. This invention uses an encoder consisting of three convolutional layers to extract features at different scales, and an upsampling layer in the decoder to restore the image spatial dimensions, preserving the key characteristics of the discharge pattern. Furthermore, the introduction of the ECA attention mechanism enables adaptive weight allocation of feature channels, enabling the network to focus on more discriminative features and improving the model's feature learning capabilities.

[0053] Furthermore, the present invention will pre-train The PDCN model is integrated into the improved SAE network, and the transfer learning mechanism is used to enhance the feature extraction capability. The PDCN model captures discharge features at different scales through multi-scale residual blocks, and incorporates physical constraints of partial discharge during training, making the extracted features more consistent with the physical laws of discharge. The introduction of 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] Crucially, this invention introduces physical parameter constraints for partial discharge during deep learning training, integrating constraints on the discharge onset voltage, discharge phase distribution, and discharge amplitude distribution into the loss function. This approach enables the network to learn not only the statistical characteristics of the data but also feature representations that conform to physical laws, thereby maintaining high recognition capabilities even when faced with unprecedented discharge patterns or complex environmental conditions. The introduction of physical constraints essentially expands the model's effective learning space, allowing it to focus not only on the "what" but also on the "why," thereby improving the model's generalization and interpretability.

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

[0056] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0057] The specific implementation method of step S01 is to construct a GIS partial discharge experimental platform that complies with the IEC60270 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, and a step-up transformer with a transformation ratio of 220V:50kV is connected to provide sufficient experimental voltage. The protective resistor adopts a high-power resistor with a resistance of 5kΩ, which is mainly used to limit the amplitude of the discharge current and prevent equipment damage. An 800pF high-voltage ceramic capacitor is selected as a coupling capacitor, which provides a transmission channel for the discharge signal and is used for measurement system calibration. The detection impedance selects a wide-band impedance with a frequency response range of 10kHz to 10MHz, which is used to convert the high-frequency discharge pulse signal into a measurable voltage signal. The GIS cavity is filled with a pressure of 0.1MPa. Gas, the pressure value is selected based on Gas at this pressure exhibits excellent insulation properties and is close to actual equipment operating conditions. The MPD 800 partial discharge detection system's center frequency is set at 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 kept to approximately 200fC through shielding, signal filtering, and grounding optimization, ensuring that the signal-to-noise ratio of the measured signal met experimental requirements. This step aims to establish a standardized experimental platform, providing a reliable foundation for subsequent defect model discharge data acquisition.

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

[0059] The specific implementation of step S03 is to preprocess the collected PRPD spectrum data. First, the image is converted from the BGR color space to the HSV color space, and the HSV threshold range of the gray and black grids is defined as: H value 0 to 180, S value 0 to 30, and V value 100 to 220. Based on this threshold range, the coordinate axis and grid line area in the image are extracted using a binarization method to generate a mask image. The binarization process can be expressed as: ; Where, The mask image is at coordinates The value at 、 、 The original image is at coordinates hue, saturation, and brightness values ​​at 、 、 、 、 、 The lower and upper limits of the HSV threshold range are respectively. The identified gridline area pixels are then replaced with the background color, while the discharge signal area remains unchanged. The processed image is then uniformly resized to 256×256 pixels and scaled using the bicubic interpolation algorithm, which can achieve smooth scaling while maintaining image details. Normalization uses the Min-Max normalization method to map pixel values ​​to the range of 0 to 1: ; Where, The normalized image is at coordinate The pixel value at ; The original image at coordinates The pixel value at ; and The PRPD spectrum feature enhancement is performed based on the physical mechanism equation of partial discharge, and the discharge starting voltage is used to and discharge amplitude Calculate the adjustment factor : ; Where, is the measured discharge amplitude; is the theoretical reference discharge amplitude; is the test voltage; is the discharge starting voltage. The adjustment coefficient is usually in the range of 1.05 to 1.3, and then Perform nonlinear contrast enhancement on the image: ; Where, The enhanced image is at coordinate The pixel value at ; The normalized image is at coordinate The purpose of this step is to eliminate irrelevant information in the image, retain key discharge features, and normalize the data into a format that is easy for the neural network 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 step size of 1 and a padding of 1; the second layer uses 128 3×3 filters with a step size of 2 and a padding of 1; and the third layer uses 256 3×3 filters with a step size of 2 and a padding of 1. The convolution operation of each layer can be expressed as: ; Where, For the Layer feature map; For the Layer convolution kernel weights; For the Layer feature map; For the Layer bias term; Represents the convolution operation; Represents the activation function, using the ReLU function: Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The calculation formula of the batch normalization layer is: ; ; Where, is the input feature; is the mean of the small batch data; is the variance of the mini-batch data; To prevent small constants from dividing by zero, the value is ; and is a learnable parameter; 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 layers of transposed convolutional layers for upsampling. The first layer uses 128 3×3 filters with a stride of 2 and a padding of 1; the second layer uses 64 3×3 filters with a stride of 2 and a padding of 1; and the third layer uses 3 3×3 filters with a stride of 1 and a padding of 1. The ECA attention mechanism is integrated between the encoder output and the decoder input. First, the channel descriptor is obtained through global average pooling: ; Where, For the Descriptors of channels; The feature map Channels at position The value at and are the height and width of the feature map respectively. Then use the convolution kernel size of The one-dimensional convolution realizes the information interaction between channels, where Value by channel number Adaptive determination: ; When the number of channels hour, The value is 9. The one-dimensional convolution operation is expressed as: ; Where, For the The weight coefficient of each channel; is the convolution kernel parameter; is the channel descriptor; For the Sigmoid activation function: Finally, the Sigmoid activation function is used to generate weight coefficients and weight the feature channels: ; Where, is the weighted feature map channels; is the weight coefficient; The original feature map channels. The network also uses pre-trained PDCN model assists 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: ; Where, is the fused feature; for PDCN model output features; To improve the output features of the SAE network encoder, the purpose of this step is to build a deep learning model with efficient feature extraction and reconstruction capabilities, and to enhance the recognition ability of partial discharge features by combining the attention mechanism and pre-training model.

[0061] The specific implementation of step S05 is to perform model training. First, the 2000 PRPD spectrogram dataset is randomly divided into a training set of 1400 and a test set of 600 at a ratio of 7:3. The training set is then divided into 1260 actual training samples and 140 validation samples at a ratio of 9:1. The model is trained using a mini-batch gradient descent method with a batch size of 32. The initial learning rate is set to 0.001. The Adam optimizer is used with parameters Set to 0.9, Set to 0.999, Set as The training process adopts the cosine annealing learning rate adjustment strategy, and the learning rate calculation formula is: ; Where, For the The learning rate of the training round; is the minimum learning rate, set to 0.01 times the initial learning rate; is the maximum learning rate, that is, the initial learning rate is 0.001; is the current training round number; is the total number of training rounds, set to 100. The loss function is designed to be the weighted sum of reconstruction loss and classification loss: ; Where, is the mean square error loss, which is used to evaluate the reconstruction quality: ; is the cross entropy loss, which is used to evaluate classification accuracy: ; To constrain the loss of partial discharge physical parameters, calculate the deviation between the predicted parameters and the theoretical parameters: ; Where, is the sample size; For the The true value of the samples; For the The predicted value of samples; is the number of categories, 4 in this case; For the The samples belong to The true probability of the class; For the The samples belong to The predicted probability of the class; is the number of physical parameters; For the The weight coefficient of each physical parameter; For the The predicted value of a physical parameter; For the Theoretical values ​​of physical parameters; For the The permissible deviation of a physical parameter. Weighting factor , , Physical parameter constraints include a discharge starting phase constraint with a threshold of ±15°; a discharge amplitude distribution constraint with a threshold of ±20%; and a discharge frequency distribution constraint with a threshold of ±15%. An early stopping strategy is used during training, stopping training 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 both learn the visual features of the PRPD spectra and conform to the physical laws of partial discharge.

[0062] Next, we conduct model testing and discharge type identification. First, we input the 600 PRPD spectra from the test set into the trained improved SAE network model. The feature processing process includes: extracting features through the encoder, with the final convolutional layer outputting a feature map of size 64×64×256; applying global average pooling to reduce dimensionality, compressing the feature map into a 1×1×256 feature vector: ; Where, For the Global features of each channel; The feature map Channels at position The value at and are the height and width of the feature map respectively. The features are processed by a fully connected layer containing 128 neurons, and the activation function is ReLU: ; Where, is the hidden layer output; is the weight matrix; is the input feature vector; is the bias vector; Represents the ReLU activation function. Finally, a Softmax output layer with 4 neurons is used to calculate the probability distribution of the four defect types: ; Where, The sample belongs to class probability; The output layer The input of each neuron; The output layer The input of each neuron. During the prediction process, the probability threshold is set to 0.75. Samples below this threshold are marked as uncertain types and require further analysis. At the same time, the recognition results are verified by the constraints of the physical parameters of partial discharge, and the consistency index CI between the predicted discharge parameters and the theoretical parameters is calculated: ; Where, is the consistency index; For the The predicted value of a physical parameter; For the Theoretical values ​​of physical parameters; For the The permissible deviation of a physical parameter; is the number of physical parameters. When the CI is lower than 0.85, a warning is triggered and manual review is recommended. The recognition result evaluation uses indicators such as accuracy, precision, recall, and F1 score. The accuracy calculation formula is: ; The accuracy calculation formula is: ; The recall calculation formula is: ; The F1 score calculation formula is: ; Where, is the number of true positive cases; is the number of true negative examples; is the number of false positives; is the number of false negatives. The overall recognition accuracy must be no less than 95%, and the F1 score for each defect type must be no less than 0.92. This step verifies the model's recognition performance and ensures the reliability and accuracy of the recognition results through physical parameter constraints.

[0063] The detailed structure of the PDCN model is as follows: The model is based on a deep convolutional neural network architecture with multi-scale residual blocks, which includes an initial convolution layer, four multi-scale residual blocks, a feature fusion layer, and a fully connected classification layer. The initial convolution layer uses 64 7×7 convolution kernels with a stride of 2 and a padding of 3 to extract features from the input 256×256×3 PRPD spectrogram, and the output size is 128×128×64. Each multi-scale residual block contains three parallel convolution branches, using convolution kernels of different sizes: 3×3, 5×5, and 7×7, to capture discharge features of different scales, respectively. The 5×5 and 7×7 convolutions are implemented by stacking multiple 3×3 convolutions to reduce the number of parameters. The calculation of the residual block can be expressed as: ; Where, is the residual block output; Input for the residual block; is the residual function, is the convolution layer parameter. The feature extraction process of the multi-scale residual block can be expressed as: ; Where, This is the result of multi-scale feature extraction; 、 、 The features extracted using 3×3, 5×5 and 7×7 convolution kernels respectively; Represents the feature concatenation operation in the channel dimension. The first multi-scale residual block outputs 64 channels, the second outputs 128 channels, the third outputs 256 channels, and the fourth outputs 512 channels. Each residual block also contains a 1×1 convolution shortcut connection to adjust the number of input and output channels to be consistent. The feature fusion layer uses channel weighting to integrate multi-level features, and the weight coefficient 、 、 The weights corresponding to shallow, middle and deep features are initially set to 0.2, 0.3 and 0.5, respectively, and are adaptively adjusted during the network training process: ; Where, is the fused feature; 、 、 These are shallow, middle, and deep features, respectively; 、 、 is the weight coefficient and satisfies The weight coefficient is based on the phase distribution density of the PRPD spectrum. , discharge amplitude concentration , Phase span Three parameters are adaptively adjusted: ; Where, is the weight matrix; is the bias vector; is the normalization function. 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. This model has approximately 8.5 million parameters. Its design philosophy is to capture the fine-grained characteristics of partial discharge through multi-scale feature extraction, while using residual connections to ensure stable gradient information transmission during deep network training.

[0064] The specific implementation of step S06 is model application, that is, inputting the processed PRPD spectrum data of the GIS equipment to be tested into the trained improved SAE network model to identify the defect type and obtain the identification result. To better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A research team applied the method of the present invention to a 35kV GIS device at 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. The schematic diagram of the experiment is shown in the figure below. Figure 2 As shown in the figure, the actual experimental deployment is as follows Figure 5 As shown in the figure, an AC power supply with a rated capacity of 100kVA is used, and the experimental voltage is provided by a step-up transformer with a transformation ratio of 220V:50kV. The platform is 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 cavity is filled with a pressure of 0.1MPa. 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 tip discharge model uses a copper needle with a curvature radius of 0.05mm and a flat plate structure, with a distance of 3mm between the needle tip and the plate, and a test voltage of 4.9kV; the surface discharge model is made of epoxy resin, with a 15mm long triangular surface discharge channel, and a test voltage of 3.5kV; the air gap discharge model consists of two parallel metal electrodes with a 0.5mm thick insulating spacer inserted between them, and a test voltage of 6.9kV; the suspended metal discharge model suspends a metal ball with a diameter of 2mm near the high-voltage electrode, at a distance of 4mm from the electrode, and a test voltage of 7.5kV. The data sample statistics for each defect type collected by the researchers are shown in Table 1: Table 1 Data collection statistics of partial discharge defect types

[0066] The collected PRPD spectra were preprocessed by color threshold segmentation in the HSV color space. The defined HSV threshold range was: H value 0 to 180, S value 0 to 30, and V value 100 to 220. The processed images were uniformly adjusted to 256 × 256 pixel size and normalized, as shown in the following example. Figure 4 Adjustment coefficient calculated based on the partial discharge physical mechanism equation 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 constructed improved SAE network model are shown in Table 3: Table 3 Structural parameters of the improved SAE network model

[0068] The 2000 PRPD spectrogram data were divided into a training set of 1400 and a test set of 600 in a ratio of 7:3. The training process used a small batch gradient descent method with a batch size of 32, an initial learning rate of 0.001, and a loss function weight coefficient of , , The changes in loss values ​​at each stage during training are shown in Table 4: Table 4 Changes in loss values ​​during model training

[0069] The evaluation indicators of the recognition results of the 600 PRPD spectrograms in the test set are shown in Table 5: Table 5 Evaluation results of recognition performance of different discharge types

[0070] Traditional methods for identifying partial discharge in GIS equipment rely primarily on manually extracting characteristic parameters from phase maps, such as the ratio of the positive and negative half-cycle discharge amplitudes, the maximum amplitude of the discharge pulse, and the distribution of the discharge phase. These parameters are then classified and identified using machine learning methods such as support vector machines, random forests, and BP neural networks. This approach suffers from the following issues: First, the feature extraction process is highly subjective, making it difficult to fully characterize the full range of discharge patterns; second, traditional machine learning methods have limited ability to fit complex nonlinear relationships; and third, they lack consideration of the physical mechanisms of discharge, resulting in weak generalization capabilities.

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

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

[0073] Table 6 Variable Explanation Table (Part I)

[0074] Table 7 Variable Explanation Table (Part II)

[0075] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying partial discharge patterns 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; Preprocess the collected PRPD spectrogram data; build an improved SAE network model, integrate the ECA attention mechanism between the encoder output and the decoder input, and use the pre-trained PDCN model assists network feature extraction; A training data set is established based on the preprocessed PRPD spectrum data to train the improved SAE network model; the processed PRPD spectrum data of the GIS equipment to be tested is input into the trained improved SAE network model to identify the defect type and obtain the recognition result.

2. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 1, characterized in that: The steps of constructing the GIS partial discharge experimental platform are specifically constructed in accordance with the IEC60270 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 universal partial discharge measurement and analysis system. The protective resistor is set to 5kΩ, the coupling capacitor is 800pF, and the GIS cavity is filled with an air pressure of 0.1MPa. Gas, the center frequency of the universal partial discharge measurement and analysis system is set to 250 kHz.

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

4. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 3 is characterized in that: The step of preprocessing the collected PRPD spectrum data specifically involves using the color threshold segmentation method of the HSV color space to remove the coordinate axes and grid lines in the original image, and then uniformly adjusting all images to a size of 256×256 and normalizing them.

5. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on the improved SAE network according to claim 4 is characterized in that: The color threshold segmentation method of the HSV color space is to convert the image from the color space to the HSV color space, define the HSV range suitable for the gray and black grids, extract the pixels in the image that meet the range through color threshold segmentation, and replace the grid lines and coordinate axes with the background color.

6. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 4, characterized in that: In the pre-processing step, the PRPD spectrum features are also enhanced based on the partial discharge physical mechanism equation, which is used to describe the partial discharge physical mechanism equation. The physical process of partial discharge in gaseous media and its relationship with the surrounding electric field distribution.

7. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 6, characterized in that: The inputs of the partial discharge physical mechanism equation include electric field intensity, gas pressure, temperature, air gap distance and dielectric relative permittivity, and the outputs are discharge starting voltage and discharge amplitude. The discharge amplitude is used for discharge amplitude calibration during the PRPD spectrum feature enhancement process.

8. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 7, characterized in that: The encoder part of the improved SAE network model contains three convolutional layers that use 64, 128, and 256 filters for feature extraction, and the decoder part restores the image space dimension through an upsampling layer.

9. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on an improved SAE network according to claim 8, characterized in that: 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. It first obtains the description of each channel through global average pooling, and then uses one-dimensional convolution to interact information between channels.

10. The method for identifying partial discharge patterns of offshore wind power GIS equipment based on improved SAE network according to claim 9, characterized in that: described The PDCN model is a A pre-trained deep convolutional neural network model for partial discharge recognition in gas environment, whose structure is a deep convolutional neural network based on multi-scale residual blocks.

Images