Multi-type GIS partial discharge diagnosis method and system based on domain adaptation

Abstract

The invention discloses a multi-type GIS partial discharge diagnosis method and system based on domain adaptation, and belongs to the technical field of electrical equipment insulation detection. In order to solve the problem that precision of an existing SF6 gas diagnosis model is reduced due to feature offset in novel environment-friendly gas, a mixed neural network model fusing CNN and Transformer is constructed, and cross-gas medium high-precision diagnosis is achieved through an unsupervised domain adaptation technology. The method comprises the following steps: building a real GIS experiment platform to collect a multi-mode partial discharge signal in mixed gas; designing a deep learning model with domain separation and domain adversarial mechanisms, and extracting domain invariant features through feature decoupling and gradient inversion; and carrying out model training by adopting an AdamW optimizer. Under the condition that a target domain sample is only 0.1% of a source domain, the diagnosis precision of the four typical defects reaches 99.8% or above, and the method is remarkably superior to a traditional method. The system can be compatible with UHF and ultrasonic signal input, and an effective technical means is provided for online monitoring of environment-friendly GIS equipment.

Description

Technical Field

[0001] The present invention belongs to the field of insulation defect detection of gas insulated switchgear (GIS), and specifically relates to a multi-type GIS partial discharge diagnosis method and system based on domain adaptation, which is suitable for online monitoring and fault early warning of partial discharge in GIS equipment. Background Art

[0002] With the expansion of power grids and the continuous advancement of grid technology, gas-insulated switchgear (GIS) has gained widespread application in power systems due to its high reliability, compact design, and low maintenance requirements. Traditional GIS primarily uses sulfur hexafluoride (SF6) as the gas insulation medium. However, SF6 is a strong greenhouse gas with serious negative impacts on global warming and ozone layer depletion. Under the Kyoto Protocol, the use of SF6 is strictly restricted. To meet environmental protection requirements, perfluoroisobutyronitrile (C4F7N) has become an ideal alternative to SF6 due to its excellent insulation properties and low global warming potential (GWP). In practical applications, C4F7N is often mixed with gases such as CO2 and N2 to optimize its insulation properties and lower its liquefaction temperature. Existing research has shown that C4F7N / CO2 mixtures exhibit excellent electrical insulation and arc extinguishing performance, and are now being used in current equipment.

[0003] While significant progress has been made in research on the physical, chemical, and electrical insulation properties of C₄FₐN / CO₂ gas mixtures, relatively little research has been conducted on partial discharge (PD), a common fault type in GIS applications. In particular, technologies based on ultrasonic or ultra-high frequency (UHF) electromagnetic wave detection remain largely unresolved. GIS PD fault diagnosis is typically performed by detecting the ultrasonic or UHF electromagnetic waves generated by PD. However, differences in the insulating gas medium result in differences in the characteristics of ultrasonic or UHF electromagnetic wave signals, rendering existing detection methods for SF₆ inapplicable.

[0004] Existing deep learning-based PD diagnosis methods (such as ResNet and LSTM) perform well with SF6 gas, but their accuracy is less than 50% to 70% when directly transferred to C4F7N mixed gas. Furthermore, they rely on a large amount of labeled data in the target domain, limiting their practical application. Furthermore, while fusion diagnosis of multimodal signals (UHF + US) can improve robustness, the characteristic differences across gas domains further complicate model adaptation.

[0005] Therefore, in order to improve the accuracy and practicality of partial discharge diagnosis in GIS, the present invention proposes a partial discharge diagnosis method based on CNN and Transformer. The present invention first constructs a partial discharge experimental platform based on a real GIS, and based on this platform, collects partial discharge signals of four typical defects in C4F7N / CO2 / O2 mixed gas, and constructs a complete partial discharge identification dataset. Then, the present invention constructs a multi-type GIS partial discharge diagnosis algorithm based on CNN and Transformer, and trains the model on the partial discharge identification dataset. Finally, the present invention conducts experimental tests on the trained partial discharge diagnosis algorithm. The results show that the model proposed by the present invention can accurately diagnose the type of partial discharge with only a small number of parameters, which is significantly better than existing methods. In summary, the multi-type GIS partial discharge diagnosis algorithm based on CNN and Transformer has high accuracy and real-time performance in partial discharge diagnosis applications, and can be compatible with ultra-high frequency electromagnetic wave signal and ultrasonic signal input at the same time. It is an effective diagnostic method with broad application prospects. Summary of the Invention

[0006] The purpose of the invention is to overcome the shortcomings of the existing technology and provide a multi-type GIS partial discharge diagnosis method and system based on domain adaptation. By fusing CNN and Transformer models and combining domain separation and domain adversarial mechanisms, cross-domain high-precision diagnosis can be achieved without the need for target domain labeling, solving the key technical problem of PD signal feature migration in environmental gas GIS.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A multi-type GIS partial discharge diagnosis method based on domain adaptation is characterized by comprising the following steps:

[0009] Build a GIS partial discharge experimental platform to collect partial discharge signals, including electromagnetic wave signals and ultrasonic signals, in a C4F7N / CO2 / O2 mixed gas environment;

[0010] Perform fast Fourier transform and normalization processing on the collected partial discharge signals to construct a partial discharge dataset containing pinpoint defects, suspension defects, particle defects, and surface discharge defects;

[0011] Constructing a partial discharge classification model based on CNN and Transformer to extract features and classify the partial discharge signal;

[0012] An unsupervised domain adaptation mechanism is introduced to eliminate the feature distribution differences between SF6 gas and C4F7N / CO2 / O2 mixed gas through the domain separation module and the domain adversarial module;

[0013] The trained model is used to diagnose the type of partial discharge in the target gas environment.

[0014] Furthermore, the GIS partial discharge experimental platform includes:

[0015] The GIS equipment is filled with a C4F7N / CO2 / O2 mixed gas at a pressure of 0.5 MPa, where the volume ratio of the mixed gas is C4F7N:CO2:O2=8.5:86:5.5;

[0016] UHF sensor and ultrasonic sensor are used to collect electromagnetic wave signals and ultrasonic signals of partial discharge respectively;

[0017] High-voltage power supply, PD detector and oscilloscope are used to stimulate partial discharge and record the signal.

[0018] Furthermore, the partial discharge signal includes four typical defect types:

[0019] For needle tip defects, a metal tip with a bottom diameter of 8 mm and a length of 20 mm is used, and the tip is 5 mm away from the low-voltage conductor;

[0020] Suspension defects are fixed between the plates using metal bolts through epoxy resin, 8 mm from the high-voltage conductor and 5 mm from the low-voltage conductor;

[0021] Particle defects were detected by placing a 0.5 mm diameter steel ball in an epoxy resin cylinder;

[0022] The surface discharge defect was formed by contacting a metal bolt with an epoxy resin cylinder with a diameter of 50 mm and a thickness of 20 mm.

[0023] Furthermore, the CNN and Transformer partial discharge classification model includes:

[0024] The encoder part consists of a one-dimensional convolutional layer with a large convolution kernel, four ResCNN Block modules, and three TransformerBlock modules, which are used to extract multi-scale features of the signal;

[0025] Classifier part: consists of two fully connected layers, which are used to output the probability distribution of defect types.

[0026] Furthermore, the unsupervised domain adaptation mechanism is introduced, specifically including: - Domain separation module: feature decoupling is achieved by minimizing the Pearson correlation coefficient between shared features and proprietary features; - Domain adversarial module: domain-invariant features are extracted through a gradient reversal layer and a domain classifier.

[0027] Furthermore, the loss function of the domain separation module is:

[0028]

[0029] in represents the Pearson correlation coefficient, represents the shared features of the source domain, represents the source domain-specific features, represents the shared features of the target domain, Represents target domain-specific features.

[0030] Furthermore, the loss function of the domain adversarial module is:

[0031]

[0032] Where, and denote the domain classification prediction of the exclusive and shared features of the source domain, and is the corresponding prediction result of the target domain. is the binary cross entropy loss function.

[0033] Second, the present invention also provides a GIS partial discharge diagnostic system for implementing the above method, which is characterized by comprising:

[0034] Signal acquisition unit: used to obtain partial discharge ultrasonic signals and UHF electromagnetic wave signals in GIS equipment;

[0035] Data processing unit: used for signal preprocessing and feature extraction;

[0036] Intelligent diagnosis unit: deploys a CNN-Transformer hybrid model trained with domain adaptation;

[0037] Result display unit: outputs partial discharge type and confidence level.

[0038] Furthermore, the intelligent diagnosis unit supports three working modes: pure UHF signal analysis mode; pure ultrasonic signal analysis mode; multimodal signal fusion analysis mode.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. Improved diagnostic accuracy and robustness: This method effectively enhances the recognition of complex partial discharge signals by combining the local feature extraction capabilities of CNNs with the global modeling capabilities of Transformers. Without requiring target domain annotation, it achieves high-precision diagnosis across gas domains through an unsupervised domain adaptation strategy. Experiments demonstrate significantly better accuracy than traditional deep learning models, improving diagnostic reliability and practicality.

[0041] 2. Compatibility with multimodal signals and improved generalization capability: This invention supports the fusion input of multimodal partial discharge signals such as ultra-high frequency (UHF) and ultrasonic (US), and uses multi-source information to enhance feature expression capabilities. It can more comprehensively characterize the characteristics of different types of discharges and significantly improve the model's adaptability and robustness to multiple types of defects under complex working conditions.

[0042] 3. No need for a large amount of target domain labeled data, low deployment cost: The unsupervised domain adaptation method that combines domain separation and domain adversarial mechanism significantly reduces the need for manual labeling of new environmental gas data such as C4F7N, lowers the model deployment threshold and data acquisition costs, and has stronger engineering scalability and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flowchart of the multi-type GIS partial discharge diagnosis method based on domain adaptation;

[0044] Figure 2 Partial discharge experimental platform;

[0045] Figure 3 Schematic diagrams of four typical defects, where (a) is a tip defect, (b) is a suspended defect, (c) is a particle defect, and (d) is a surface defect.

[0046] Figure 4 The structure of the PD signal classification model;

[0047] Figure 5 T-SNE visualization, where (a) is UHF signal and (b) is ultrasonic signal;

[0048] Figure 6 Cross-domain diagnostic models;

[0049] Figure 7 The impact of the source domain / target domain data ratio on model performance. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments, but this should not limit the scope of protection of the present invention.

[0051] like Figure 1 As shown in FIG, a multi-type GIS partial discharge diagnosis method based on domain adaptation includes the following steps:

[0052] Step 1. GIS partial discharge experimental platform construction and data collection

[0053] like Figure 2As shown, the present invention constructs a partial discharge (PD) experimental platform based on a 252kV GIS to collect PD signals from typical defects in a C4F7N / CO2 / O2 gas mixture. The high-voltage power supply is a corona-free AC power supply. This embodiment uses a UHF electromagnetic wave sensor placed on the inside of the handhole cover and an ultrasonic sensor attached to the outer wall of the GIS to capture the UHF electromagnetic wave and ultrasonic signals of PD. A Haefely DDX 9121b PD detector is used to detect the initial voltage of the PD. Finally, the captured PD signal is recorded using a LeCroy oscilloscope. To align with actual application scenarios, the pressure of the C4F7N / CO2 gas mixture filled in the GIS is 0.5MPa, and the proportions of C4F7N / CO2 / O2 in the C4F7N / CO2 / O2 mixture are 8.5% / 86% / 5.5%.

[0054] Step 2. Signal acquisition and preprocessing

[0055] The following are the settings of four typical defect models. The schematic diagram of each defect is as follows: Figure 3 As shown. The needle tip defect model uses a metal tip with a bottom diameter of 8mm and a length of 20mm to simulate the tip defect inside the GIS. The metal tip is 5mm away from the low-voltage conductor. For suspended defects, the present invention uses epoxy resin to fix the metal bolts between the upper and lower plates, wherein the metal screws are 8mm away from the high-voltage conductor and 5mm away from the low-voltage conductor. For particle defects, small steel balls with a diameter of about 0.5mm are used to simulate free metal particles. These small steel balls are placed on the lower plate and surrounded by a cylinder made of epoxy resin to prevent them from rolling off. The surface discharge defect is achieved by contacting the metal bolts leading out of the upper plate with the upper and lower bottom surfaces of the epoxy resin cylinder with a diameter of 50mm and a thickness of 20mm on the lower plate, respectively.

[0056] During the experiments, to ensure stable partial discharge (PD), the applied voltage for each defect was increased by 10 kV. For each defect type, the partial discharge (PD) signal was collected in a C₄FₐN / CO₂ gas mixture. An oscilloscope was used to simultaneously record the signals from both the ultrasonic and UHF sensors. 10,000 signal sets were collected under each experimental condition. The ultrasonic signal was sampled at a rate of 500 kHz for 20 milliseconds, while the UHF signal was sampled at a rate of 5 GHz for 2 microseconds. The obtained time-domain signals were then subjected to a fast Fourier transform (FFT) and normalized. Finally, a PD recognition dataset was constructed using the processed signals. The dataset was randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0057] Step 3. Model building and training

[0058] In order to realize the automatic recognition of PD signals, a PD signal classification model is built using the basic one-dimensional convolutional layer and the Transformer module. The input of the model is the spectrum of the PD signal, and the output is the classification probability. The structure of the model is as follows: Figure 4 shown.

[0059] The PD signal classification model can be divided into two parts: an encoder and a classifier. The encoder is used to encode the input spectrum to generate high-dimensional features, while the classifier is responsible for regressing the signal category from the high-dimensional features. In the encoder part, a one-dimensional convolution layer with a large convolution kernel is first used to capture the preliminary features of the input spectrum. Here, the present invention selects a larger convolution kernel to improve the receptive field, which is very important for capturing subtle features in the spectrum. Then, four consecutive ResCNNBlock modules are used to increase the dimension of the features and shorten the feature length at the same time. In the last part of the encoder, three Transformer Blocks with global attention are used to capture long-distance dependencies within the data to further enhance the representation ability of the features. After the features are extracted by the encoder, the classifier part of the network performs the final category prediction through two linear layers.

[0060] This embodiment addresses the domain shift problem, which hinders traditional algorithms from effectively addressing the characteristic differences between SF6 gas and the environmentally friendly C4F7N / CO2 / O2 gas mixture. A domain adaptation module is designed to achieve high-precision diagnosis across gas domains. To further explore the differences in PD signal characteristics between the two gases, the present invention introduces the t-distributed Stochastic Neighbor Embedding (T-SNE) nonlinear dimensionality reduction algorithm. This algorithm visualizes the high-dimensional intermediate features output by the encoder in two-dimensional space to observe the characteristic distribution differences of the same defect type in different gases.

[0061] The specific implementation method is as follows: First, the PD classification model trained based on SF6 gas is used to encode the PD signals of SF6 and C4F7N / CO2 / O2 respectively to obtain the high-dimensional feature representations corresponding to the two gases. Then, the T-SNE algorithm is used to reduce the above high-dimensional features to two dimensions for visualization analysis. Through T-SNE visualization (such as Figure 5 As shown in the figure, the significant differences in the high-dimensional feature space between the PD signals of the same type of defects in SF6 gas and C4F7N / CO2 / O2 mixed gas are intuitively presented, thereby verifying why the traditional model trained on SF6 gas cannot be directly applied to the diagnosis task of C4F7N / CO2 / O2 mixed gas, and providing a theoretical basis for the subsequent construction of domain adaptation models.

[0062] Through the above-mentioned visualization analysis, the present invention further designs a domain separation module and a domain adversarial module to actively extract and eliminate the differences between the PD signals of the two gases at the feature level, ensuring that the constructed PD diagnosis model can maintain high recognition accuracy in both gases.

[0063] Step 4. Introducing unsupervised domain adaptation mechanism

[0064] To address the inconsistent feature distributions (i.e., domain shift) caused by different gas media, an unsupervised domain adaptation mechanism is introduced based on the original PD classification model. This results in a proposed partial discharge diagnosis algorithm that integrates domain separation and domain adversarial modules. This method effectively generalizes the SF6 gas training model to C4F7N / CO2 / O2 mixed gas conditions. Furthermore, cross-domain transfer can be achieved without requiring labeled target domain data, requiring only a small number of unlabeled samples.

[0065] like Figure 6 As shown in Figure 1, the diagnostic model proposed in this embodiment includes three types of encoders (source domain encoder, target domain encoder, and shared encoder), a domain classifier, a gradient reversal layer (GRL), and a final classifier. The core concept is to input partial discharge signals from the source domain (SF6 gas) and target domain (C4F7N / CO2 / O2 mixed gas) into three sets of encoders to extract features. The features extracted from the shared encoder are used as the basis for classification. The features undergo domain separation and adversarial optimization in the structure to achieve cross-domain recognition.

[0066] First, the structures of the three encoders are consistent with those used in the original PD classification model. The only difference is the parameter initialization method: the shared encoder is initialized using the model parameters trained on SF6 gas, while the source domain and target domain dedicated encoders are randomly initialized. The high-dimensional features output by each encoder can be expressed as:

[0067] Source domain shared features:

[0068] Source domain-specific features:

[0069] Target domain shared features:

[0070] Target domain-specific features:

[0071] In order to enhance the separability of shared features and proprietary features, the Pearson correlation coefficient is introduced as a constraint to minimize the correlation between shared and proprietary features, thereby achieving feature decoupling. The corresponding feature loss function is:

[0072]

[0073] in Represents the Pearson correlation coefficient, the smaller the value, the more independent the features.

[0074] At the same time, in order to achieve cross-domain consistency of the shared feature space, a domain adversarial module is introduced, in which the GRL (gradient reversal layer) changes the gradient direction during backpropagation, prompting the shared encoder to extract domain-indistinguishable shared features. Its mathematical expression is:

[0075]

[0076] In the forward propagation, GRL is a unit map; in the back propagation, its gradient is multiplied by -1, which prompts the shared encoder to extract domain-independent features.

[0077] After GRL, the shared features are input into the domain classifier and trained adversarially using the binary cross entropy loss function, so that the model can both distinguish domain sources and optimize the universality of shared features. The domain classification loss function is defined as:

[0078]

[0079] in and denote the domain classification prediction of the exclusive and shared features of the source domain, and is the corresponding prediction result of the target domain.

[0080] To ensure classification accuracy, the source domain shared features are used to perform category prediction through the classifier, and the classification loss is defined as cross entropy:

[0081]

[0082] in represents the true label, represents the model prediction probability, is the sample size, is the number of categories.

[0083] The total loss function of the final model is composed of the above three parts, and its weighted form is as follows:

[0084]

[0085] in and Hyperparameters for adjusting feature decoupling and domain adversarial weights.

[0086] The training was completed on an NVIDIA RTX 4060 GPU using the AdamW optimizer, with an initial learning rate of 0.0001 and hyperparameters , Even with very few PD signal samples in the target domain (the number of target domain samples is 0.1% of that in the source domain), the model can still achieve high-precision identification of partial discharge types in C4F7N / CO2 / O2 gases, effectively improving the model's adaptability to new environmentally friendly gas GIS.

[0087] Step 5. Performance Verification

[0088] method SF6 diagnostic accuracy (%) Environmentally friendly gas diagnostic accuracy (%) Parameter quantity DA Model (UHF+US) 99.97 99.80 3.3 DA Model (UHF) 99.98 99.99 3.3 DA model (US) 99.96 99.84 3.3 ResNet (UHF+US) 99.99 68.58 30.7 LSTM (UHF+US) 90.41 49.84 0.0026 Transformer (UHF+US) 99.86 84.62 26.8 GoogLeNet (UHF+US) 99.99 74.07 3.4

[0089] Table 1 Diagnostic accuracy of different methods for partial discharge in various types of GIS

[0090] In order to verify the practicality and superiority of the unsupervised domain-adaptive partial discharge diagnosis algorithm proposed in this invention that integrates domain separation and domain adversarial mechanisms in the new environmentally friendly gas GIS, the present invention designed and carried out systematic model comparison experiments and ablation experiments.

[0091] First, we used a test set of signals collected from a C₄FₐN / CO₂ / O₂ gas mixture to evaluate the performance of the proposed model (hereinafter referred to as the DA model, or Domain-Adaptive Model). We compared the model with four mainstream deep learning methods (including ResNet, LSTM, GoogLeNet, and Transformer) and a PD signal classification model without domain adaptation. The results are shown in Table 1.

[0092] As shown in the table, the proposed DA model achieves over 99% classification accuracy in the target domain (C4F7N / CO2 / O2 gases) under all three input modes (UHF signal, US signal, and UHF+US signal fusion), significantly outperforming traditional methods that do not incorporate domain adaptation. Specifically, the DA model in the fusion input mode achieves recognition accuracies of 99.97% and 99.80% in the SF6 and C4F7N / CO2 / O2 domains, respectively. This demonstrates that this method effectively mitigates the accuracy degradation caused by differences in feature distribution in the target domain while maintaining performance in the source domain.

[0093] Furthermore, to evaluate the robustness of this method under the condition of scarce target domain samples, the present invention conducted a sensitivity experiment on the source / target domain sample ratio. By adjusting the number of unlabeled target domain samples involved in training, the model's recognition accuracy on the target domain (C4F7N / CO2 / O2) under different data ratios was tested. The results are shown in Figure 2. Figure 7 shown.

[0094] Experimental results show that even when the number of target domain samples is only 0.1% of the source domain, the DA model still achieves 99.99% and 99.84% recognition accuracy in UHF and US input modes, respectively, demonstrating extremely high sample utilization efficiency and cross-domain transfer capabilities. However, when the target domain data is further reduced to 0.05% or even 0.01%, the model's recognition accuracy in the fusion input mode drops significantly, indicating that the fusion of multimodal features is more sensitive to the number of training samples. Therefore, to ensure model performance, it is recommended that the ratio of target domain data to source domain data should be no less than 0.1%.

[0095] In summary, the DA model can accurately identify four typical types of partial discharge defects in C₄FₐN / CO₂ / O₂ gas environments while maintaining minimal target domain data input, significantly outperforming traditional methods. Its strong generalization and high recognition accuracy demonstrate the value of the proposed method in the field of online partial discharge monitoring and fault diagnosis in environmentally friendly GIS.

[0096] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A multi-type GIS partial discharge diagnosis method based on domain adaptation, characterized by: The following steps are involved: Build a GIS partial discharge experimental platform to collect partial discharge signals, including electromagnetic wave signals and ultrasonic signals, in a C4F7N / CO2 / O2 mixed gas environment; Perform fast Fourier transform and normalization processing on the collected partial discharge signals to construct a partial discharge dataset containing pinpoint defects, suspension defects, particle defects, and surface discharge defects; Constructing a partial discharge classification model based on CNN and Transformer to extract features and classify the partial discharge signal; An unsupervised domain adaptation mechanism is introduced to eliminate the feature distribution differences between SF6 gas and C4F7N / CO2 / O2 mixed gas through the domain separation module and the domain adversarial module; The unsupervised domain adaptation mechanism introduced includes: a separation module: achieving feature decoupling by minimizing the Pearson correlation coefficient between shared features and exclusive features; and a domain adversarial module: extracting domain-invariant features through a gradient reversal layer and a domain classifier; The trained model is used to diagnose the type of partial discharge in the target gas environment.

2. The multi-type GIS partial discharge diagnosis method based on domain adaptation according to claim 1 is characterized in that: The GIS partial discharge experimental platform includes: The GIS equipment is filled with a C4F7N / CO2 / O2 mixed gas at a pressure of 0.5 MPa, where the volume ratio of the mixed gas is C4F7N:CO2:O2=8.5:86:5.5; UHF sensor and ultrasonic sensor are used to collect electromagnetic wave signals and ultrasonic signals of partial discharge respectively; High-voltage power supply, PD detector and oscilloscope are used to stimulate partial discharge and record the signal.

3. The multi-type GIS partial discharge diagnosis method based on domain adaptation according to claim 1 is characterized in that: The partial discharge signals include four typical defect types: For needle tip defects, a metal tip with a bottom diameter of 8 mm and a length of 20 mm is used, and the tip is 5 mm away from the low-voltage conductor; Suspension defects are fixed between the plates using metal bolts through epoxy resin, 8 mm from the high-voltage conductor and 5 mm from the low-voltage conductor; Particle defects were detected by placing a 0.5 mm diameter steel ball in an epoxy resin cylinder; The surface discharge defect was formed by contacting a metal bolt with an epoxy resin cylinder with a diameter of 50 mm and a thickness of 20 mm.

4. The multi-type GIS partial discharge diagnosis method based on domain adaptation according to claim 1 is characterized in that: The CNN and Transformer partial discharge classification model includes: The encoder part consists of a one-dimensional convolutional layer with a large convolution kernel, four ResCNN Block modules, and three TransformerBlock modules, which are used to extract multi-scale features of the signal; Classifier part: consists of two fully connected layers, which are used to output the probability distribution of defect types.

5. The multi-type GIS partial discharge diagnosis method based on domain adaptation according to claim 1 is characterized in that: The loss function of the domain separation module is: in represents the Pearson correlation coefficient, represents the shared features of the source domain, represents the source domain-specific features, represents the shared features of the target domain, Represents target domain-specific features.

6. The multi-type GIS partial discharge diagnosis method based on domain adaptation according to claim 1 is characterized in that: The loss function of the domain adversarial module is: Where, and denote the domain classification prediction of the exclusive and shared features of the source domain, and is the corresponding prediction result of the target domain, is the binary cross entropy loss function.

7. A GIS partial discharge diagnostic system implementing the method according to any one of claims 1 to 6, characterized in that: include: Signal acquisition unit: used to obtain partial discharge ultrasonic signals and UHF electromagnetic wave signals in GIS equipment; Data processing unit: used for signal preprocessing and feature extraction; Intelligent diagnosis unit: deploys a CNN-Transformer hybrid model trained with domain adaptation; Result display unit: outputs partial discharge type and confidence level.

8. The GIS partial discharge diagnostic system according to claim 7, characterized in that: The intelligent diagnosis unit supports three working modes: pure UHF signal analysis mode; pure ultrasonic signal analysis mode; multi-modal signal fusion analysis mode.

Images