Partial discharge identification method and system for gas insulated switchgear

Abstract

A gas insulated switchgear partial discharge identification method and system based on unsupervised domain adversarial transfer learning, aiming at the problem that the recognition accuracy of the traditional SF6 gas partial discharge identification model significantly decreases under the condition of new environmental protection gas (C4F7N / CO2 / O2 mixed gas), a double classifier unsupervised domain adversarial transfer learning (DCUDA) model is used, through the cooperative optimization of the encoder, double classifier and domain discriminator, combined with the three-stage training strategy (source domain pre-training, classifier difference maximization and adversarial optimization), the feature self-adaptive alignment under the condition of strong domain shift is realized. This method comprehensively utilizes the UHF and UL dual-mode signals, extracts time domain, frequency domain and statistical multi-dimensional features, and under the condition of no labeled samples in the target domain, the recognition accuracy in the new gas environment is greatly improved, the model construction cost of the new environmental protection GIS equipment is reduced, and an efficient and reliable technical scheme is provided for the intelligent operation and maintenance of power equipment.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of gas insulated switchgear (GIS) insulation defect detection, and particularly relates to a gas insulated switchgear partial discharge identification method and system based on unsupervised domain adversarial transfer learning, especially novel environment-friendly SF6 gas and C4F7N / CO2 / O2 mixed gas. BACKGROUND

[0002] Gas insulated switchgear (GIS) has been widely used in modern power systems due to its small size, convenient maintenance, and high reliability. However, during manufacturing, transportation, installation, and operation, GIS inevitably has insulation defects, which can easily cause different types of partial discharge (PD). Partial discharge can cause insulation performance degradation and even equipment failure or safety accidents, so accurate identification of partial discharge types is of great significance for the stable operation and safety of GIS.

[0003] Currently, the commonly used insulation gas in GIS is sulfur hexafluoride (SF6), which has excellent insulation performance and arc extinguishing ability. However, its global warming potential (GWP) is as high as about 23500 times (based on carbon dioxide), making it a strong greenhouse gas with serious environmental impact. Therefore, the use of SF6 gas is strictly limited, and there is an urgent need to develop alternative gases. In recent years, environmentally friendly insulation gases such as heptafluoroisobutyl cyanide (C4F7N) have become one of the important substitutes for SF6 gas due to their excellent insulation properties and lower environmental impact. Usually, C4F7N is mixed with CO2 to reduce the liquefaction temperature and improve the insulation strength; further adding a certain proportion of O2 can effectively improve the insulation performance of the mixed gas and inhibit the decomposition of C4F7N. Therefore, C4F7N / CO2 / O2 mixed gas is considered as one of the ideal solutions to replace SF6 gas.

[0004] The partial discharge detection methods of GIS mainly include pulse current method (PCM), ultra-high frequency (UHF) method, ultrasonic (UL) method and optical detection method. Among them, the UHF method and the UL method are most widely used because of their online detection capability. However, the existing PD identification models are mostly established based on data under SF6 gas conditions. Although these models show high identification accuracy under SF6 conditions, when applied to C4F7N / CO2 / O2 mixed gas, due to the obvious domain difference (Domain Shift) between the partial discharge signals of the two gases, the recognition accuracy of the traditional model under the condition of new type of insulating gas is significantly reduced.

[0005] To solve the above problems, the usual way is to construct an independent PD identification model for each type of insulating gas. However, this method has the following disadvantages:

[0006] (1) The independent construction of identification models for different insulating gases significantly increases the cost of model development and maintenance.

[0007] (2) The PD data of new type of insulating gas such as C4F7N / CO2 / O2 mixed gas is relatively scarce, and the data volume is not enough to fully train a new identification model, which seriously restricts the promotion and application of new type of insulating gas.

[0008] Therefore, how to develop a high-precision PD identification method that can be applied to both traditional SF6 gas and new type of insulating gas (such as C4F7N / CO2 / O2 mixed gas) under limited data conditions has become one of the key problems to be solved in the current GIS partial discharge monitoring technology.

[0009] In view of the above problems, the present application proposes a new type of environmentally friendly gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning, which can effectively improve the accuracy and reliability of the new type of environmentally friendly gas GIS partial discharge detection, thereby providing a new type of technical solution for the safe operation of new type of environmentally friendly gas GIS equipment, and has significant engineering application value. SUMMARY

[0010] The purpose of the present application is to solve the problems of existing partial discharge identification models, such as the decline of recognition accuracy, the lack of model generalization ability and the scarcity of target domain samples when facing new environmentally friendly insulating gases (such as C4F7N / CO2 / O2 mixed gas), and to propose a gas insulated switchgear partial discharge identification method based on double classifier unsupervised domain adversarial transfer learning. This method introduces a cross-domain feature extraction mechanism and an adversarial learning strategy, and realizes high-precision identification of partial discharge signals under various insulating gases on the premise of only needing a small amount of target domain data, effectively improving the adaptability, robustness and generalization ability of the model. In order to achieve the above purpose, the present application adopts the following technical solutions:

[0011] On the one hand, the present application provides a gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning, characterized in that it comprises the following steps: S1. Constructing a partial discharge experimental platform, simulating four typical discharge defects of needle-point discharge, suspension discharge, particle discharge and surface discharge under SF6 gas and C4F7N / CO2 / O2 mixed gas conditions, synchronously collecting ultra-high frequency (UHF) and ultrasonic (UL) partial discharge signal data to form source domain and target domain data sets; S2. Extracting multi-dimensional features of the PD signals collected in step S1, including time domain features, frequency domain features and statistical features, and constructing a structured feature set for transfer learning;

[0012] S3. Based on the feature set extracted in step S2, a double classifier unsupervised domain adversarial transfer learning (DCUDA) model is constructed, which comprises:

[0013] - an encoder for extracting high-dimensional features shared across domains;

[0014] - a double classifier for guiding feature alignment through classification difference maximization;

[0015] - a domain discriminator for realizing domain-invariant feature learning through adversarial training;

[0016] S4. The DCUDA model of step S3 is optimized using a three-stage training strategy, which is executed in the following order:

[0017] - pre-training stage: use the source domain data to train the encoder and the double classifier to establish the initial classification ability;

[0018] - classifier difference maximization stage: freeze the encoder and optimize the double classifier to expand the output difference of the target domain, identify the domain offset area;

[0019] - adversarial optimization stage: freeze the double classifier, jointly train the encoder and the domain discriminator, minimize the domain discriminant loss and the classifier difference, and realize feature space alignment.

[0020] Further, the step S5 comprises: based on the model trained in the step S4, verifying the cross-domain feature alignment effect by using mixed signal input and t-SNE visualization analysis, and evaluating the classification performance through a confusion matrix.

[0021] Further, in the step S2:

[0022] Time domain features, including peak value, charge quantity, pulse width, signal duration and initial rising rate, are used to characterize the transient characteristics of the discharge waveform;

[0023] Frequency domain features, including dominant frequency, average frequency, spectral energy density, spectral peak value, spectral kurtosis and skewness, are used to analyze the influence of the gas medium on the spectral distribution;

[0024] Statistical features, including signal distribution symmetry and energy concentration, are used to quantify the distribution difference between domains and provide a difference measurement basis for the adversarial training in the step S3.

[0025] Further, in the step S3:

[0026] The encoder is composed of one-dimensional convolution (Conv1D), residual convolutional network (ResCNN) and Transformer module, which is used to extract high-dimensional features of the PD signal, and the output of the encoder is input into the dual classifier and the domain discriminator at the same time, forming a closed-loop optimization link of feature extraction-classification discrimination-domain confrontation;

[0027] The dual classifier is composed of two classifiers with the same structure but not sharing parameters, each classifier including an average pooling layer and a fully connected layer, and the difference loss of the dual classifier guides the encoder to adjust the feature extraction strategy in reverse;

[0028] The domain discriminator is used to distinguish the source domain and target domain features, and realizes the adversarial training through the gradient reversal layer (GRL), and the domain discriminator and the encoder constitute an adversarial game through the gradient reversal layer (GRL), forcing the encoder to generate domain-invariant features.

[0029] On the other hand, the application also provides a partial discharge identification system for a gas insulated switchgear, characterized in that it comprises:

[0030] The signal acquisition module is configured to synchronously acquire original signals of four typical discharge defects under SF6 gas and C4F7N / CO2 / O2 mixed gas conditions through UHF sensors and UL sensors, and output source domain dataset SF6 and target domain dataset C4F7N / CO2 / O2 containing gas type labels to the feature engineering module;

[0031] The feature engineering module is configured to receive the raw data of the signal acquisition module, extract time domain, frequency domain and statistical features, classify the generated feature set according to gas types, mark the source domain features with discharge type labels, retain only domain labels for the target domain features, and output a structured feature matrix to the model training module and the performance verification module.

[0032] The DCUDA model training module is configured to receive the feature matrix of the feature engineering module, load a DCUDA model composed of an encoder, a double classifier and a domain discriminator, output an optimized model after training, and realize cross-domain feature alignment and partial discharge type recognition.

[0033] Further, the performance verification module is configured to receive the target domain features of the feature engineering module and the optimized model of the DCUDA model training module, perform hybrid signal testing, t-SNE visualization and confusion matrix generation, feed back classification accuracy and feature alignment degree indicators to the DCUDA model training module, and output a final evaluation report to trigger model deployment or iterative training.

[0034] Further, the feature engineering module comprises:

[0035] The signal preprocessing unit is configured to receive the raw data of the signal acquisition module and perform noise reduction and normalization processing.

[0036] The feature extraction unit is configured to extract time domain, frequency domain and statistical features from the preprocessed signals.

[0037] The feature labeling unit is configured to add discharge type labels to the source domain data and add domain labels to the target domain data.

[0038] Further, the performance verification module comprises:

[0039] The online testing unit is configured to receive real-time field collected data for model verification.

[0040] The visualization analysis unit is configured to generate t-SNE feature distribution graphs and confusion matrices.

[0041] The performance evaluation unit is configured to calculate and store key indicators such as accuracy and recall rate.

[0042] Thirdly, the application also provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the above method.

[0043] Compared with the prior art, the application has the following advantages:

[0044] 1. Realize high-precision partial discharge recognition across gas domains, significantly improve model generalization ability: the application realizes domain adaptation between SF6 gas and C4F7N / CO2 / O2 mixed gas without target domain labeled data by introducing a double classifier unsupervised domain adversarial transfer learning (DCUDA) framework, solves the problem of significant decline in precision of traditional models in new environmental gas scene, greatly improves the cross-domain recognition ability and practical application range of the model.

[0045] 2. Fusion of multi-modal signal and multi-dimensional feature, effectively enhance recognition robustness and reliability: the application comprehensively utilizes UHF and UL double modal partial discharge signal, and combines time domain, frequency domain and statistical characteristics to construct high-quality feature set, effectively captures the key features of partial discharge signal, improves the model's ability to distinguish different discharge types under complex working conditions, has stronger environmental adaptability and engineering robustness.

[0046] 3. Reduce the cost of new gas model construction, improve the efficiency of engineering deployment: through end-to-end transfer learning strategy, the application greatly reduces the dependence of new gas PD samples, realizes the rapid adaptation and deployment of the model in the target domain, avoids the high cost of repeatedly training the recognition model for each gas, has good scalability and promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 It is a schematic diagram of a gas insulated switchgear partial discharge recognition method based on unsupervised domain adversarial transfer learning, wherein (a) is a schematic diagram of an experimental platform, and (b) is a flowchart;

[0048] Figure 2 Four kinds of simulated partial discharge defect schematic diagram;

[0049] Figure 3 Time domain partial discharge signal, wherein (a) is UHF signal, and (b) is UL signal;

[0050] Figure 4 Partial discharge UHF signal spectrum;

[0051] Figure 5 Partial discharge ultrasonic signal spectrum, wherein (a) is UHF signal, and (b) is UL signal;

[0052] Figure 6 Encoder structure;

[0053] Figure 7 t-SNE visualization, wherein (a) SF 6， (b) C4F7N / CO2 / O2;

[0054] Figure 8 Classifier structure;

[0055] Figure 9 domain discriminator structure;

[0056] Figure 10 DCUDA training process. DETAILED DESCRIPTION

[0057] The technical solutions of the present application will be described in detail below in combination with the drawings and examples, but the protection scope of the present application should not be limited thereto.

[0058] A gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning, comprising the following steps:

[0059] S1. Construct a partial discharge experimental platform, based on real GIS equipment, simulate four typical discharge defects (needle, suspension, particle and surface) under SF6 gas and C4F7N / CO2 / O2 mixed gas conditions, respectively, and collect ultra-high frequency (UHF) and ultrasonic (UL) partial discharge signal data.

[0060] S2. Perform feature analysis and processing on the acquired PD signals, use time domain waveform analysis, frequency domain Fourier transform and statistical feature extraction methods, etc., to mine the differences between different insulation gas discharge signals, and construct a data feature set for modeling. Through multi-dimensional feature extraction (time domain, frequency domain and statistical features), a high-quality feature set is constructed, providing a data basis for model training.

[0061] S3. Construct a double classifier unsupervised domain adversarial transfer learning (DCUDA) model, which is composed of an encoder, a double classifier and a domain discriminator. Through a multi-stage training strategy, the high-dimensional features are optimized in an adversarial manner, and the feature alignment between the source domain and the target domain is realized.

[0062] S4. Three-stage training strategy

[0063] Pre-training stage: use the source domain data to train the encoder and the double classifier to ensure the high-precision classification ability of the model in the source domain.

[0064] Classifier training stage: freeze the encoder, maximize the output difference of the double classifier on the target domain data, and provide an optimization direction for subsequent adversarial training.

[0065] Adversarial optimization stage: freeze the classifier, train the encoder and the domain discriminator, minimize the domain discriminant loss and the classifier output difference, and realize cross-domain feature alignment.

[0066] S5. Performance verification

[0067] The model performance is verified by using mixed signal input, t-SNE visualization analysis, confusion matrix evaluation and the like, and performance tests are carried out in combination with different target domain sample proportions, signal-to-noise ratio conditions and signal type input modes to evaluate the performance of the model in terms of recognition accuracy, robustness and sample dependency.

[0068] The specific method of step (1) is:

[0069] To realize effective acquisition and analysis of partial discharge (PD) signals under the condition of new insulation gas, the present application designs and builds a set of partial discharge experimental platform based on actual 252 kV gas-insulated switchgear (GIS), as shown in Figure 1 The experimental platform can simulate typical discharge defects under controllable conditions and carry out comparative experiments in traditional SF6 gas and new C4F7N / CO2 / O2 mixed gas environments.

[0070] Experimental gas and pressure setting

[0071] The experiment is carried out under two insulation gas conditions: one is the traditional SF6 gas, and the other is the environmentally friendly mixed gas with a volume fraction ratio of 8.5% C4F7N / 86% CO2 / 5.5% O2. Both gases are maintained at an insulation pressure of 0.5 MPa during the experiment to ensure comparability.

[0072] Typical defect construction

[0073] To comprehensively study various discharge behaviors, the platform is provided with four typical partial discharge defects, including needle defect, floating defect, particle defect and surface defect, as shown in Figure 2

[0074] · Needle discharge: simulated by a metal needle with a diameter of 8 mm and a length of 20 mm.

[0075] · Floating discharge: a metal bolt is fixed by epoxy resin at a distance of 8 mm from the upper conductor and 5 mm from the lower conductor.

[0076] · Particle discharge: a metal microsphere with a diameter of 0.5 mm is placed on the surface of the conductor and wrapped in an epoxy tube.

[0077] · Surface discharge: a cylindrical epoxy block with a diameter of 50 mm and a thickness of 20 mm is placed between the conductors, and the upper and lower electrodes are contacted by bolts. ​

[0078] These structures simulate typical insulation defects that may occur in GIS devices in operation, ensuring the engineering representativeness of experimental data.

[0079] Sensor arrangement and signal acquisition

[0080] Two types of non-intrusive sensors are used in the experimental platform for PD signal acquisition:

[0081] • UHF sensor: frequency range 0.3-1.5 GHz, installed on the GIS handle hole cover plate, used to collect electromagnetic radiation signals;

[0082] • UL sensor: frequency range 20-100 kHz, adhered to the GIS shell surface, used to collect ultrasonic signals.

[0083] The distance between the sensor and the defect is uniformly set to 30 cm to control the influence of spatial variables. The system uses a Haefely DDX 9121b high-precision PD detector to measure the initial discharge voltage, and all signals are recorded synchronously by a LeCroy 204XI-A high-speed oscilloscope to ensure data accuracy and consistency.

[0084] Data acquisition strategy

[0085] Under each gas, each defect type, and each sensor combination, 10,000 PD signal data are collected, and a high-quality PD data set containing 160,000 samples is constructed. This data set not only covers multiple discharge types and working conditions, but also provides a solid data foundation for subsequent signal analysis and model training.

[0086] The specific method of step (2) is:

[0087] To construct a high-precision partial discharge recognition model suitable for various insulation gas environments, the present application conducts systematic multi-dimensional feature analysis on the collected PD signals. The analysis includes time domain waveform structure, frequency domain spectrum distribution, and statistical characteristic differences, aiming to explore the signal differences and commonalities exhibited by various discharge defects in different gas media.

[0088] Time domain feature analysis

[0089] By comparing the original PD signal waveforms collected by UHF and UL sensors, the time domain response characteristics of different discharge defect types in SF6 gas and C4F7N / CO2 / O2 mixed gas are identified (see Figure 3 ). The results show that:

[0090] • The UHF signals generally exhibit high-speed pulses with abrupt changes near time zero, followed by rapid decay, reflecting the transient nature of charge acceleration and radiation in PD. The decay rate of particle discharge is significantly slower than other discharge types, indicating a more persistent charge release process.

[0091] • The UL signal waveforms have strong gas dependence. In SF6 gas, the UL waveforms of different discharge types have high distinguishability, while in C4F7N / CO2 / O2 gas, the signal waveforms tend to be consistent, showing reduced ability to distinguish defect types.

[0092] This phenomenon indicates that UHF signals are more stable in the two gas environments and have cross-domain universality, while UL signals are more sensitive to gas properties and are suitable for extracting domain-specific features.

[0093] Frequency domain feature analysis

[0094] Fast Fourier Transform (FFT) is applied to the time-domain PD signals to obtain their frequency spectrum distribution, revealing the energy characteristics of different discharge types in the frequency dimension. The analysis results are shown in Figure 4 and Figure 5 .

[0095] • The frequency spectrum of UHF signals is mainly concentrated in the 0.3-1.5 GHz band, and the spectral shape difference between SF6 gas and C4F7N / CO2 / O2 gas is small. This is due to the fact that UHF signals are formed by high-frequency electromagnetic radiation, and their frequency components are greatly affected by PD transient current, while the conductivity and dielectric constant of the two gases are close, resulting in similar excitation mechanisms.

[0096] • Although UHF signals are similar in overall shape, statistical analysis shows that there are certain differences in the high-order characteristics of the spectrum. The inventors calculated the kurtosis and skewness of the UHF spectrum, and observed significant differences in floating discharge and other types, with specific values shown in Table 1.

[0097] Table 1 Kurtosis and skewness of UHF partial discharge signals

[0098]

[0099] In contrast, UL signals in the frequency domain show obvious gas dependence. In the 20-30 kHz frequency band, the PD signals generated by SF6 gas have higher energy than those generated by C4F7N / CO2 / O2 gas (see Figure 5 ). This difference is due to the influence of gas molecular mass and acoustic impedance on the propagation of sound waves formed after PD. SF6 molecules have a larger mass, resulting in lower vibration frequency and more concentrated energy in the low frequency band.

[0100] Comprehensive feature construction

[0101] Based on the comprehensive time-domain and frequency-domain analysis results, the following multi-dimensional features are designed for modeling:

[0102] Time-domain features: peak value, charge quantity, pulse width, signal duration, initial rising rate, etc.

[0103] Frequency-domain features: main frequency, average frequency, spectral energy density, spectral peak value, spectral kurtosis and skewness

[0104] Statistical features: signal distribution symmetry, energy concentration, etc.

[0105] Label information: defect type identification and gas domain identification as auxiliary training input.

[0106] This feature system can effectively capture the signal differences between different gas environments and discharge types, and provide a transferable structured feature space basis for subsequent transfer learning models.

[0107] The specific method of step (3) is:

[0108] In order to overcome the problem of degradation of recognition performance caused by significant domain differences in partial discharge signal features under different insulation gas conditions, the present application proposes a dual-classifier unsupervised domain adversarial model (Dual-Classifier Unsupervised Domain Adversarial, DCUDA) based on unsupervised transfer learning framework. By introducing an adversarial training strategy and feature consistency discrimination mechanism, the model realizes the alignment of features between the source domain and the target domain, thereby significantly improving the recognition accuracy and robustness of the model in the target domain under the premise of only requiring a small amount of labeled samples in the target domain.

[0109] The core structure of the DCUDA model consists of three modules: an encoder, two parallel classifiers, and a domain discriminator. First, the encoder is responsible for feature extraction of the input PD signal (which can be UHF or UL signal, or a mixture of the two). In order to fully capture the time sequence structure and cross-time scale information contained in the partial discharge signal, the encoder integrates various neural network modules such as one-dimensional convolution (Conv1D), residual convolution network (ResCNN) and Transformer structure (see Figure 6 ). Among them, the initial 1D convolution layer uses a larger convolution kernel (size 7) to increase the receptive field and extract local context features; then four layers of ResCNN structure are introduced, which can not only deepen the network structure, but also alleviate the gradient vanishing and improve the convergence stability through residual connection; finally, three layers of multi-head attention Transformer module are connected, which significantly enhances the model's ability to perceive long-range dependencies and non-local structures in PD signals.

[0110] In order to intuitively present the distribution of the high-dimensional characteristics of the PD signal under different gas conditions, the t-SNE algorithm is used to reduce the dimension and visualize the characteristics of the encoder output (see Figure 7 ). In SF6 gas, the feature clustering effect of different defect categories is good, and the boundary between categories is clear, indicating that the model has good separability; while in C4F7N / CO2 / O2 mixed gas, the feature distribution is significantly mixed, and the boundary between different categories is blurred, indicating that the model trained directly using SF6 is difficult to generalize to the target domain, representing a significant domain shift phenomenon.

[0111] To solve the above problems, the present application sets two classifiers with the same structure but not sharing parameters in parallel after the encoder (see Figure 8 ). Each classifier is composed of an average pooling layer and two linear fully connected layers. The two classifiers are trained to be consistent in the source domain data, but are guided to output different classification results as much as possible in the target domain, thereby providing training signals for the encoder to learn shared domain invariant features under the premise that the target domain data lacks labels. The greater the difference between the predictions of the two classifiers for the same target domain sample, the less consistent the target domain feature distribution with the source domain, and vice versa, indicating a higher degree of domain alignment. Therefore, by minimizing this difference in the subsequent training stage, the encoder is effectively guided to learn feature representations with cross-domain consistency.

[0112] At the same time, to further improve the consistency of the feature space for the source domain and the target domain, the present application designs a domain discriminator module for source / target domain discrimination of the features extracted by the encoder (see Figure 9 ). In the forward propagation process, the features first pass through the gradient reversal layer (Gradient Reversal Layer, GRL) into the discriminator. The role of GRL is to automatically reverse the gradient sign in the backpropagation phase, thereby forcing the encoder to generate "domain indistinguishable" features that are difficult to be identified by the domain discriminator. The domain discriminator is composed of a residual convolution module and a ReLU activation function inside, outputs the domain discrimination probability, and is optimized through the binary cross-entropy loss function. The discriminator and the encoder form a game relationship: the discriminator tries to maximize the discrimination ability between the source domain and the target domain, while the encoder tries to minimize this ability, i.e. to achieve cross-domain feature alignment.

[0113] The overall training process of the DCUDA model is divided into three stages, as follows Figure 10The first stage is a pre-training stage, in which the encoder and the two classifiers are trained using a large amount of SF6 source domain data, so that they have good classification ability on the source domain. The second stage freezes the encoder and only trains the two classifiers, with the goal of maximizing the output difference of the two classifiers on the target domain data, thereby expanding the discriminant boundary and providing an optimization direction for subsequent adversarial training. In the third stage, the classifiers are frozen, and the encoder and the domain discriminator are retrained. At this time, the encoder is guided to extract shared features that are consistent for the outputs of the two classifiers and cannot be identified by the domain discriminator, thereby finally realizing the alignment and migration of the source domain and the target domain in the feature space and solving the problem of inconsistent feature distribution caused by different gas compositions.

[0114] The specific method of step (4) is:

[0115] To comprehensively evaluate the partial discharge (PD) recognition performance of the double-classifier unsupervised domain adversarial transfer learning (DCUDA) model proposed in the application under different gas medium conditions, the application constructs a systematic performance verification method and experimental device. This method not only covers the overall classification accuracy test of the model on the source domain and the target domain, but also further considers the adaptability of the model to data proportion changes, signal input methods, and noise interference and other actual working conditions.

[0116] In the model training and testing process, the UHF and UL original PD signals collected by the experimental platform are first subjected to Fourier transform by the "data processing module", and the spectral features thereof are extracted as model input data. The PD signals under SF6 gas are used as source domain data, and the PD signals under C4F7N / CO2 / O2 gas are used as target domain data, and the two are constructed into a training set and a test set in an adjustable proportion, wherein the proportion of target domain samples in the training set is adjusted as a hyperparameter. Under the standard setting, the proportion of source domain and target domain test data is fixed at 1:1 to ensure the fairness of the evaluation indicators.

[0117] In the "training module", the model undergoes three stages of pre-training, classifier training and adversarial optimization, and finally completes the optimal fitting of the DCUDA model structure parameters. During the training and optimization process, the classification loss, the classifier output difference loss and the domain discrimination loss are minimized to realize cross-domain feature alignment and class separation. After optimization, the "test module" is entered, the test set data is input into the model and the class prediction results are output, the classification accuracy, confusion matrix, t-SNE visualization distribution and other performance indicators are calculated by comparing with the true labels, to quantify the generalization ability and migration performance of the model under actual working conditions. The final test results are shown in Table 2.

[0118] Table 2 Model test results

[0119]

[0120] To verify the robustness of the model, the present application also tests the noise robustness of the model under different signal-to-noise ratios. In the original signal, Gaussian noise with SNR of 25 dB and 5 dB is superimposed respectively, and the results show that the DCUDA model still maintains 97.6% of the target domain recognition accuracy under high noise interference, showing good anti-interference ability.

[0121] In addition, to evaluate the influence of the proportion of target domain data on the performance of the model, the present application system tests the classification accuracy when the proportion of target domain samples is 0.1, 0.05, 0.01, 0.005, 0.001 and 0.0005 respectively. The results are shown in Table 3, although the target domain samples are extremely rare (for example, only 0.001 times of the source domain samples), the DCUDA model can still maintain a high accuracy of 98.79% in the target domain, indicating that the model has a small dependence on target domain data and has strong low sample generalization ability.

[0122] Table 3 Influence of data proportion on model performance

[0123]

[0124] Further, the present application also uses the t-SNE dimensionality reduction visualization method to compare and analyze the distribution of the DCUDA model and other comparative models (such as ResNet, Transformer, LSTM) in the source domain and target domain feature space. The results show that although the traditional model has good classification ability in the source domain, it shows obvious feature domain offset phenomenon in the target domain; while the DCUDA model realizes the high overlap of the source domain and target domain features, indicating that it realizes effective domain alignment in the deep space.

[0125] In summary, the present application maximizes the output difference of the target domain by using double classifiers, and realizes cross-domain feature alignment under unsupervised conditions by combining the adversarial training of the domain discriminator, solving the limitation of traditional methods that rely on target domain labeled data. Combined with Conv1D, ResCNN and Transformer modules, the local and global features of the PD signal are fully extracted, and the representation ability of the model for complex signals is enhanced. The encoder, classifier and domain discriminator are optimized in stages, and the feature alignment is gradually realized, which significantly improves the convergence speed and generalization performance of the model. The present application provides an efficient and reliable technical solution for the recognition of new environmental gas GIS equipment partial discharge, and has important engineering application value.

[0126] Although the specific embodiments of the present application have been described above in combination with the drawings, it is not a limitation on the scope of protection of the present application, and those skilled in the art should understand that various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the scope of protection of the present application.

Claims

1. A gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning, characterized in that, Comprising the following steps: S1. Constructing a partial discharge experimental platform, simulating four typical discharge defects of needle-plate discharge, floating discharge, particle discharge and surface discharge under SF6 gas and C4F7N / CO2 / O2 mixed gas conditions, synchronously collecting ultra-high frequency (UHF) and ultrasonic (UL) partial discharge signal data to form source domain and target domain data sets; S2. Extracting multi-dimensional features from the partial discharge signals collected in step S1, including time domain features, frequency domain features and statistical features, and constructing a structured feature set for transfer learning; S3. Based on the feature set extracted in step S2, a double classifier unsupervised domain adversarial transfer learning (DCUDA) model is constructed, which includes: - an encoder for extracting high-dimensional features shared across domains; - a double classifier for guiding feature alignment through classification difference maximization; - a domain discriminator for realizing domain-invariant feature learning through adversarial training; S4. A three-stage training strategy is used to optimize the double classifier unsupervised domain adversarial transfer learning model of step S3, which is executed in turn: - pre-training stage: use the source domain data to train the encoder and the double classifier to establish the initial classification ability; - classifier difference maximization stage: freeze the encoder and optimize the double classifier to expand the output difference of the target domain to identify the domain offset area; - adversarial optimization stage: freeze the double classifier, jointly train the encoder and the domain discriminator, minimize the domain discriminator loss and the classifier difference, and realize feature space alignment.

2. The gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning according to claim 1, characterized in that, Further comprising: S5. Based on the model trained in step S4, mixed signal input and t-SNE visualization analysis are used to verify the cross-domain feature alignment effect, and the confusion matrix is used to evaluate the classification performance.

3. The gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning according to claim 1, characterized in that, In step S2: Time domain features include peak value, charge quantity, pulse width, signal duration and initial rise rate, which are used to characterize the transient characteristics of discharge waveform; Frequency domain features include main frequency, average frequency, spectral energy density, spectral peak value, spectral kurtosis and skewness, which are used to analyze the influence of gas medium on spectral distribution; Statistical features include signal distribution symmetry and energy concentration, which are used to quantify the distribution difference between domains and provide a difference measurement basis for the adversarial training of step S3.

4. The gas insulated switchgear partial discharge identification method based on unsupervised domain adversarial transfer learning according to claim 1, characterized in that, In step S3: The encoder is composed of one-dimensional convolution (Conv1D), residual convolutional network (ResCNN) and Transformer module, which is used to extract high-dimensional features of partial discharge signals, and the output of the encoder is input into the double classifier and the domain discriminator to form a closed-loop optimization link of feature extraction-classification discrimination-domain adversarial; The double classifier is composed of two classifiers with the same structure but not sharing parameters, each classifier includes an average pooling layer and a fully connected layer, and the difference loss of the double classifier guides the encoder to adjust the feature extraction strategy; The domain discriminator is used to distinguish the source domain and target domain features, and realizes adversarial training through the gradient reversal layer (GRL), and the domain discriminator and the encoder form an adversarial game through the gradient reversal layer (GRL) to force the encoder to generate domain-invariant features.

5. A partial discharge identification system for a gas insulated switchgear, characterized in that Comprising: The signal acquisition module is configured to synchronously acquire original signals of four typical discharge defects under SF6 gas and C4F7N / CO2 / O2 mixed gas conditions through an ultrahigh frequency sensor and an ultrasonic sensor, and output source domain data set SF6 and target domain data set C4F7N / CO2 / O2 containing a gas type label to the feature engineering module; The feature engineering module is used for receiving original data of the signal acquisition module, extracting time domain, frequency domain and statistical features, classifying generated feature sets according to gas types, marking discharge types in source domain features, retaining only domain labels in target domain features, and outputting a structured feature matrix to the double classifier unsupervised domain adversarial transfer learning model training module and the performance verification module; The double classifier unsupervised domain adversarial transfer learning model training module is used for receiving the feature matrix of the feature engineering module, loading a double classifier unsupervised domain adversarial transfer learning model composed of an encoder, a double classifier and a domain discriminator, outputting an optimized model after training, and realizing cross-domain feature alignment and partial discharge type recognition.

6. The partial discharge recognition system for gas insulated switchgear according to claim 5, characterized in that The performance verification module is used for receiving target domain features of the feature engineering module and the optimized model of the double classifier unsupervised domain adversarial transfer learning model training module, performing mixed signal testing, t-SNE visualization and confusion matrix generation, feeding classification accuracy and feature alignment degree indicators back to the double classifier unsupervised domain adversarial transfer learning model training module, and outputting a final evaluation report to trigger model deployment or iterative training.

7. The partial discharge recognition system for gas insulated switchgear according to claim 5, characterized in that, The feature engineering module comprises: A signal preprocessing unit is used for receiving original data of the signal acquisition module and performing noise reduction and normalization processing; A feature extraction unit is used for extracting time domain, frequency domain and statistical features from the preprocessed signals; A feature labeling unit is used for adding discharge type labels to source domain data and adding domain labels to target domain data.

8. The partial discharge recognition system for gas insulated switchgear according to claim 5, characterized in that, The performance verification module comprises: An online testing unit is used for real-time receiving of on-site collected data for model verification; A visualization analysis unit is used for generating t-SNE feature distribution graphs and confusion matrices; A performance evaluation unit is used for calculating and storing accuracy and recall rate key indicators.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the method of any one of claims 1-4.

Images