Office signal classification method based on physical constraint generative adversarial network

Abstract

In order to solve the problem of inaccurate classification caused by fixed time-frequency resolution and lack of physical rationality in sample enhancement in a traditional classification method, the invention relates to a partial discharge signal classification method based on a physical constraint generative adversarial network, and the method comprises the following steps: extracting a partial discharge signal and a historical partial discharge signal, and carrying out the double-window time-frequency coding of the partial discharge signal and the historical partial discharge signal; correspondingly obtaining a fused time-frequency graph and a historical fused time-frequency graph; determining a corresponding discharge type label based on the historical partial discharge signal, establishing a physical constraint generative adversarial network, and training the physical constraint generative adversarial network in the historical fusion time-frequency graph based on the discharge type label; and inputting the fusion time-frequency graph into the trained physical constraint generative adversarial network to obtain a fusion feature graph corresponding to the fusion time-frequency graph, performing time sequence modeling on the fusion feature graph, mapping the fusion feature graph into a time sequence feature vector, and performing classification on the corresponding partial discharge signals based on the time sequence feature vector.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge signal classification, and in particular to a partial discharge signal classification method based on a physical constraint generative adversarial network. Background Art

[0002] Partial discharge signal classification typically involves analyzing electromagnetic wave signals generated by insulation defects in power equipment to determine the discharge type. Partial discharge signals are transient and non-stationary, so accurately extracting their time-frequency characteristics is crucial for accurate classification.

[0003] However, traditional PD signal classification methods generally use fixed window functions for time-frequency analysis. Limited by a single resolution setting, these methods struggle to simultaneously capture both high-frequency resonant components and narrow pulse transient details across a wide frequency range, resulting in insufficient time-frequency representation accuracy. Furthermore, existing methods primarily rely on artificial augmentation strategies for sample expansion, lacking modeling and constraints on the physical characteristics of the discharge. Generated samples can easily deviate from the true discharge mechanism, leading to training bias and thus compromising model generalization performance. These issues limit the accuracy of traditional classification methods for PD signal classification. Summary of the Invention

[0004] In view of the above analysis, the main purpose of the present invention is to provide a partial discharge signal classification method based on a physical constraint generative adversarial network to solve the problem that traditional classification methods have fixed time-frequency resolution and lack of physical rationality in sample enhancement, resulting in inaccurate classification.

[0005] In this regard, the present invention provides a partial discharge signal classification method based on a physical constraint generative adversarial network, comprising the following steps: a. extracting partial discharge signals and historical partial discharge signals and performing dual-window time-frequency encoding on each, thereby obtaining a fused time-frequency graph and a historical fused time-frequency graph; b. determining a corresponding discharge type label based on the historical partial discharge signal, establishing a physical constraint generative adversarial network, and training the physical constraint generative adversarial network based on the discharge type label and the historical fused time-frequency graph; c. inputting the fused time-frequency graph into the trained physical constraint generative adversarial network to obtain a fused feature graph corresponding to the fused time-frequency graph, performing time series modeling on the fused feature graph and mapping it into a time series feature vector, and classifying the corresponding partial discharge signal based on the time series feature vector.

[0006] As preferred, in the step a, the process of extracting the historical partial discharge signals for double-window time-frequency coding comprises: performing wide-window short-time Fourier transform and narrow-window short-time Fourier transform on the historical partial discharge signals respectively, and obtaining corresponding historical wide-window time-frequency spectrograms and historical narrow-window time-frequency spectrograms; splicing the amplitudes of the corresponding historical wide-window time-frequency spectrograms and the amplitudes of the corresponding historical narrow-window time-frequency spectrograms along the corresponding channel axis, generating a frequency band mask matrix based on a convolutional neural network and an activation function, and obtaining a corresponding frequency band mask complementary matrix based on the frequency band mask matrix; multiplying the corresponding historical wide-window time-frequency spectrograms and the corresponding frequency band mask matrix element by element to obtain a first weighted spectrogram, multiplying the corresponding historical narrow-window time-frequency spectrograms and the corresponding frequency band mask complementary matrix element by element to obtain a second weighted spectrogram, and adding the first weighted spectrogram and the second weighted spectrogram to obtain a historical fusion time-frequency graph.

[0007] As preferred, the step b comprises: b1. establishing the physically constrained generative adversarial network comprises establishing a generator network and a discriminator network; the generator network is configured to generate a synthetic time-frequency graph based on a random noise vector and the discharge type label, and the discriminator network is configured to generate a authenticity probability based on the discharge type label and the corresponding historical fusion time-frequency graph, and generate a synthetic probability based on the discharge type label and the corresponding synthetic time-frequency graph; b2. defining an adversarial loss function and a physical constraint loss function, and training the generator network and the discriminator network in combination with the historical fusion time-frequency graph and the corresponding discharge type label to obtain the trained physically constrained generative adversarial network.

[0008] As further preferred, the process of defining the adversarial loss function in the step b2 comprises: generating the random noise vector based on a preset Gaussian distribution; extracting deep features of the historical fusion time-frequency graph and the synthetic time-frequency graph based on the convolutional layers of the discriminator network, and outputting the authenticity probability and the synthetic probability corresponding to the historical fusion time-frequency graph and the synthetic time-frequency graph respectively in combination with the discharge type label input into the discriminator network; defining the adversarial loss function corresponding to the discriminator network and the generator network based on the authenticity probability and the synthetic probability respectively.

[0009] As further preferred, the process of defining the physical constraint loss function in the step b2 comprises: calculating the amplitude distribution matching loss of the amplitude distribution parameter corresponding to all historical partial discharge signals based on a divergence algorithm, calculating the phase concentration loss of the phase concentration parameter corresponding to all historical partial discharge signals based on a mean square error algorithm, and defining the physical constraint loss function based on the weighted amplitude distribution matching loss and the phase concentration loss.

[0010] As a further preferred embodiment, in step b2, the generator network and the discriminator network are trained in combination with the historical fusion time-frequency graph and the corresponding discharge type label, including: when training the discriminator network, the weight parameters of the generator network are fixed, and the total discriminator loss function is obtained based on the adversarial loss function corresponding to the discriminator network and the physical constraint loss function, the total discriminator loss function is minimized and the weight parameters of the discriminator network are updated; when training the generator network, the weight parameters of the discriminator network are fixed, and the total generator loss function is obtained based on the adversarial loss function corresponding to the generator network and the physical constraint loss function, the total generator loss function is minimized and the weight parameters of the generator network are updated; the discriminator network and the generator network are trained alternately until the total discriminator loss function and the total generator loss function reach convergence conditions, and the weight parameters corresponding to the current generator network and the discriminator network are frozen.

[0011] Preferably, step c includes: c1. inputting the fused time-frequency graph into the trained physical constraint generative adversarial network to extract deep features and output a corresponding fused feature graph; c2. slicing the fused feature graph and then performing time series encoding and time series modeling, wherein the time series modeling includes concatenating the sliced ​​fused feature graphs according to the feature dependencies of the time series encoding to obtain the time series feature vector; and c3. constructing a knowledge distillation classification network, and classifying all partial discharge signals in the knowledge distillation classification network based on the time series feature vector.

[0012] As a further preferred embodiment, the step c2 specifically includes: slicing the fused feature map along the time dimension to obtain a time slice sequence, and processing the time slice sequence separately from the first time series processing branch and the second time series processing branch through a bidirectional long short-term memory network; extracting time series dependent features from the time slice sequence in time sequence based on the first time series processing branch, and outputting them as a first feature sequence; extracting time series dependent features from the time slice sequence in time reverse order based on the second time series processing branch, and outputting them as a second feature sequence; intercepting the final time step state of the first feature sequence and splicing it with the initial time step state of the second feature sequence to obtain the time series feature vector.

[0013] As a further preferred, the step c3 specifically comprises: constructing a knowledge distillation classification network, including constructing a teacher classification network and a student classification network; the teacher classification network is configured to perform multi-dimensional space modeling based on the time sequence feature vector and obtain a first discharge type probability vector; the student classification network is configured to perform lightweight feature extraction based on the time sequence feature vector and obtain a second discharge type probability vector; taking the first discharge type probability vector as a soft supervision target, constructing a distillation loss function to measure the distribution difference between the first discharge type probability vector and the second discharge type probability vector, minimizing the distillation loss function and training the student classification network, so that the trained student classification network analyzes discharge mode features based on the time sequence feature vector to perform classification on the partial discharge signal.

[0014] As a further preferred, the distillation loss function is constructed to measure the distribution difference between the first discharge type probability vector and the second discharge type probability vector, including: calculating the divergence loss between the first discharge type probability vector and the second discharge type probability vector, extracting the time sequence attention map of the teacher network and the student network, and calculating the mean square error loss; the distillation loss function is constructed by the weighted sum of the divergence loss and the mean square error loss.

[0015] The partial discharge signal classification method based on the physical constraint generative adversarial network has the following beneficial effects: Firstly, unlike the traditional classification method based on fixed window time-frequency analysis, which leads to single time-frequency resolution, the present application dynamically fuses the time-frequency spectrum extracted by wide and narrow window functions through a double window time-frequency coding strategy, which can adaptively select the optimal resolution in different frequency bands, and can capture high-frequency resonance characteristics and narrow pulse transient details, significantly improving the accuracy of time-frequency representation and signal analysis ability.

[0016] Secondly, in view of the limitation that the traditional classification method lacks physical rationality in the sample enhancement process, the present application constructs a physical constraint generative adversarial network, introduces discharge type labels and amplitude distribution, phase concentration and other physical prior knowledge as a double supervision mechanism in the discriminator, ensures that the generated samples are highly consistent with the real partial discharge signal in statistical characteristics and physical laws, effectively avoids the misclassification problem caused by "pseudo samples", and improves the authenticity and diversity of the training data.

[0017] Thirdly, the present invention uses a multi-scale convolutional network to extract features from the fused time-frequency graphs. Combined with time series modeling techniques, this method models the temporal dependencies of partial discharge pulse trains, further exploring the temporal evolution of the signals. This, supported by physically enhanced samples, significantly improves the model's ability to recognize complex discharge signals, demonstrating enhanced robustness and generalization in the classification accuracy of rare discharge types. This overall approach effectively addresses the classification inaccuracies inherent in traditional methods, which suffer from fixed time-frequency resolution and a lack of physical rationality in sample enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a partial discharge signal classification method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be described in more detail below with reference to the accompanying drawings. It should be noted that the following description of the present invention with reference to the accompanying drawings is only illustrative and not restrictive.

[0020] Where possible, the various embodiments described below may be recombined with each other to form other embodiments not shown in the following description; the various technical features described below may also be recombined with each other to form other embodiments not shown in the following description.

[0021] Please refer to the attached Figure 1 .

[0022] Example: To address the problem of inaccurate classification caused by the fixed time-frequency resolution and lack of physical rationality in sample enhancement in traditional classification methods, this embodiment provides a partial discharge signal classification method based on a physical constraint generative adversarial network. The specific steps include: a. Extract the PD signal and historical PD signal and perform dual-window time-frequency encoding on each, resulting in a fused time-frequency map and a historical fused time-frequency map. This dual-window fusion strategy overcomes the time-frequency resolution limitations of traditional single-window functions. It adaptively selects the optimal resolution across different frequency bands, significantly improving the accuracy of the time-frequency representation and providing higher-quality input features for subsequent classification.

[0023] b. Based on historical partial discharge signals, the corresponding discharge type labels are determined and a physically constrained generative adversarial network is established. This network is trained based on the discharge type labels and the historical fusion time-frequency graphs. This step introduces physical prior knowledge, such as discharge type labels, to guide the sample generation process, avoiding the "pseudo-sample" problem that can occur in traditional data augmentation methods. This improves the diversity and authenticity of the training data, thereby enhancing the model's generalization ability and classification robustness.

[0024] c. Input the fused time-frequency graph into a trained physical constraint generative adversarial network to obtain a fused feature graph corresponding to the fused time-frequency graph. Time series modeling is performed on the fused feature graph and mapped into a time series feature vector. Based on the time series feature vector, the corresponding partial discharge signal is classified. Multi-scale feature extraction in this step enhances the model's adaptability to complex partial discharge signal morphologies, while time series modeling fully exploits the signal's evolution over time. The combination of these two significantly improves classification accuracy and stability.

[0025] The present invention breaks through the limitation of fixed time-frequency resolution in traditional classification methods through a dual-window time-frequency fusion strategy, realizes adaptive optimization of time-frequency resolution in different frequency bands, and thus significantly improves the accuracy of time-frequency representation of partial discharge signals. At the same time, it introduces a physical constraint generative adversarial network to fuse physical prior knowledge such as discharge type labels in the data enhancement process, effectively avoiding the training bias caused by "pseudo-samples" in traditional methods and enhancing the generalization ability and classification robustness of the model. On this basis, combined with multi-scale feature extraction and time series modeling technology, the complex characteristics and evolution laws of partial discharge signals in the time and frequency domains are further explored. The overall solution effectively solves the problem of inaccurate classification caused by the fixed time-frequency resolution and lack of physical rationality of sample enhancement in traditional methods.

[0026] In order to improve the time-frequency resolution matching capability of the partial discharge signal in different frequency bands and enhance the feature expression capability of the subsequent classification model, in a preferred embodiment, the process of extracting the historical partial discharge signal for dual-window time-frequency encoding in step a includes: Perform wide window short-time Fourier transform and narrow window short-time Fourier transform respectively, and obtain the historical wide window time spectrum diagram and historical narrow window time spectrum diagram accordingly: Wide window short-time Fourier transform: ; Narrow window short-time Fourier transform: . By performing wide-window short-time Fourier transform (STFT) and narrow-window short-time Fourier transform on the historical partial discharge signal respectively, a wide-window time-spectrum diagram with high frequency resolution and a narrow-window time-spectrum diagram with high time resolution are obtained, thereby retaining the key local characteristics of the signal at different scales. The amplitude of the corresponding historical wide-window time-spectrum diagram and the amplitude of the corresponding historical narrow-window time-spectrum diagram are extracted and spliced ​​along the corresponding channel axis. The amplitude information of the wide-window and narrow-window time-spectrum diagrams is extracted and spliced ​​along the channel axis to form a fused input feature, which provides comprehensive amplitude information support for subsequent mask generation and enhances the accuracy and adaptability of frequency band selection. The frequency band mask matrix is ​​generated based on the convolutional neural network and the activation function, and the corresponding frequency band mask complement matrix is ​​obtained based on the frequency band mask matrix: ; ,in is the frequency band mask matrix, It is the frequency band mask complement matrix. The frequency band mask matrix is ​​generated based on the convolutional neural network combined with the Sigmoid activation function. It can dynamically allocate fusion weights according to the energy distribution characteristics of different frequency bands, so that the fused time-frequency diagram can adaptively select the optimal resolution in different frequency bands, thereby improving the fineness and physical interpretability of the time-frequency representation. The first weighted spectrum is obtained by multiplying the corresponding wide-window time-frequency spectrum with the corresponding frequency band mask matrix element by element. The corresponding frequency band mask complement matrix is ​​further generated according to the frequency band mask matrix, which realizes the complementary weighting mechanism for the wide-window and narrow-window time-frequency diagrams, ensuring that the information of the two is evenly retained during the fusion process, avoiding information loss or redundancy. The second weighted spectrum is obtained by multiplying the corresponding narrow-window time-frequency spectrum with the corresponding frequency band mask complement matrix element by element, and the fused time-frequency diagram is obtained by adding the first weighted spectrum and the second weighted spectrum. The first weighted spectrogram is obtained by element-by-element multiplication of the wide-window time-frequency spectrum with the frequency band mask. The second weighted spectrogram is obtained by element-by-element multiplication of the narrow-window time-frequency spectrum with the frequency band complement mask. The two are then added together to obtain the final fused time-frequency spectrum. This fusion result achieves an optimal balance between time resolution and frequency resolution, significantly improving the quality of feature representation and the robustness of the classification model. The same steps as above can be used to perform dual-window time-frequency encoding on the partial discharge signal.

[0027] To improve the accuracy of partial discharge signal classification and enhance the model's robustness and generalization capabilities under different operating conditions, in a preferred embodiment, step b includes: b1. Establishing a physically constrained generative adversarial network (GAN) includes establishing a generator network and a discriminator network. The generator network is configured to generate synthetic time-frequency maps based on a random noise vector and a discharge type label, and the discriminator network is configured to generate authenticity probabilities based on the discharge type label and the corresponding historical fused time-frequency map, as well as synthetic probabilities based on the discharge type label and the corresponding synthetic time-frequency map. This step provides accurate data support for subsequent steps, ensuring that each partial discharge signal is correctly classified and identified. b2. Defining an adversarial loss function and a physically constrained loss function, and training the generator network and the discriminator network using the historical fused time-frequency maps and the corresponding discharge type labels to obtain a trained GAN. This step utilizes PC-GAN and a physically constrained loss function for training, ensuring that these maps meet the constraints of the actual physical process while generating realistic enhanced time-frequency maps. This approach not only improves the quality and authenticity of generated samples, but also, by introducing physical constraints, enhances the model's understanding of real-world application scenarios, further improving its performance and reliability in real-world operations. Furthermore, this type of model training effectively expands the training dataset, alleviating overfitting issues caused by insufficient data, and ultimately improving the model's generalization and stability.

[0028] In order to improve the physical rationality and generalization ability of the generative adversarial network in the process of local discharge signal enhancement, in a further preferred embodiment, in step b2, the process of defining the adversarial loss function includes: generating a random noise vector based on a preset Gaussian distribution, and using a Gaussian distribution to generate random noise provides a diverse input for the generator, thereby enhancing the diversity of generated samples and helping to prevent model overfitting. The convolution layer based on the discriminator network extracts the deep features of the historical fused time-frequency graph and the synthetic time-frequency graph respectively, and outputs the authenticity probability and synthesis probability corresponding to the historical fused time-frequency graph and the synthetic time-frequency graph respectively in combination with the discharge type label corresponding to the input discriminator network. Extracting deep features through the convolution layer and combining the discharge type label for authenticity assessment not only enhances the model's ability to understand local details and patterns, but also ensures that the generated time-frequency graph is closer to the actual physical phenomenon, thereby improving the realism and reliability of the generated signal. Based on the authenticity probability and synthesis probability, the adversarial loss functions corresponding to the discriminator network and the generator network are defined respectively:

[0029]

[0030] in This is a real historical fusion time-frequency diagram. is a random noise vector, is the discharge type label, is the generator network, is the discriminator network, is the training batch size. This step can effectively balance the game relationship between the discriminator and the generator, prompting the generator to generate more refined and physically meaningful partial discharge signals, while improving the generalization performance of the model on different data sets, so that the entire system can not only accurately identify and enhance known types of partial discharge signals, but also maintain high adaptability and stability for signals in unknown or changing environments. This embodiment generates a random noise vector based on a preset Gaussian distribution, and uses the convolutional layer of the discriminator network to extract the deep features of the historical fusion time-frequency graph and the synthetic time-frequency graph respectively, and defines the adversarial loss function by combining the corresponding discharge type label output authenticity probability and synthesis probability. This method significantly improves the physical rationality and generalization ability of the generative adversarial network in partial discharge signal enhancement.

[0031] To ensure the accuracy of the physical parameters, in a further preferred embodiment, in step b2, the process of defining the physical constraint loss function includes: calculating the amplitude distribution matching loss of the amplitude distribution parameters corresponding to all historical partial discharge signals based on the divergence algorithm, and performing physical constraints on the generated samples based on the two dimensions of the amplitude statistical characteristics and phase distribution law of the historical partial discharge signals, thereby avoiding the physical distortion or feature deviation problems that may occur in traditional data enhancement methods and significantly improving the physical credibility of the output samples of the generative adversarial network. The phase concentration loss of the phase concentration parameters corresponding to all partial discharge signals is calculated based on the mean square error algorithm, and the physical constraint loss function is defined based on the weighted combination of the amplitude distribution matching loss and the phase concentration loss. The weighted fusion loss design in this step enables the model to have a certain generalization ability while retaining the essential characteristics of the discharge type, thereby improving the classification accuracy while enhancing the model's adaptability and robustness to actual complex environments. In this embodiment, by introducing the amplitude distribution matching loss calculated based on the divergence algorithm and combining it with the mean square error algorithm to calculate the phase concentration loss of the phase concentration parameter.

[0032] To improve the physical consistency and sample quality of the generative adversarial network during partial discharge signal enhancement, in a further preferred embodiment, in step b2, the generator network and discriminator network are trained using historical fusion time-frequency graphs and corresponding discharge type labels. This includes: when training the discriminator network, the weight parameters of the generator network are fixed, and the total discriminator loss function is calculated based on the discriminator network's corresponding adversarial loss function and the physical constraint loss function. The total discriminator loss function is minimized and the weight parameters of the discriminator network are updated. This design, by introducing the physical constraint loss, enhances the discriminator's ability to determine the physical plausibility of generated samples, improving the stability and accuracy of model training. When training the generator network, the weight parameters of the discriminator network are fixed, and the total generator loss function is calculated based on the generator network's corresponding adversarial loss function and the physical constraint loss function. The total generator loss function is minimized and the weight parameters of the generator network are updated. By jointly optimizing the adversarial loss and the physical constraint loss, the generator ensures that the output samples conform to the physical laws of partial discharge signals while improving generation quality, effectively avoiding the problem of "pseudo-samples." The discriminator network and the generator network are trained alternately until both the total discriminator loss function and the total generator loss function reach convergence, freezing the corresponding weight parameters of the current generator network and the discriminator network. This alternating optimization strategy achieves a dynamic balance between the discriminator and the generator, ensuring that the generated samples have a high degree of authenticity and consistency in both statistical distribution and physical properties, providing high-quality enhanced data for subsequent classification tasks. This embodiment helps improve the robustness and generalization performance of the entire partial discharge signal classification system, especially in real-world engineering scenarios with small samples or severe noise interference, demonstrating greater adaptability and classification accuracy.

[0033] To improve the accuracy of partial discharge signal classification and model generalization, in a preferred embodiment, step c includes: c1. Inputting the fused time-frequency graph into a trained physically constrained generative adversarial network (PCN) to extract deep features and output the corresponding fused feature graph. This step utilizes the trained PCN for feature extraction, which not only captures the joint distribution characteristics of the PD signal in time and frequency but also preserves the high-dimensional semantic information under physical consistency constraints. This provides high-quality feature input for subsequent time series modeling, significantly improving the model's ability to recognize complex discharge patterns. c2. Slicing the fused feature graph and then performing time series encoding and time series modeling. Time series modeling involves concatenating the sliced ​​fused feature graph according to the feature dependencies of the time series encoding to obtain a time series feature vector. By slicing and time series modeling the fused feature graph, the evolution patterns and phase dependencies of the PD signal in the time dimension are fully explored, enhancing the model's ability to perceive the dynamic changes in the discharge signal. Furthermore, the concatenation method based on time series encoding effectively preserves temporal structural information, making the final time series feature vector more discriminative and generalizable. c3. A knowledge distillation classification network is constructed, and all partial discharge signals are classified within the network based on the time series feature vectors. This step, while maintaining a lightweight model, leverages the teacher network's strong classification capabilities and attention mechanism to guide the student network in learning more accurate probability distributions and internal representations, further improving classification accuracy and model robustness. This embodiment combines the KL divergence with the mean squared error of the attention graph distillation loss function design, resulting in superior performance of the student network in both output consistency and internal mechanism imitation, significantly enhancing the network's generalization and stability.

[0034] To enhance the model's ability to capture the temporal dynamics of partial discharge signals, in a further preferred embodiment, step c2 specifically includes slicing the fused feature map along the time dimension to obtain a time slice sequence. The time slice sequence is then processed separately using a bidirectional long short-term memory network (LSTM) from both the first and second temporal processing branches. The use of a bidirectional LSTM allows for simultaneous capture of temporal dependencies within the time slice sequence from both the positive and negative directions, enabling the model to not only understand the temporal trends of partial discharge signals but also identify the impact of past events on the future, thereby enhancing its ability to capture the temporal dynamics of partial discharge signals. The first temporal processing branch extracts temporal dependency features from the time slice sequence in temporal order and outputs them as a first feature sequence. The second temporal processing branch extracts temporal dependency features from the time slice sequence in temporal reverse order and outputs them as a second feature sequence. By concatenating the final time step state of the first feature sequence with the initial time step state of the second feature sequence, the temporal order and reverse order information of the partial discharge signal can be comprehensively considered, constructing a richer and more comprehensive feature representation that helps the model more accurately identify different types of partial discharge patterns. The final time step state of the first feature sequence is intercepted and concatenated with the initial time step state of the second feature sequence to obtain a time series feature vector. This bidirectional processing effectively captures long-term dependencies in the signal, which is particularly important in situations with complex background noise or large differences between signals. It helps improve the model's robustness and generalization ability, allowing it to maintain high classification accuracy even when faced with new, unseen data.

[0035] To improve classification performance and inference efficiency after model compression, in a further preferred embodiment, step c3 specifically includes constructing a knowledge distillation classification network, including a teacher classification network and a student classification network. The teacher classification network is configured to perform multidimensional spatial modeling based on the time series feature vector and obtain a first discharge type probability vector. Through complex multi-level feature learning, the teacher network is able to capture the subtle and complex pattern characteristics of the partial discharge signal and provide high-quality soft labels as supervision information. The student classification network is configured to perform lightweight feature extraction based on the time series feature vector and obtain a second discharge type probability vector. This design aims to maintain classification accuracy while significantly reducing computational cost and model size, thereby improving inference efficiency. Using the first discharge type probability vector as the soft supervision target, a distillation loss function is constructed to measure the distribution difference between the first and second discharge type probability vectors. This not only allows the student network to learn the knowledge of the teacher network but also smooths the decision boundary through soft labels, enhancing the model's generalization ability. The distillation loss function is minimized and the student classification network is trained so that the trained student classification network parses the discharge pattern characteristics based on the time series feature vector to classify the partial discharge signal. Ultimately, a highly efficient and accurate partial discharge signal classification system was implemented, capable of quickly and accurately identifying different types of discharge phenomena in resource-constrained environments. This embodiment, through the aforementioned implementation, ensures that the model has good classification performance and response speed, even in practical application scenarios with limited computing resources, while maintaining a high level of accuracy, meeting the needs of real-time monitoring and fault diagnosis.

[0036] In order to improve the learning accuracy and generalization ability of the student network in the knowledge distillation process, in a further preferred embodiment, a distillation loss function is constructed to measure the distribution difference between the first discharge type probability vector and the second discharge type probability vector, including: calculating the divergence loss between the first discharge type probability vector and the second discharge type probability vector. This step introduces divergence loss by modeling the distribution difference between the first discharge type probability vector output by the teacher network and the second discharge type probability vector output by the student network, which can more accurately align the probability distribution of the two on the discharge type and enhance classification consistency. The temporal attention map of the teacher network and the student network is extracted, and the mean square error loss is calculated. This step further extracts the temporal attention map of the teacher network and the student network and calculates its mean square error loss, so that the student network not only learns the distribution of the output label, but also imitates the attention mechanism of the teacher network in the time dimension, thereby improving its understanding and expression of the temporal characteristics of the partial discharge signal. The distillation loss function is constructed with the weighted sum of the divergence loss and the mean square error loss. This step constructs a distillation loss function by weightedly fusing the two aforementioned losses. This guides the student network's learning process from both the output distribution and internal representation levels, significantly enhancing the model's generalization and robustness while reducing model complexity and maintaining high classification accuracy. This embodiment effectively improves the student network's learning accuracy and generalization ability during the knowledge distillation process by constructing a distillation loss function that incorporates both divergence loss and mean squared error loss.

[0037] It should be understood that the embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A partial discharge signal classification method based on a physical constraint generative adversarial network, characterized in that: include: a. Extract the PD signal and historical PD signal and perform dual-window time-frequency encoding on them, respectively, to obtain a fused time-frequency graph and a historical fused time-frequency graph; b. determining a corresponding discharge type label based on the historical partial discharge signal, establishing a physical constraint generative adversarial network, and training the physical constraint generative adversarial network based on the discharge type label and the historical fused time-frequency graph; c. Input the fused time-frequency graph into the trained physical constraint generative adversarial network to obtain a fused feature graph corresponding to the fused time-frequency graph, perform time series modeling on the fused feature graph and map it into a time series feature vector, and classify the corresponding partial discharge signal based on the time series feature vector.

2. The partial discharge signal classification method according to claim 1, characterized in that: In step a, the process of extracting the historical partial discharge signal and performing dual-window time-frequency coding includes: Performing wide-window short-time Fourier transform and narrow-window short-time Fourier transform on the historical partial discharge signal, and obtaining a historical wide-window time-frequency spectrum diagram and a historical narrow-window time-frequency spectrum diagram accordingly; After extracting the amplitude of the corresponding historical wide window spectrum graph and the amplitude of the corresponding historical narrow window spectrum graph along the corresponding channel axis, a frequency band mask matrix is ​​generated based on the convolutional neural network and the activation function, and the corresponding frequency band mask complement matrix is ​​obtained based on the frequency band mask matrix; The historical wide-window time-frequency spectrum is multiplied element-by-element by the corresponding frequency band mask matrix to obtain a first weighted spectrum, the historical narrow-window time-frequency spectrum is multiplied element-by-element by the corresponding frequency band mask complement matrix to obtain a second weighted spectrum, and the first weighted spectrum and the second weighted spectrum are added to obtain a historical fused time-frequency spectrum.

3. The partial discharge signal classification method according to claim 1, characterized in that: The step b comprises: b1. Establishing the physical constraint generative adversarial network includes establishing a generator network and a discriminator network; the generator network is configured to generate a synthetic time-frequency map based on a random noise vector and the discharge type label, and the discriminator network is configured to generate a authenticity probability based on the discharge type label and the corresponding historical fused time-frequency map, and to generate a synthesis probability based on the discharge type label and the corresponding synthetic time-frequency map; b2. Define an adversarial loss function and a physical constraint loss function, and train the generator network and the discriminator network using the historical fusion time-frequency graph and the corresponding discharge type labels to obtain the trained physical constraint generative adversarial network.

4. The partial discharge signal classification method according to claim 3, characterized in that: The process of defining the adversarial loss function in step b2 includes: The random noise vector is generated based on a preset Gaussian distribution; the deep features of the historical fused time-frequency graph and the synthetic time-frequency graph are respectively extracted based on the convolutional layer of the discriminator network, and the authenticity probability and synthesis probability corresponding to the historical fused time-frequency graph and the synthetic time-frequency graph are respectively output in combination with the discharge type label corresponding to the input of the discriminator network; the adversarial loss functions corresponding to the discriminator network and the generator network are respectively defined based on the authenticity probability and the synthesis probability.

5. The partial discharge signal classification method according to claim 4, characterized in that: The process of defining the physical constraint loss function in step b2 includes: The amplitude distribution matching loss of the amplitude distribution parameters corresponding to all historical partial discharge signals is calculated based on the divergence algorithm, the phase concentration loss of the phase concentration parameters corresponding to all historical partial discharge signals is calculated based on the mean square error algorithm, and the physical constraint loss function is defined based on the weighted combination of the amplitude distribution matching loss and the phase concentration loss.

6. The partial discharge signal classification method according to claim 5, characterized in that: In step b2, the training of the generator network and the discriminator network by combining the historical fusion time-frequency graph and the corresponding discharge type labels includes: When training the discriminator network, the weight parameters of the generator network are fixed, and a total discriminator loss function is obtained based on the adversarial loss function corresponding to the discriminator network and the physical constraint loss function, the total discriminator loss function is minimized, and the weight parameters of the discriminator network are updated; When training the generator network, the weight parameters of the discriminator network are fixed, and a total generator loss function is obtained based on the adversarial loss function corresponding to the generator network and the physical constraint loss function, the total generator loss function is minimized, and the weight parameters of the generator network are updated; The discriminator network and the generator network are trained alternately until both the total loss function of the discriminator and the total loss function of the generator reach convergence conditions, and the weight parameters corresponding to the current generator network and the discriminator network are frozen.

7. The partial discharge signal classification method according to claims 1 to 6, characterized in that: The step c comprises: c1. Input the fused time-frequency graph into the trained physical constraint generative adversarial network, extract deep features and output the corresponding fused feature graph; c2. performing time series coding and time series modeling on the fused feature map after slicing, wherein the time series modeling includes concatenating the sliced ​​fused feature map according to the feature dependency of the time series coding to obtain the time series feature vector; c3. Constructing a knowledge distillation classification network, and classifying all partial discharge signals based on the time series feature vectors in the knowledge distillation classification network.

8. The partial discharge signal classification method according to claim 7, characterized in that: The step c2 specifically includes: Slice the fused feature map along the time dimension to obtain a time slice sequence, and process the time slice sequence respectively from the first time series processing branch and the second time series processing branch through a bidirectional long short-term memory network; Extracting time-series dependency features from the time slice sequence in time order based on the first time-series processing branch, and outputting the features as a first feature sequence; extracting time-series dependency features from the time slice sequence in time reverse order based on the second time-series processing branch, and outputting the features as a second feature sequence; The final time step state of the first feature sequence is intercepted and concatenated with the initial time step state of the second feature sequence to obtain the time series feature vector.

9. The partial discharge signal classification method according to claim 7, characterized in that: The step c3 specifically includes: Construct a knowledge distillation classification network, including building a teacher classification network and a student classification network; The teacher classification network is configured to perform multidimensional space modeling based on the time series feature vector and obtain a first discharge type probability vector; the student classification network is configured to perform lightweight feature extraction based on the time series feature vector and obtain a second discharge type probability vector; Taking the first discharge type probability vector as a soft supervision target, a distillation loss function is constructed to measure the distribution difference between the first discharge type probability vector and the second discharge type probability vector, the distillation loss function is minimized and the student classification network is trained so that the trained student classification network parses the discharge pattern features based on the time series feature vector to perform classification on the partial discharge signal.

10. The partial discharge signal classification method according to claim 9, characterized in that: Constructing a distillation loss function to measure the distribution difference between the first discharge type probability vector and the second discharge type probability vector includes: Calculate the divergence loss between the first discharge type probability vector and the second discharge type probability vector, extract the temporal attention map of the teacher network and the student network, and calculate the mean square error loss; construct the distillation loss function as a weighted sum of the divergence loss and the mean square error loss.

Images