A method for generating generator partial discharge samples by physical constraint and reinforcement learning
By introducing multidimensional physical feature constraints and reinforcement learning into generative adversarial networks, partial discharge samples of generators that conform to physical laws are generated, solving the problem of sample inconsistency in existing technologies and improving the diagnostic performance of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-14
AI Technical Summary
Existing methods for generating generator partial discharge samples based on generative adversarial networks (GANs) lack physical consistency, resulting in generated samples that do not conform to the basic physical laws of partial discharge patterns, thus affecting the generalization ability and reliability of the model.
By constructing a multidimensional physical feature constraint space and introducing physical consistency as a reinforcement learning reward feedback into the training process of the generative adversarial network, the samples generated by the generator must simultaneously satisfy the discriminator's true/false discrimination criteria and the conditions of the multidimensional physical feature constraint space. Reinforcement learning algorithms are then used to optimize the generator parameters.
The generated simulated PRPD samples have the same physical and statistical properties as the real samples, which solves the problem of scarce and unevenly distributed generator stator insulation fault samples and improves the generalization ability and accuracy of the diagnostic model.
Smart Images
Figure CN122065036B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for generating partial discharge samples of generators using physical constraints and reinforcement learning, belonging to the field of high-voltage equipment condition assessment and artificial intelligence data augmentation technology. Background Technology
[0002] Generators are core equipment in power systems, and the assessment of their stator insulation condition directly affects the safe and stable operation of the unit. Partial discharge (PD) detection is an important means of assessing the stator insulation condition of generators. With the development of artificial intelligence technology, deep learning-based PD pattern recognition methods have been widely studied and have demonstrated superior performance in the field of generator PD pattern recognition.
[0003] In practical engineering applications, the training of deep learning models relies heavily on a large amount of high-quality labeled sample data. As valuable equipment, generators face the following challenges in obtaining partial discharge samples: (1) In actual operation, the frequency of partial discharge in generators is limited, and it is difficult to cover all defect types; (2) Simulating various defects through experimental means is costly and time-consuming, resulting in small-scale and unbalanced partial discharge sample data, constituting a typical small sample problem. Based on this, using generative adversarial networks to expand samples has become an effective technical approach.
[0004] Existing data augmentation methods based on Generative Adversarial Networks (GANs) can generate visually realistic PRPD samples, but they often lack physical consistency. That is, the generated samples may violate the basic physical laws of partial discharge pattern distribution, such as phase distribution misalignment and missing polarity effects. Using such samples without physical characteristics to train pattern recognition models not only fails to improve model performance, but may also introduce noise, reducing the model's generalization ability and reliability.
[0005] Therefore, how to generate high-quality augmented samples that both conform to the statistical distribution of images and strictly follow the physical mechanism of partial discharge is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a generator partial discharge sample generation method based on physical constraints and reinforcement learning. By constructing a multi-dimensional physical feature constraint space and introducing physical consistency as a reinforcement learning reward feedback into the training process of the generative adversarial network, high-quality generator partial discharge augmentation samples with physical consistency are generated.
[0007] The objective of this invention is achieved by the following technical solution:
[0008] A method for generating generator partial discharge samples using physical constraints and reinforcement learning includes the following steps:
[0009] S1: Obtain the original partial discharge pulse signal of the generator stator insulation and convert it into a real PRPD sample with phase-amplitude-number of discharge pulses to form the original real sample set;
[0010] S2: Perform statistical analysis on the original real sample set, extract key physical features of real PRPD samples that reflect the physical mechanism of partial discharge, and construct a multi-dimensional physical feature constraint space based on the statistical distribution of key physical features of real PRPD samples.
[0011] S3: Construct a generative adversarial network, which includes a generator for generating simulated PRPD samples, a discriminator for distinguishing between genuine and fake graphs, and a physical verification module for calculating the physical feature deviation of simulated PRPD samples in a multidimensional physical feature constraint space.
[0012] S4: Construct a reinforcement learning training environment, model the generator's generation and training process as a reinforcement learning policy, and regard the generator's generation of simulated PRPD samples as actions; construct a comprehensive reward function, which is a weighted fusion of the true / false probability reward output by the discriminator, the physical feature deviation penalty output by the physical verification module, and the hard constraint penalty.
[0013] S5: During the training iteration, the reinforcement learning algorithm is used to calculate the gradient and update the generator parameters according to the comprehensive reward function until the simulated PRPD samples generated by the generator simultaneously satisfy the discriminator's true / false discrimination criteria and the constraint conditions of the multidimensional physical feature constraint space, thus obtaining a converged generator.
[0014] S6: Using a convergent generator, generate batches of simulated PRPD samples with generator partial discharge extension.
[0015] Furthermore, the specific steps of S1 include:
[0016] S1.1: Use rated voltage to obtain the partial discharge signal of typical defects in generator stator bars. The partial discharge signal is synchronously phase triggered by the applied voltage, and the single sampling period is 20ms of the power frequency period.
[0017] S1.2: Power frequency cycle The discharge amplitude range is divided into M phase windows, linearly normalized to [0,1], and divided into N amplitude levels; M is the number of phase windows, and N is the number of amplitude levels;
[0018] S1.3: Count the number of pulses falling within the i-th phase window and the j-th amplitude level within a set sampling period, constructing a dimension of... The matrix X forms the real PRPD sample.
[0019] Furthermore, in step S2, the key physical characteristics of the real PRPD sample include at least the discharge repetition rate polarity ratio characteristic, phase distribution characteristic, and amplitude asymmetry characteristic.
[0020] The polarity ratio characteristic of the discharge repetition rate The calculation is shown in formula (1);
[0021]
[0022] in, This represents the average number of discharge pulses during the positive half-cycle in the PRPD sample. This represents the average number of discharge pulses during the negative half-cycle in the PRPD sample.
[0023] The phase distribution characteristics include the centroid and phase dispersion of the phase distribution of the partial discharge pulse cluster during the positive and negative half-cycles of the power frequency; the centroid of the phase distribution during the positive half-cycle. The calculation is shown in formula (2), where the centroid of the negative half-cycle phase distribution is... The positive half-cycle phase dispersion is calculated as shown in formula (3). The negative half-cycle phase dispersion is calculated as shown in formula (4). The calculation is shown in formula (5);
[0024]
[0025]
[0026] in, This represents the number of pulses counted within the i-th phase window and the j-th amplitude level of the corresponding matrix X. The center phase of the i-th phase window;
[0027]
[0028]
[0029] The amplitude asymmetry feature The calculation is shown in formula (6);
[0030]
[0031] in, This represents the maximum discharge amplitude during the positive half-cycle. This represents the maximum discharge amplitude during the negative half-cycle.
[0032] The multidimensional physical feature constraint space in S2 is constructed based on the mean and covariance matrix of the feature vectors of the original real sample set, and Mahalanobis distance is used as the metric.
[0033] Furthermore, step S3 specifically includes:
[0034] S3.1: The generator adopts a deep neural network architecture including an encoder, a decoder, and skip connections; the input of the generator is a random noise vector that follows a normal distribution, and the output is a simulated PRPD sample with the same dimension as the original real PRPD sample;
[0035] S3.2: The discriminator adopts a convolutional neural network architecture, with PRPD samples as input and the output being the probability value that the input PRPD sample is a real PRPD sample;
[0036] S3.3: The physical verification module is embedded between the generator and the discriminator. It is used to receive simulated PRPD samples generated by the generator, extract the physical feature vector of the simulated PRPD samples, calculate the Mahalanobis distance of the physical feature vector relative to the multidimensional physical feature constraint space, and normalize it using the Sigmoid function to map it to the physical feature deviation in the range [0,1]. Meanwhile, the physical verification module is also used to detect hard constraint violations, which include: the difference between the positive and negative half-cycle phase centroids exceeding the preset physical reasonable range, the generated simulated PRPD sample being inconsistent with the unipolar mode of the preset defect type, and the cumulative discharge amplitude in the zero-crossing region exceeding the preset safety threshold; if the physical verification module detects a violation, it outputs a penalty flag.
[0037] Furthermore, in step S4, the comprehensive reward function R is constructed as shown in formula (7);
[0038]
[0039] in, , This represents the probability value of the true PRPD sample output by the discriminator, where D is the discriminator function, G is the generator function, and z is the random noise vector. This represents the logarithmic reward corresponding to the probability value; Indicates deviation in physical characteristics; As a hard constraint penalty, when the physical verification module outputs a penalty flag... The default penalty constant is used; otherwise, it is 0. , , These are the weighting coefficients.
[0040] Further, step S5 specifically includes:
[0041] S5.1: Initialize generator and discriminator parameters;
[0042] S5.2: For a single iteration, sample B random noise vectors from the noise distribution and generate B simulated PRPD samples using the current generator;
[0043] S5.3: Use a discriminator to distinguish between real PRPD samples and generated simulated PRPD samples, and update the discriminator parameters to minimize the cross-entropy loss;
[0044] S5.4: Calculate the physical feature deviation and hard constraint penalty of each generated simulated PRPD sample using the physical verification module, and calculate the comprehensive reward value of the simulated PRPD sample by combining the discriminator output;
[0045] S5.5: The REINFORCE algorithm, a reinforcement learning algorithm based on policy gradients, is used to calculate the gradient and update the generator parameters. The calculation is shown in formula (8), and the parameter update is shown in formula (9);
[0046]
[0047]
[0048] in, Generator parameters The gradient operator, i.e., taking the partial derivative with respect to the parameters; The generator is represented by the b-th random noise vector. Generate simulated PRPD samples Generation strategy; The reward for the current sample. This is the historical reward baseline value. Here, B is the generator parameter, and B is the number of samples. This represents the learning rate of the generator;
[0049] S5.6: Repeat the above S5.2-S5.5 process until the convergence condition is met. The convergence condition is: the discriminator probability values of the simulated PRPD samples all exceed the first preset threshold, the physical feature deviations are all lower than the second preset threshold, and the hard constraint penalty is 0. The first preset threshold is used to measure the realism of the generated simulated PRPD samples at the visual / statistical distribution level, and the second preset threshold is used to measure whether the physical characteristics of the generated simulated PRPD samples conform to the statistical laws of real PRPD samples.
[0050] This invention embeds the physical and statistical laws of dielectric discharge as physical constraints in reinforcement learning into a generative adversarial network, enabling the generated simulated PRPD samples to possess the same physical and statistical characteristics as real PRPD samples. This effectively solves the problem of scarce and unevenly distributed generator stator insulation fault samples, providing a high-quality data foundation for improving the generalization ability and accuracy of diagnostic models. Attached Figure Description
[0051] Figure 1 This is a schematic diagram illustrating the process of the present invention.
[0052] Figure 2 This is a schematic diagram illustrating the process of the present invention. Detailed Implementation
[0053] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. Contents not described in detail herein are prior art and structures known to those skilled in the art. It is understandable that certain well-known structures and their descriptions may be omitted from the drawings.
[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0055] like Figure 1 , Figure 2 As shown in this embodiment, a generator partial discharge sample generation method based on physical constraints and reinforcement learning includes the following steps:
[0056] S1: Obtain the original partial discharge pulse signal of the generator stator insulation and convert it into a real PRPD sample with phase-amplitude-number of discharge pulses to form the original real sample set; the specific steps are as follows:
[0057] S1.1: The stator bars of the generator containing typical defects are pressurized with rated voltage, and the partial discharge pulse signal is obtained by using a 2nF high-voltage coupling capacitor. The sampling rate of the acquisition system is 125MS / s and the amplitude resolution is 14 bits. The power frequency phase signal is used as the synchronous trigger source and the single sampling period is 20ms.
[0058] S1.2: Power frequency cycle Divided into M=180 phase windows, each phase window is The discharge amplitude range Linearly normalized to [0,1] and divided into N=180 amplitude levels;
[0059] S1.3: Statistically analyze data falling within the specified sampling period of 50 periods. i The first phase window and the first j Number of pulses within each amplitude level The construction dimension is A matrix X, where elements The statistical method is shown in formula (10); the values in the matrix are linearly normalized to [0,1] to form real PRPD samples; multiple real PRPD samples form the original real sample set;
[0060] ；
[0061] Where K is the set sampling period, and in this example K=50; This represents the total number of pulses falling within the i-th phase window and the j-th amplitude level during the t-th sampling period; Represents the matrix of the first j Line 1 i The values in the column.
[0062] S2: Perform statistical analysis on the original real sample set to extract key physical features of real PRPD samples that reflect the physical mechanism of partial discharge. Construct a multi-dimensional physical feature constraint space based on the statistical distribution of the key physical features of real PRPD samples. The key physical features of real PRPD samples include at least the discharge repetition rate polarity ratio feature, phase distribution feature, and amplitude asymmetry feature.
[0063] The polarity ratio characteristic of the discharge repetition rate The calculation is shown in formula (1);
[0064]
[0065] in, This represents the average number of discharge pulses during the positive half-cycle in the PRPD sample. This represents the average number of discharge pulses during the negative half-cycle in the PRPD sample.
[0066] The phase distribution characteristics include the centroid and phase dispersion of the phase distribution of the partial discharge pulse cluster during the positive and negative half-cycles of the power frequency; the centroid of the phase distribution during the positive half-cycle. The calculation is shown in formula (2), where the centroid of the negative half-cycle phase distribution is... The positive half-cycle phase dispersion is calculated as shown in formula (3). The negative half-cycle phase dispersion is calculated as shown in formula (4). The calculation is shown in formula (5);
[0067]
[0068]
[0069] in, This represents the number of pulses counted within the i-th phase window and the j-th amplitude level of the corresponding matrix X. The center phase of the i-th phase window;
[0070]
[0071]
[0072] The amplitude asymmetry feature The calculation is shown in formula (6);
[0073]
[0074] in, This represents the maximum discharge amplitude during the positive half-cycle. This represents the maximum discharge amplitude during the negative half-cycle.
[0075] Extract the above features to form a feature vector Calculate the mean of the feature vectors of all samples in the original sample set. Covariance Matrix A multidimensional physical feature constraint space is constructed, and Mahalanobis distance is used as the metric.
[0076] S3: Construct a Generative Adversarial Network (GAN). The GAN includes a generator for generating simulated PRPD samples, a discriminator for distinguishing genuine from fake graphs, and a physical verification module for calculating the physical feature deviations of simulated PRPD samples within a multidimensional physical feature constraint space. The specific steps are as follows:
[0077] S3.1: The generator employs a neural network architecture based on depthwise transposed convolution and residual skip connections. The input is a 128-dimensional random noise vector following a standard normal distribution, and the output is a resolution of [resolution value missing]. The simulated PRPD samples; where: (1) the random noise vector is first projected and reshaped by a fully connected layer into a sample with 512 channels and a size of (2) The initial feature tensor; (3) Feature decoding is performed through a four-level cascaded upsampling module. Each level of the module uses a transposed convolutional layer with a stride of 2, 3, 3, 2, and the feature map size is determined by the following steps. The path gradually expands, with convolutional layers using stride 2. The convolutional kernel, the convolutional layer with a stride of 3 uses The convolution kernels are configured with batch normalization and ReLU activation functions in each layer. The residual modules with skip connections are embedded in each upsampling stage to preserve the high-frequency edge features of the partial discharge pulse; (3) The output pixel values are constrained to the [0,1] range by the Sigmoid activation function to form a simulated PRPD sample with the same dimension as the real PRPD sample.
[0078] S3.2: The discriminator adopts a deep convolutional neural network architecture, and its input is a dimension. The PRPD samples are output as the probability value that the input PRPD samples are real PRPD samples; where: (1) the input map is feature extracted through a four-level cascaded downsampling module, and each module passes through a convolutional layer with a stride of 2, 3, 3, 2 to extract the feature dimensions of the input map. The path is gradually compressed, with the convolutional layer using a stride of 2. The convolutional kernel, the convolutional layer with a stride of 3 uses The convolution kernels are used, and after each convolution layer, a Leaky ReLU activation function and a Dropout layer (dropout rate 0.25) are used to extract high-order semantic features and prevent overfitting; (2) the convolution kernels are used to extract high-order semantic features and prevent overfitting. The feature map is unfolded and mapped to a scalar through a fully connected layer. The Sigmoid activation function outputs a confidence value in the range [0, 1], which is used to represent the probability that the input sample is a real generator partial discharge pattern.
[0079] S3.3: The physical verification module is embedded between the generator and the discriminator. It is used to receive simulated PRPD samples generated by the generator, extract the physical feature vector of the simulated PRPD samples, calculate the Mahalanobis distance of the physical feature vector relative to the multidimensional physical feature constraint space, and normalize it using the Sigmoid function to map it to the physical feature deviation in the range [0,1]. Simultaneously, the physical verification module is also used to detect hard constraint violations, including: the difference between the positive and negative half-cycle phase centroids exceeding a preset physical reasonable range, inconsistencies between the generated simulated PRPD sample and the unipolar mode of a preset defect type, and the cumulative discharge amplitude exceeding a preset safety threshold within the zero-crossing region. If the physical verification module detects a violation, it outputs a penalty flag. Its main function is to process input dimensions... The PRPD samples were subjected to physical consistency testing; specifically: (1) the physical verification module was based on the characteristics of the discharge repetition rate polarity ratio in S2. Centroid of positive half-cycle phase distribution Centroid of negative half-cycle phase distribution Positive half-cycle phase dispersion Negative half-cycle phase dispersion Amplitude asymmetry characteristics Formulas (1)-(6) are used to calculate the key eigenvectors of the input matrix X. ; Mean of feature vectors based on real samples Covariance Matrix Using the formula Calculate the Mahalanobis distance of the generated samples relative to the true physical distribution, and normalize it using the Sigmoid function to map it to the physical feature deviation within the range [0,1]. Then, the feedback is given to the comprehensive reward function, where T is the transpose operator; (2) Based on the hard constraint violation of the three generator insulation defect mechanism, simulate PRPD sample judgment is performed, and the penalty flag is output for the violation. Hard constraint violations include: (1) detecting the difference between the centroids of the positive and negative half-cycle phases, with the following judgment condition: That is, the allowable deviation range is If it exceeds this range, it is considered a violation; (2) For the bipolar distribution defect type, detect whether there is a unipolar missing, and the judgment condition is: If it exceeds this range, it is judged as a violation; (3) Detect the background noise in the zero-crossing area of the power frequency voltage. The zero-crossing area mainly includes , and If the average amplitude within the zero-crossing region exceeds 20% of the overall average amplitude, it is considered a violation.
[0080] S4: Construct a reinforcement learning training environment, model the generator's generation and training process as a reinforcement learning policy, and treat the generator's generation of simulated PRPD samples as actions; construct a comprehensive reward function, which is a weighted fusion of the true / false probability reward output by the discriminator, the physical feature bias penalty output by the physical verification module, and the hard constraint penalty; the reinforcement learning elements are defined as follows:
[0081] Intelligent Agent: The generator acts as the intelligent agent, continuously optimizing its own parameters. Generate simulated PRPD samples that maximize the overall reward signal;
[0082] Environment: Consists of a discriminator, a physical verification module, and a data distribution of real PRPD samples. It receives samples generated by the agent and feeds back a comprehensive reward value.
[0083] State: The random noise vector input to the generator. The state space is a 128-dimensional standard normal distribution space. Each randomly sampled noise vector represents the initial state of a normal generation task.
[0084] Action: The generator generates a dimension based on the current state. Simulated PRPD samples with values in the range [0,1];
[0085] Strategy: The strategy is directly determined by the generator parameters. The decision is that the training process is the process of finding the optimal policy parameters through the policy gradient algorithm;
[0086] Reward: The comprehensive reward function R is constructed as shown in formula (7);
[0087]
[0088] in, , This represents the probability value of the true PRPD sample output by the discriminator, where D is the discriminator function, G is the generator function, and z is the random noise vector. This represents the logarithmic reward corresponding to the probability value; Indicates deviation in physical characteristics; As a hard constraint penalty, when the physical verification module outputs a penalty flag... The default penalty constant is used; otherwise, it is 0. , , These are the weighting coefficients; in this embodiment =0.5、 =0.2、 =0.3.
[0089] S5: During the training iteration, a reinforcement learning algorithm is used to calculate the gradient based on the comprehensive reward function and update the generator parameters until the simulated PRPD samples generated by the generator simultaneously satisfy the discriminator's true / false discrimination criteria and the constraints of the multi-dimensional physical feature constraint space, thus obtaining a converged generator; the specific steps are as follows:
[0090] S5.1: Initialize generator parameters, discriminator parameters, initialize experience replay buffer to store historical reward values, set batch size B=64, learning rate settings: generator learning rate is 0.0001, discriminator learning rate is 0.0001, and optimizer is Adam for both.
[0091] S5.2: Sample B random noise vectors from the noise distribution. Generate B simulated PRPD samples using the current generator.
[0092] S5.3: Randomly sample B real PRPD samples from the real PRPD sample set. Calculate the discriminator's loss on real PRPD samples. Loss on generating simulated PRPD samples Calculate total loss Backpropagation updates the discriminator parameters; where For the discriminator function, This is the b-th real PRPD sample; This is the b-th simulated PRPD sample;
[0093] S5.4: Calculate the physical characteristic deviation of each generated simulated PRPD sample using the physical verification module. and hard constraints and penalties Combined with the discriminator output, the comprehensive reward value of each simulated PRPD sample is calculated based on formula (7);
[0094] S5.5: The REINFORCE algorithm, a reinforcement learning algorithm based on policy gradients, is used to calculate the gradient and update the generator parameters. The calculation is shown in formula (8), and the parameter update is shown in formula (9);
[0095]
[0096]
[0097] in, Generator parameters The gradient operator, i.e., taking the partial derivative with respect to the parameters; The generator is represented by the b-th random noise vector. Generate simulated PRPD samples Generation strategy; The reward for the current sample. This is the historical reward baseline value. Here, B is the generator parameter, and B is the number of samples. Indicates the learning rate of the generator; Updated to the exponential moving average of the current batch of rewards;
[0098] S5.6: Repeat the above S5.2-S5.5 process until the convergence condition is met. The convergence condition is: the discriminator probability value of the simulated PRPD sample exceeds the first preset threshold of 0.8, that is, the discriminator output probability is stable at around 0.8; the physical feature deviation is lower than the second preset threshold of 0.2; the hard constraint penalty is 0; stop training and save the trained generator parameters.
[0099] S6: Input a normally distributed random noise vector into the trained convergent generator and output batches of generator partial discharge augmentation samples with physical consistency for subsequent training of the diagnostic model.
Claims
1. A method for generating generator partial discharge samples using physical constraints and reinforcement learning, characterized in that, Includes the following steps: S1: Obtain the original partial discharge pulse signal of the generator stator insulation and convert it into a real PRPD sample with phase-amplitude-number of discharge pulses to form the original real sample set; S2: Perform statistical analysis on the original real sample set, extract key physical features of real PRPD samples that reflect the physical mechanism of partial discharge, and construct a multi-dimensional physical feature constraint space based on the statistical distribution of key physical features of real PRPD samples. The multidimensional physical feature constraint space is constructed based on the mean and covariance matrix of the physical feature vectors of the original real sample set, and Mahalanobis distance is used as the metric. S3: Construct a Generative Adversarial Network (GAN), which includes a generator for generating simulated PRPD samples, a discriminator for distinguishing genuine from fake graphs, and a physical verification module for calculating the physical feature deviations of simulated PRPD samples within a multidimensional physical feature constraint space; specifically including: S3.1: The generator adopts a deep neural network architecture including an encoder, a decoder, and skip connections; the input of the generator is a random noise vector that follows a normal distribution, and the output is a simulated PRPD sample with the same dimension as the original real PRPD sample; S3.2: The discriminator adopts a convolutional neural network architecture, with PRPD samples as input and the output being the probability value that the input PRPD sample is a real PRPD sample; S3.3: The physical verification module is embedded between the generator and the discriminator. It is used to receive simulated PRPD samples generated by the generator, extract the physical feature vector of the simulated PRPD samples, calculate the Mahalanobis distance of the physical feature vector relative to the multidimensional physical feature constraint space, and normalize it using the Sigmoid function to map it to the physical feature deviation in the range [0,1]. Meanwhile, the physical verification module is also used to detect hard constraint violations, which include: the difference between the positive and negative half-cycle phase centroids exceeding the preset physical reasonable range, the generated simulated PRPD sample being inconsistent with the unipolar mode of the preset defect type, and the cumulative discharge amplitude in the zero-crossing region exceeding the preset safety threshold; if the physical verification module detects a violation, it outputs a penalty flag. S4: Construct a reinforcement learning training environment, model the generator's generation and training process as a reinforcement learning policy, and regard the generator's generation of simulated PRPD samples as actions; construct a comprehensive reward function, which is a weighted fusion of the true / false probability reward output by the discriminator, the physical feature deviation penalty output by the physical verification module, and the hard constraint penalty. S5: During the training iteration, the reinforcement learning algorithm is used to calculate the gradient and update the generator parameters according to the comprehensive reward function until the simulated PRPD samples generated by the generator simultaneously satisfy the discriminator's true / false discrimination criteria and the constraint conditions of the multidimensional physical feature constraint space, thus obtaining a converged generator. S6: Using a convergent generator, generate batches of simulated PRPD samples with generator partial discharge extension.
2. The generator partial discharge sample generation method based on physical constraints and reinforcement learning according to claim 1, characterized in that, S1 specifically includes: S1.1: Use rated voltage to obtain the partial discharge signal of typical defects in generator stator bars. The partial discharge signal is synchronously phase triggered by the applied voltage, and the single sampling period is 20ms of the power frequency period. S1.2: Power frequency cycle The discharge amplitude range is divided into M phase windows, linearly normalized to [0,1], and divided into N amplitude levels; M is the number of phase windows, and N is the number of amplitude levels; S1.3: Count the number of pulses falling within the i-th phase window and the j-th amplitude level within a set sampling period, constructing a dimension of... The matrix X forms the real PRPD sample.
3. The generator partial discharge sample generation method based on physical constraints and reinforcement learning according to claim 2, characterized in that, The key physical characteristics of the real PRPD sample in S2 include at least the discharge repetition rate polarity ratio characteristic, phase distribution characteristic, and amplitude asymmetry characteristic; The polarity ratio characteristic of the discharge repetition rate The calculation is shown in formula (1); in, This represents the average number of discharge pulses during the positive half-cycle in the PRPD sample. This represents the average number of discharge pulses during the negative half-cycle in the PRPD sample. The phase distribution characteristics include the centroid and phase dispersion of the phase distribution of the partial discharge pulse cluster during the positive and negative half-cycles of the power frequency; the centroid of the phase distribution during the positive half-cycle. The calculation is shown in formula (2), where the centroid of the negative half-cycle phase distribution is... The positive half-cycle phase dispersion is calculated as shown in formula (3). The negative half-cycle phase dispersion is calculated as shown in formula (4). The calculation is shown in formula (5); in, This represents the number of pulses counted within the i-th phase window and the j-th amplitude level of the corresponding matrix X. The center phase of the i-th phase window; The amplitude asymmetry feature The calculation is shown in formula (6); in, This represents the maximum discharge amplitude during the positive half-cycle. This represents the maximum discharge amplitude during the negative half-cycle.
4. The generator partial discharge sample generation method based on physical constraints and reinforcement learning according to claim 1, characterized in that, In step S4, the comprehensive reward function R is constructed as shown in formula (7); in, , This represents the probability value of the true PRPD sample output by the discriminator, where D is the discriminator function, G is the generator function, and z is the random noise vector. This represents the logarithmic reward corresponding to the probability value; Indicates deviation in physical characteristics; As a hard constraint penalty, when the physical verification module outputs a penalty flag... The default penalty constant is used; otherwise, it is 0. , , These are the weighting coefficients.
5. The generator partial discharge sample generation method based on physical constraints and reinforcement learning according to claim 4, characterized in that, Step S5 specifically includes: S5.1: Initialize generator and discriminator parameters; S5.2: For a single iteration, sample B random noise vectors from the noise distribution and generate B simulated PRPD samples using the current generator; S5.3: Use a discriminator to distinguish between real PRPD samples and generated simulated PRPD samples, and update the discriminator parameters to minimize the cross-entropy loss; S5.4: Calculate the physical feature deviation and hard constraint penalty of each generated simulated PRPD sample using the physical verification module, and calculate the comprehensive reward value of the simulated PRPD sample by combining the discriminator output; S5.5: The REINFORCE algorithm, a reinforcement learning algorithm based on policy gradients, is used to calculate the gradient and update the generator parameters. The calculation is shown in formula (8), and the parameter update is shown in formula (9); in, Generator parameters The gradient operator, i.e., taking the partial derivative with respect to the parameters; The generator is represented by the b-th random noise vector. Generate simulated PRPD samples Generation strategy; The reward for the current sample. This is the historical reward baseline value. Here, B is the generator parameter, and B is the number of samples. This represents the learning rate of the generator; S5.6: Repeat the above S5.2-S5.5 process until the convergence condition is met. The convergence condition is: the discriminator probability values of the simulated PRPD samples all exceed the first preset threshold, the physical feature deviations are all lower than the second preset threshold, and the hard constraint penalty is 0. The first preset threshold is used to measure the realism of the generated simulated PRPD samples at the visual or statistical distribution level, and the second preset threshold is used to measure whether the physical characteristics of the generated simulated PRPD samples conform to the statistical laws of real PRPD samples.