Partial discharge signal identification positioning model training method

Abstract

The invention discloses a partial discharge signal identification and positioning model training method, which comprises the following steps: acquiring a partial discharge signal, and carrying out denoising processing on the partial discharge signal to obtain denoised sample data; dividing the de-noised sample data into a training set and a test set, and training a partial discharge signal identification and positioning model by using a partial discharge signal in the training set as an input and a discharge position corresponding to the partial discharge signal as an output; defining an execution action to represent a discharge position corresponding to the partial discharge signal predicted by the partial discharge signal identification positioning model under the time step t, and storing the correct execution action to a memory storage area; and randomly extracting part of data from the memory storage area, performing adversarial training by utilizing an adversarial data generation model to generate virtual adversarial data, and training a partial discharge signal identification and positioning model based on the data in the memory storage area and the virtual adversarial data.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power equipment discharge detection, and in particular to a local discharge signal recognition and positioning model training method. Background Art

[0002] Partial discharge is a local arc discharge phenomenon inside electrical equipment, which often causes damage to electrical equipment and even fires. The traditional gas insulated metal enclosed switchgear (GIS) partial discharge detection is to identify and locate the possible partial discharge phenomenon in the equipment through a series of traditional signal processing and feature extraction technologies. Use devices such as capacitive sensors or radio interference (Ultra High Frequency, UHF) sensors to collect discharge signals in GIS equipment in real time, filter to remove noise, analyze and extract signal features, and combine the acquisition time difference (Time of Flight, TOF) of the signal at different sensors or other positioning algorithms, such as time difference positioning method, to determine the specific location of the discharge source. Commonly used algorithms in traditional methods also include traditional machine learning methods such as support vector machines (SVM) or rule-based classification methods. The above processing method is implemented in a real-time system to continuously monitor the discharge of GIS equipment and detect potential faults in a timely manner.

[0003] However, when implementing partial discharge detection based on discharge signals, the acquisition cost of discharge signal data is high and the quality is often inconsistent, which will lead to insufficient training data or uneven data distribution, thus affecting the accuracy and generalization ability of the model. In addition, since discharge signals are often interfered by environmental noise, they may also be attacked by malicious attackers by injecting specific noise, and current deep learning models often show poor robustness to these adversarial attacks and noise.

[0004] In the related art, in a fault detection method based on the discharge signal of an insulating switchgear proposed in a patent application document with publication number CN114254668A, when training the model, the original local discharge sample signal is subjected to data enhancement processing to obtain a virtual local discharge sample signal, and the virtual local discharge sample signal and the initial local discharge sample signal are used as local discharge sample signals to train a waveform feature extraction model and an abnormality recognition model; however, in this scheme, the virtual samples are generated based on the basic signal data according to a preset threshold value and are used to expand the samples, but the expanded samples contain discharge signals and error signals, which are both used for subsequent feature recognition training, and no optimization is performed on the vulnerability of the model.

[0005] In the paper "A Review of Experience Replay in Off-Strategy Deep Reinforcement Learning, Hu Zijian et al., Acta Automatica Sinica", a model experience augmentation algorithm is proposed. The construction of an environmental model can generate a large amount of simulated experience while reducing the interaction between the agent and the environment. The paper proposes to design an experience pool to temporarily store the interaction experience between the agent and the environment, so that the agent can select appropriate experience from it to learn and update its own action strategy. It solves the problem of sample correlation in reinforcement learning, that is, how to use past experience to improve the efficiency and stability of training. It focuses on how to break the correlation of training samples and improve sample utilization. In implementation, the experience replay pool stores the agent's historical experience (state-action-reward-next state pairs) and randomly samples from it during the training process, which requires the design of a suitable data storage structure and sampling strategy. Summary of the invention

[0006] The technical problem to be solved by the present invention is how to enhance the recognition capability of a local discharge signal recognition and positioning model for different discharge signal characteristics and noises.

[0007] The present invention solves the above technical problems by the following technical means:

[0008] A partial discharge signal recognition and positioning model training method is proposed, the method comprising:

[0009] Collecting partial discharge signals and performing denoising processing on the partial discharge signals to obtain denoised sample data;

[0010] The denoised sample data is divided into a training set and a test set, and the partial discharge signal in the training set is used as input and the discharge position corresponding to the partial discharge signal is used as output to train the partial discharge signal recognition and positioning model;

[0011] An execution action is defined to represent a discharge position corresponding to a local discharge signal predicted by the local discharge signal recognition and positioning model at a time step t, and a correct execution action is stored in a memory storage area;

[0012] Part of the data is randomly extracted from the memory storage area and adversarial training is performed using an adversarial data generation model to generate virtual adversarial data, wherein the adversarial data generation model is:

[0013]

[0014] In the formula, is the expected data parameter, ο is the data parameter under the current training state, α is the anti-disturbance coefficient, δ1 is the lateral state disturbance difference, δ2 is the longitudinal state disturbance difference, is the expected signal state, p is the signal state under the current training state, Υ(ο,p) is the data under the training state, Υ(ο+δ1,p+δ2) is the perturbation adversarial data, Generate data for expectations;

[0015] The local discharge signal recognition and positioning model is trained based on the data in the memory storage area and the virtual confrontation data.

[0016] Furthermore, the collecting of partial discharge signals and denoising of the partial discharge signals to obtain denoised sample data include:

[0017] The sensor is used to collect the partial discharge signal x(t) of the power equipment at uniform time intervals;

[0018] Define carrier signal f b is the carrier frequency, is the initial phase of the carrier signal;

[0019] Define y(t)=x(t)×b(t), where the Fourier transforms of x(t), y(t), and b(t) are X(ω), Y(ω), and B(ω), respectively;

[0020] Define C(ω)=Y(ω)*B(ω), and perform inverse Fourier transform on C(ω) to obtain C(t):

[0021]

[0022] In the formula, f c is the frequency of the composite signal in the time domain, is the phase of the composite signal in the time domain;

[0023] Decompose C(t) to get x r (t) = A i e jΔT(t-1) and And for x r (t) and x o (t) respectively perform Fourier transform to obtain X r (ω),X o (ω);

[0024] Define signal X rz (ω) = X r (ω), X oz (ω) = X o (ω)-C(ω z ), X z (ω)=(X rz (ω)+X oz (ω))B(ω);

[0025] X z (ω) Perform inverse Fourier transform to obtain the denoised sample data.

[0026] Furthermore, the method of using the partial discharge signal in the training set as input and the discharge position corresponding to the partial discharge signal as output to train the partial discharge signal recognition and positioning model includes:

[0027] Define the training set G = {(m1,n1),(m2,n2),...,(m r ,n r )}, where m r represents the rth local discharge signal in the training set G, n r Indicates the partial discharge signal m r The corresponding discharge position, r represents the number of discharge signals in the training set G;

[0028] Define the state domain P = {p1, p2, ..., p r}, the partial discharge signal is recorded as the system state, and the system state p at time step t r m corresponding to the training data set G r ;

[0029] The partial discharge signal in the training set is used as input and the discharge position corresponding to the partial discharge signal is used as output to train the partial discharge signal recognition and positioning model.

[0030] Furthermore, the definition of the execution action represents the discharge position corresponding to the local discharge signal predicted by the local discharge signal recognition and positioning model at the time step t, and the correct execution action is stored in the memory storage area, including:

[0031] Define execution domain Q = {q1,q2,…,q r}, define the position coordinates (x, y, z) of the partial discharge signal, and execute action q t represents the discharge position guessed by the local discharge signal recognition and positioning model at time step t, where represents the discharge position coordinates predicted by the local discharge signal recognition and positioning model;

[0032] After completing a system execution action, when the evaluation function is used to evaluate whether the execution action is correct, the correct execution action data is Stored in the memory storage area, the system state p selects the execution action q to get the evaluation function to evaluate to true and enter the next state s represents the evaluation result of the system execution action.

[0033] Furthermore, the method further comprises:

[0034] After completing a system execution action, the data in the training set G is randomly shuffled.

[0035] Furthermore, the evaluation function formula is expressed as:

[0036]

[0037] In the formula, x tr ,y tr With z tr Represent the partial discharge signal m r Corresponding discharge position n r The x-coordinate, y-coordinate, and z-coordinate of .

[0038] Furthermore, the loss function L of the adversarial data generation model is dk for:

[0039]

[0040] Where μ is the loss weight and N[] is defined as the position measurement function.

[0041] Furthermore, the perturbation gradient l and perturbation amplitude h of the adversarial data generation model are respectively:

[0042]

[0043]

[0044] In the formula, is the adversarial gradient function, δ is the perturbation difference, ∈ is the perturbation limit, and ||||2 is the norm calculation symbol.

[0045] Further, when the evaluation function is used to evaluate that the execution action is incorrect, it is determined that there is an error in the partial discharge signal recognition and positioning model switching execution action, and the execution action is optimized using the action optimization function, and the operation action optimization function is:

[0046]

[0047] In the formula, the sigmoid function makes q t Normalized to [0,1], f t () is the state switching action function, Q t To perform the action after optimization, q t Make a guess about the signal location for the model.

[0048] Furthermore, after training the partial discharge signal recognition and positioning model based on the data in the memory storage area and the virtual confrontation data, the method further includes:

[0049] Using the test set to test the trained partial discharge signal recognition and positioning model;

[0050] The training results are compared with the test results to obtain the final desired discharge signal recognition and positioning model.

[0051] The advantages of the present invention are:

[0052] (1) During the training process of the deep learning model, the present invention introduces virtual adversarial training to generate challenging adversarial samples, so that the model can be exposed to more changes and situations in different environments during the training process, thereby enhancing the model's ability to recognize different discharge signal characteristics and noise. Virtual adversarial training can help the model resist adversarial attacks and noise injection on discharge signal classification, improve the robustness and security of the model, and ensure that it can run stably in practical applications. The memory storage area can store historical data or model status. By combining virtual adversarial training and memory storage area optimization, the model parameters can be optimized more effectively, and the training efficiency and the generalization ability of the model can be improved.

[0053] (2) Introducing adversarial samples and adaptively sensing the disturbance direction to which the neural network is most sensitive to classification behavior, and updating the model by reducing the output difference between samples before and after the disturbance, can achieve the effect of strengthening the decision boundary and make the boundaries between different types of samples clearer, thereby effectively improving the model's ability to resist interference.

[0054] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a partial discharge signal recognition and positioning model training method proposed in one embodiment of the present invention;

[0056] Figure 2 It is a principle block diagram of a partial discharge signal recognition and positioning model training method in one embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] like Figure 1As shown, an embodiment of the present invention provides a method for training a partial discharge signal recognition and positioning model, the method comprising the following steps:

[0059] S10, collecting partial discharge signals, and performing denoising processing on the partial discharge signals to obtain denoised sample data;

[0060] S20, dividing the denoised sample data into a training set and a test set, and using the partial discharge signal in the training set as input and the discharge position corresponding to the partial discharge signal as output to train the partial discharge signal recognition and positioning model;

[0061] S30, defining an execution action to represent a discharge position corresponding to a partial discharge signal predicted by the partial discharge signal recognition and positioning model at a time step t, and storing the correct execution action in a memory storage area;

[0062] S40, randomly extracting part of the data from the memory storage area and performing adversarial training using an adversarial data generation model to generate virtual adversarial data, wherein the adversarial data generation model is:

[0063]

[0064] In the formula, is the expected data parameter, ο is the data parameter under the current training state, α is the anti-disturbance coefficient, δ1 is the lateral state disturbance difference, δ2 is the longitudinal state disturbance difference, is the expected signal state, p is the signal state under the current training state, Υ(ο,p) is the data under the training state, Υ(ο+δ1,p+δ2) is the perturbation adversarial data, Generate data for expectations;

[0065] Specifically, this embodiment introduces adversarial perturbations into existing partial discharge signal data to generate adversarial data. These perturbations are intended to challenge and test the robustness of existing models. When designing adversarial samples, a loss function is usually defined to measure the effectiveness of the perturbations, so that the adversarial samples can cause the model to misclassify as much as possible while maintaining the similarity between the adversarial samples and the original samples. Optimization algorithms (such as gradient descent) are specifically used to create adversarial perturbations. The optimization algorithm allows fine-tuning of the perturbations so that the generated adversarial samples can challenge the model more effectively, find effective adversarial perturbations more quickly, and improve training and generation efficiency. By adjusting the loss function and optimization strategy, it can adapt to different adversarial sample generation requirements.

[0066] These samples are used to train and optimize the recognition model so that it can better handle interference in the actual environment. The generated adversarial samples can reveal the model's weaknesses when facing interference, which helps researchers and engineers understand the shortcomings of the model and thus improve and optimize the model. In the identification of partial discharge signals, generating adversarial data can help improve the model's ability to recognize different types of partial discharge signals, making it more reliable in fault diagnosis of actual equipment.

[0067] S50: Training the partial discharge signal recognition and positioning model based on the data in the memory storage area and the virtual adversarial data.

[0068] It should be noted that this embodiment uses the correct sample data in the memory storage area for adversarial training to generate virtual adversarial data, so that the model can be continuously exposed to more challenging situations during the training process, which helps to improve the adversarial nature of the model. Adversarial training can make the model more robust when facing similar misleading situations. In addition, the memory storage area is used to store the correct operations and the correct data is used to regularly train the continuously updated system model to enhance the model's adaptability and dynamic adjustment capabilities. If there is no memory storage area, the model may only rely on the current training data, resulting in low data utilization, which may limit the model's learning ability, especially when the amount of data is insufficient or the data distribution changes greatly. The lack of diverse training data and the challenging training environment may lead to a decrease in the model's generalization ability. The model may be more likely to fall into overfitting or over-adaptation to a specific data distribution, thereby performing poorly when faced with new data or changes.

[0069] In the present embodiment, during the training process of the deep learning model, by introducing virtual adversarial training to generate challenging adversarial samples, the model can be exposed to more changes and situations in different environments during the training process, thereby enhancing the model's ability to recognize different discharge signal characteristics and noise. Virtual adversarial training can help the model resist adversarial attacks and noise injection for discharge signal classification, improve the robustness and security of the model, and ensure that it can run stably in practical applications. The memory storage area can store historical data or model status. By combining virtual adversarial training and optimization of the memory storage area, the model parameters can be optimized more effectively, and the training efficiency and the generalization ability of the model can be improved.

[0070] As a further preferred technical solution, the step S10: collecting partial discharge signals and performing denoising on the partial discharge signals to obtain denoised sample data, comprises the following steps:

[0071] S11. Use sensors to collect partial discharge signals x(t) of power equipment at uniform time intervals, where:

[0072]

[0073] In the formula, A i is the starting amplitude of the sampling data, is the starting phase of the sampling data, ΔT is the sampling period, and t=1,2,…,N is the number of times the signal is collected.

[0074] S12. Define carrier signal f b is the carrier frequency, is the initial phase of the carrier signal;

[0075] S13. Define y(t)=x(t)×b(t), where the Fourier transforms of x(t), y(t), and b(t) are X(ω), Y(ω), and B(ω), respectively;

[0076] S14. Define C(ω)=Y(ω)*B(ω), and perform inverse Fourier transform on C(ω) to obtain C(t):

[0077]

[0078] In the formula, f c is the frequency of the composite signal in the time domain, is the phase of the composite signal in the time domain, C(ω) is the composite carrier signal, and C(t) is the composite signal in the time domain;

[0079] S15. Decompose C(t) to get x r (t) = A i e jΔT(t-1) and And for x r (t) and x o (t) respectively perform Fourier transform to obtain X r (ω),X o (ω), where x r (t) is the main signal part of the time domain composite signal, x o (t) is the auxiliary signal part of the composite signal in the time domain, and ΔT is the sampling period;

[0080] S16. Define signal X rz (ω) = X r (ω), X oz (ω) = X o (ω)-C(ω z ), X z (ω)=(X rz (ω)+X oz (ω))B(ω), where X rz (ω) is the main signal part of the denoised signal, X oz (ω) is the auxiliary signal part of the denoising signal, Xz (ω) is the denoised signal;

[0081] S17, to X z (ω) Perform inverse Fourier transform to obtain the denoised sample data.

[0082] This embodiment extracts the main components of the signal by fitting the exponential decay or oscillation components of the signal, which helps to identify and separate useful signal components from complex signals; by introducing regularization terms and adopting an iterative inversion calculation method, noise can be effectively removed and the signal can be restored, reducing the impact of noise on the results. The combination of the two methods can remove noise more accurately, help optimize the denoising process, and thus improve the quality of the signal.

[0083] Therefore, this method can accurately extract the main features of the signal, improve the accuracy of signal recovery, and reduce noise interference. It is suitable for signals with complex frequency characteristics, has strong adaptability, and can effectively extract signals in the case of multiple frequency components.

[0084] It should be noted that after denoising, the data are randomly divided into training data set G and test data set F according to a ratio of 3:1.

[0085] As a further preferred technical solution, in step S20: using the partial discharge signal in the training set as input and the discharge position corresponding to the partial discharge signal as output to train the partial discharge signal recognition and positioning model includes the following steps:

[0086] S21, define the training set G = {(m1,n1),(m2,n2),...,(m r ,n r )}, where m r represents the rth local discharge signal in the training set G, n r Indicates the partial discharge signal m r The corresponding discharge position, r represents the number of discharge signals in the training set G;

[0087] S22, define the state domain P = {p1, p2, ..., p r}, the partial discharge signal is recorded as the system state, and the system state p at time step t r m corresponding to the training data set G r ;

[0088] S23, using the partial discharge signal in the training set as input and the discharge position corresponding to the partial discharge signal as output to train the partial discharge signal recognition and positioning model.

[0089] It should be noted that the partial discharge signal recognition and localization model adopted in this embodiment is a deep learning model (Quantum Neural Network, QNN).

[0090] As a further preferred technical solution, step S30: Define an execution action to represent the discharge position corresponding to the partial discharge signal predicted by the partial discharge signal recognition and localization model at time step t, and store the correct execution action in the memory storage area, including the following steps:

[0091] S31. Define the execution domain Q = {q1, q2, …, q r}, define the partial discharge signal position coordinates (x, y, z), and the execution action q t represents the guessed discharge position of the partial discharge signal recognition and localization model at time step t, where represents the predicted discharge position coordinates of the partial discharge signal recognition and localization model;

[0092] S32. After completing a system execution action, when the execution action is evaluated as correct by using the evaluation function, store the correct execution action data into the memory storage area. The system state p selects the execution action q and is evaluated as true by the evaluation function and enters the next state s represents the evaluation result of the system execution action.

[0093] It should be noted that in this embodiment, by defining the memory storage area W for storing the execution actions evaluated as correct, randomly extracting part of the data from the memory storage area W to participate in the adversarial training to generate virtual adversarial data, training the model using the data in the memory storage area every custom k1 time steps, and training the model using the virtual adversarial data every custom k2 time steps, the training of the target discharge signal recognition model is realized. The values are customized according to the model training progress. Generally, k1 < k2 < the step length of training the model with the training set data.

[0094] In this embodiment, the implementation of the designed memory storage area usually involves establishing a data structure (such as a queue, an array, a database) to store experiences, generating adversarial samples (applying small perturbations to the input data), and using these adversarial samples together with the original samples to train the model.

[0095] As a further preferred technical solution, the method further includes:

[0096] After completing a system execution action, randomly shuffle the data in the training set G.

[0097] It should be noted that, after completing a system execution action, this embodiment randomly shuffles the data in the training set, which can help reduce the risk of overfitting of the model and improve the generalization ability of the model. By shuffling the data, it can be ensured that each batch of samples is extracted from different parts of the data set, thereby reducing the impact of the data order on model training. This method helps the model better learn the general characteristics of the data, rather than memorizing the order or specific patterns of the data.

[0098] As a further preferred technical solution, the evaluation function formula is expressed as:

[0099]

[0100] In the formula, x tr ,y tr With z tr Represent the partial discharge signal m r Corresponding discharge position n r The x-coordinate, y-coordinate, and z-coordinate of .

[0101] As a further preferred technical solution, the loss function L of the adversarial data generation model is dk for:

[0102]

[0103] Where μ is the loss weight and N[] is defined as the position measurement function.

[0104] It should be noted that this embodiment uses the loss function of the adversarial data generation model to measure the gap between the prediction model and the actual model to optimize the model and minimize this gap, so that the model's prediction is more accurate.

[0105] By calculating the difference between the predicted value and the actual value, the loss function provides a criterion for quantifying the performance of the model, provides optimization direction for model training, and helps adjust model parameters to reduce prediction errors.

[0106] As a further preferred technical solution, the perturbation gradient l and perturbation amplitude h of the adversarial data generation model are respectively:

[0107]

[0108]

[0109] In the formula, is the adversarial gradient function, δ is the perturbation difference, ∈ is the perturbation limit, and ||||2 is the norm calculation symbol.

[0110] In this embodiment, the perturbation gradient is set to indicate how to adjust the input data in the input space to maximize the loss function of the model. Using the perturbation gradient, adversarial samples can be generated. Adversarial samples are generated by fine-tuning the input data along the gradient direction. These samples will cause the model to make incorrect predictions, thereby testing and improving the robustness of the model. Using gradient information to generate adversarial samples can ensure that the generated samples have a significant impact on the loss function of the model, effectively exposing the weaknesses of the model. The gradient direction provides the precise generation direction of the adversarial samples, making the generated samples more targeted and effectively testing the robustness of the model in a specific direction.

[0111] The perturbation amplitude is used to control the size of the perturbation, thereby controlling the degree of change in the adversarial sample. By adjusting the perturbation amplitude, the difficulty of generating adversarial samples can be balanced. If the amplitude is too large, the adversarial sample may deviate from the distribution of the original sample, resulting in unstable training; if the amplitude is too small, it may not be able to effectively challenge the model. Limiting the perturbation amplitude helps the generated adversarial samples to be more feasible in practical applications. Controlling the perturbation amplitude can adjust the strength of the adversarial sample to ensure that it is sufficient to challenge the model but does not deviate from being invalid.

[0112] As a further preferred technical solution, when the evaluation function is used to evaluate that the execution action is incorrect, it is determined that there is an error in the partial discharge signal recognition and positioning model switching execution action, and the execution action is optimized using an action optimization function, and the action optimization function is:

[0113]

[0114] In the formula, the sigmoid function makes q t Normalized to [0,1], f t () is the state switching action function, Q t To perform the action after optimization, q t Make a guess about the signal location for the model.

[0115] In this embodiment, the action optimization function is used to update the strategy based on the reward signal, which provides feedback on the effect of the system model taking actions in the environment. The optimization function aims to maximize the cumulative reward and helps the system model select the optimal action by evaluating the impact of different actions on the cumulative reward.

[0116] The system model can continuously adjust its action strategy based on the feedback from the environment. By optimizing the action selection strategy, the performance of the system model in the environment can be significantly improved, such as higher scores, fewer errors, or higher efficiency. The optimization function enables the system model to effectively learn and adapt in a complex and dynamic environment, improving the flexibility of the system. Through the optimization function, the system model can find the best strategy faster, reducing training time and consumption of computing resources.

[0117] It should be noted that the sample data generated by virtual adversarial means in this embodiment is generated by adding perturbation data to the sample data that has been identified as correct, and corresponding adversarial samples are generated according to the model training progress and defects. The generated data is used for subsequent targeted error training of the model to enable the model to evolve in a targeted manner, which helps to improve the robustness and overall performance of the model on adversarial samples and enables the model to perform better when facing specific defects.

[0118] As a further preferred technical solution, Figure 2 As shown, after the partial discharge signal recognition and positioning model is trained based on the data in the memory storage area and the virtual confrontation data, the method further includes:

[0119] Using the test set to test the trained partial discharge signal recognition and positioning model;

[0120] The training results are compared with the test results to obtain the final desired discharge signal recognition and positioning model.

[0121] The adversarial memory storage optimization algorithm (Virtual Defense-Quantum Neural Network, VD-QQN) proposed in this embodiment combines virtual adversarial training and optimization of the memory storage area W. It mainly solves the problem of how to effectively store and manage the data generated during the training process, and the robustness of the model to input perturbations, that is, how to make the model still maintain high performance when facing adversarial samples. It can more effectively optimize model parameters, improve training efficiency and model generalization ability, and introduce adversarial samples and adaptively perceive the disturbance direction to which the neural network is most sensitive to classification behavior, and update the model by reducing the output difference of samples before and after the disturbance. It can achieve the effect of strengthening the decision boundary, making the boundary between different types of samples clearer, thereby effectively improving the model's ability to resist interference.

[0122] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0123] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0124] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A partial discharge signal recognition and positioning model training method, characterized in that: The method comprises: Collecting partial discharge signals and performing denoising processing on the partial discharge signals to obtain denoised sample data; The denoised sample data is divided into a training set and a test set, and the partial discharge signal in the training set is used as input and the discharge position corresponding to the partial discharge signal is used as output to train the partial discharge signal recognition and positioning model; An execution action is defined to represent a discharge position corresponding to a local discharge signal predicted by the local discharge signal recognition and positioning model at a time step t, and a correct execution action is stored in a memory storage area; Part of the data is randomly extracted from the memory storage area and adversarial training is performed using an adversarial data generation model to generate virtual adversarial data, wherein the adversarial data generation model is: In the formula, is the expected data parameter, ο is the data parameter under the current training state, α is the anti-disturbance coefficient, δ1 is the lateral state disturbance difference, δ2 is the longitudinal state disturbance difference, p is the signal state under the current training state, is the expected signal state, Υ(ο,p) is the data under the training state, Υ(ο+δ1,p+δ2) is the perturbation adversarial data, Generate data for expectations; The local discharge signal recognition and positioning model is trained based on the data in the memory storage area and the virtual adversarial data.

2. The partial discharge signal recognition and positioning model training method according to claim 1, characterized in that: The collecting of partial discharge signals and denoising the partial discharge signals to obtain denoised sample data include: The sensor is used to collect the partial discharge signal x(t) of the power equipment at uniform time intervals; Define carrier signal f b is the carrier frequency, is the initial phase of the carrier signal; Define y(t)=x(t)×b(t), where the Fourier transforms of x(t), y(t), and b(t) are X(ω), Y(ω), and B(ω), respectively; Define C(ω)=Y(ω)*B(ω), and perform inverse Fourier transform on C(ω) to obtain C(t): In the formula, f c is the frequency of the composite signal in the time domain, is the phase of the composite signal in the time domain; Decompose C(t) to get x r (t) = A i e jΔT(t-1) and And for x r (t) and x o (t) respectively perform Fourier transform to obtain X r (ω),X o (ω); define signalX rz (ω)=X r (ω), X oz (ω)=X o (ω)-C(ω z ),X z (ω)=(X rz (ω)+X oz (ω))B(ω); X z (ω) Perform inverse Fourier transform to obtain the denoised sample data.

3. The partial discharge signal recognition and positioning model training method according to claim 1, characterized in that: The method uses the local discharge signal in the training set as input and the discharge position corresponding to the local discharge signal as output to train the local discharge signal recognition and positioning model, including: Define the training set G = {(m1,n1),(m2,n2),...,(m r ,n r )}, where m r represents the rth local discharge signal in the training set G, n r Indicates the partial discharge signal m r The corresponding discharge position, r represents the number of discharge signals in the training set G; Define the state domain P = {p1, p2, ..., p r }, the partial discharge signal is recorded as the system state, and the system state p at time step t r m corresponding to the training data set G r ; The partial discharge signal in the training set is used as input and the discharge position corresponding to the partial discharge signal is used as output to train the partial discharge signal recognition and positioning model.

4. The partial discharge signal recognition and positioning model training method according to claim 1, characterized in that: The defined execution action represents the discharge position corresponding to the local discharge signal predicted by the local discharge signal recognition and positioning model at the time step t, and the correct execution action is stored in the memory storage area, including: Define execution domain Q = {q1,q2,…,q r }, define the position coordinates (x, y, z) of the partial discharge signal, and execute action q t represents the discharge position guessed by the local discharge signal recognition and positioning model at time step t, where represents the discharge position coordinates predicted by the local discharge signal recognition and positioning model; After completing a system execution action, when the evaluation function is used to evaluate whether the execution action is correct, the correct execution action data is Stored in the memory storage area, the system state p selects the execution action q to get the evaluation function to evaluate to true and enter the next state s represents the evaluation result of the system execution action.

5. The partial discharge signal recognition and positioning model training method according to claim 4, characterized in that: The method further comprises: After completing a system execution action, the data in the training set G is randomly shuffled.

6. The partial discharge signal recognition and positioning model training method according to claim 4, characterized in that: The evaluation function formula is expressed as: In the formula, x tr ,y tr With z tr Represent the partial discharge signal m r Corresponding discharge position n r The x-coordinate, y-coordinate, and z-coordinate of .

7. The partial discharge signal recognition and positioning model training method according to claim 1, characterized in that: The loss function L of the adversarial data generation model dk for: Where μ is the loss weight and N[ ] is defined as the position measurement function.

8. The partial discharge signal recognition and positioning model training method according to claim 1, characterized in that: The perturbation gradient l and perturbation amplitude h of the adversarial data generation model are respectively: In the formula, is the adversarial gradient function, δ is the perturbation difference, ∈ is the perturbation limit, and || ||2 is the norm calculation symbol.

9. The partial discharge signal recognition and positioning model training method according to claim 5, characterized in that: When the evaluation function is used to evaluate that the execution action is incorrect, it is determined that there is an error in the partial discharge signal recognition and positioning model switching execution action, and the execution action is optimized using the action optimization function, and the operation action optimization function is: In the formula, the sigmoid function makes q t Normalized to [0,1], f t ( ) is the state switching action function, Q t To perform the action after optimization, q t Make a guess about the signal location for the model.

10. The partial discharge signal recognition and positioning model training method according to any one of claims 1 to 9, characterized in that: After training the partial discharge signal recognition and positioning model based on the data in the memory storage area and the virtual confrontation data, the method further includes: Using the test set to test the trained partial discharge signal recognition and positioning model; The training results are compared with the test results to obtain the final desired discharge signal recognition and positioning model.

Images