Gyrotron abnormal test data identification method and system

Abstract

The present invention discloses a method and system for identifying abnormal gyrotron test data, the method comprising: step 1, constructing a gyrotron test data set, and initializing the parameters of a feature extractor and a Gaussian mixture model; step 2, performing preheating training, using a feature extractor to characterize and learn the gyrotron test data, obtaining the potential features of the data, and continuously judging whether the preheating training is completed, if not completed, continuously optimizing the learning target, if completed, entering step 3, wherein the feature extractor is a deep neural network; step 3, after the preheating is completed, using a Gaussian mixture model to perform distribution modeling on the potential features of the gyrotron test data; step 4, using a Gaussian mixture model to calculate the distribution confidence of the gyrotron test data, and identifying abnormal gyrotron test data; step 5, repeating steps 2-4 until the target training function converges. The present invention not only significantly improves the accuracy of the gyrotron fault identification task, but also can identify abnormal data points in the gyrotron test data, providing a strong guarantee for ensuring the stable operation of the ECRH system.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and electronic cyclotron heating, and in particular to gyrotron fault type prediction and gyrotron abnormal test data identification. Background Art

[0002] The gyrotron is a core component of the electron cyclotron resonance heating (ECRH) system. It is a vacuum electronic device based on the principle of electron cyclotron resonance emission, and its main task is to generate high-power microwave energy. The stable operation of the gyrotron is crucial to the normal operation of the ECRH system. However, during the high-power, long-pulse operation of the gyrotron, due to the influence of various factors such as mode competition and power supply fluctuations, the system may experience faults such as radio frequency oscillation, resulting in instability of the output power. This instability will not only have a negative impact on the performance of the gyrotron, but may also pose a serious threat to the effective operation of the entire ECRH system. To address this issue, some researchers have begun to use deep learning-based methods to predict potential failures that may occur during the operation of the gyrotron, so as to take preventive measures in advance to ensure the stable operation of the ECRH system.

[0003] At present, deep learning-based gyrotron fault identification algorithms rely heavily on historical test data acquired by the gyrotron test platform. The performance of these algorithms is significantly affected by data quality. The existence of this dependence requires the high quality of the collected gyrotron test data. In other words, ensuring high-quality gyrotron test data sets becomes a key factor in improving the performance of gyrotron fault identification algorithms. However, during the gyrotron test data collection process, it may be affected by human and environmental factors. For example, recording errors by researchers and changes in humidity, light, and temperature in the environment may lead to the acquisition of unreasonable gyrotron test data. These interference factors not only affect the accuracy and reliability of the data, but may also have a negative impact on the performance of subsequent fault identification algorithms.

[0004] Existing abnormal data identification methods usually use big data to build a data set for the corresponding task, and then learn the data features through some specific machine learning algorithms to identify abnormal data points.

[0005] The Chinese patent with publication number CN118797243A discloses a method and system for cleaning and eliminating abnormal data for big data analysis. The method constructs a sales data sequence at each time point in a preset time period and clusters it to obtain clusters of all time points. The abnormal data is identified based on the difference between the reference sequence and the target cluster, solving the problem of mutual influence of multi-channel pharmaceutical sales data. However, the performance of the clustering method is unstable, which may affect the overall stability of the abnormal data cleaning system.

[0006] The Chinese patent with publication number CN116894151A discloses a method for cleaning abnormal data of smart grid based on artificial intelligence. The method uses the LOF algorithm to detect and clean abnormal data by analyzing the characteristic vector of collected electric load data. However, due to the uncertainty of data distribution, the local density between abnormal data and adjacent normal data may be relatively close, which leads to the unstable performance of the LOF algorithm, thereby reducing the detection accuracy of abnormal data. Summary of the invention

[0007] The technical problem to be solved by the present invention is how to accurately detect and identify abnormal test data of a gyrotron and improve the accuracy of gyrotron fault identification, thereby ensuring the stable operation of an ECRH system.

[0008] The present invention solves the above technical problems through the following technical means: a method for identifying abnormal test data of a gyrotron, which is carried out in the following steps:

[0009] Step 1: Construct a gyrotron test data set and initialize the parameters of the feature extractor and Gaussian mixture model;

[0010] Step 2: Perform warm-up training, use a feature extractor to characterize and learn the gyrotron test data, obtain the potential features of the data, and continuously determine whether the warm-up training is completed. If not, continue to optimize the learning objectives. If completed, proceed to step 3, where the feature extractor is a deep neural network.

[0011] Step 3: After the warm-up is completed, the Gaussian mixture model is used to perform distribution modeling on the potential features of the learned gyrotron test data;

[0012] Step 4: Use the Gaussian mixture model to calculate the distribution confidence of the gyrotron test data and identify abnormal gyrotron test data;

[0013] Step 5: Repeat steps 2-4 until the target training function converges.

[0014] As a further optimized technical solution, the gyrotron test data set is ,in Indicates the gyrotron parameters related to possible gyrotron failure. represents the working state of the gyrotron, and the Gaussian mixture model is , the distribution of its fitting Depend on and Two parameters determine:

[0015] ,

[0016] in, is a discrete latent variable that determines the assignment of observations to Gaussian mixture components; The observation data belongs to The probability of a sub-model; is the latent feature of the gyrotron test data after neural network learning; and They are the mean and standard deviation of the data learned by the Gaussian mixture model; is a given parameter and Post-prediction The probability distribution expression of .

[0017] As a further optimized technical solution, step 1 specifically includes:

[0018] Step 1.1: First, build a gyrotron test dataset ,in Indicates the gyrotron parameters related to possible gyrotron failure. Indicates the working status of the gyrotron.

[0019] Step 1.2 Initialize the feature extractor The parameters of the deep neural network and the parameters of the Gaussian mixture model.

[0020] As a further optimized technical solution, step 2 specifically includes:

[0021] Step 2.1: Assume the gyrotron test data set There is no abnormal data in the feature extractor. The optimization direction is used to find a better solution. After the preset training rounds are completed, the warm-up training ends. Before the warm-up training ends, only the parameter optimization learning of the deep neural network is performed.

[0022] Step 2.2: Optimize feature extractor The parameters of the deep neural network are optimized as follows:

[0023] ,

[0024] in, is the hyperparameter weight factor; is the prediction classification loss defined; is the contrastive learning loss defined; It is the result of fault prediction by multi-layer perceptron on the potential features of gyrotron test data; Indicates the working status of the gyrotron; is the latent feature of the gyrotron test data after learning; is the hyperparameter scaling factor used in contrastive learning; and For the positive and negative samples set using contrastive learning, only the sample itself is set as the positive sample, and the rest of the samples are negative samples.

[0025] As a further optimized technical solution, the feature extractor It mainly includes three core parts, namely shallow feature extraction layer, deep feature extraction layer and prediction classification layer;

[0026] The shallow feature extraction layer includes a first linear layer, a first ReLU activation function, and a second linear layer connected in sequence, and the input end of the first linear layer inputs gyrotron test data;

[0027] The deep feature extraction layer includes two important units: a weighted quantization module and a self-gating module;

[0028] The reweighting module comprises a third linear layer, a pooling layer, a first multiplication unit, a second ReLU activation function, and a first addition unit connected in sequence, wherein the output end of the second linear layer is respectively connected to the input ends of the third linear layer, the first addition unit, and the first multiplication unit, the output end of the first multiplication unit is connected to the second ReLU activation function, and the output end of the first addition unit is connected to the self-gating module element;

[0029] The self-gating module includes a fourth linear layer, a third ReLU activation function, a second multiplication unit, a fourth ReLU activation function, and a second addition unit connected in sequence, wherein the output end of the first addition unit is simultaneously connected to the fourth linear layer, the second multiplication unit, and the second addition unit, the output end of the second multiplication unit is connected to the fourth ReLU activation function, and the output end of the second addition unit is connected to the prediction classification layer;

[0030] The prediction classification layer includes a fifth linear layer, which uses the high-dimensional gyrotron potential feature representation learned by the deep feature extraction layer to make a final prediction, output the result of gyrotron fault identification, and obtain the potential features of the gyrotron test data after learning. : .

[0031] As a further optimized technical solution, step 3 specifically includes:

[0032] In the distribution modeling process, the feature extractor Latent features of gyrotron data after learning With discrete latent variables Perform correlation and then optimize the Gaussian mixture model Parameters and :

[0033] ,

[0034] ,

[0035] in, represents the gyrotron operation state predicted by the model; represents L2 regularization; is the latent feature of the gyrotron test data after learning; are the parameters of the neural network; and Gaussian mixture model Learned data mean and variance; T represents the matrix transpose operation.

[0036] As a further optimized technical solution, step 4 specifically includes:

[0037] Step 4.1: Complete the Gaussian mixture model Parameters and After optimization, the latent features of the gyrotron test data after learning are calculated. The posterior probability of , to obtain the confidence score, confidence The calculation formula is:

[0038]

[0039] in, and All have been L2 regularized; Indicates the gyrotron parameters related to possible gyrotron failure at a given In the case of Gaussian mixture model predict Belong to The probability of a class, here there are only two categories: abnormal and normal;

[0040] Step 4.2: Based on the calculated confidence score, Update and move the gyrotron anomaly test data from the gyrotron test dataset Filtered and excluded from the training step.

[0041] The present invention also provides a gyrotron abnormal test data identification system, comprising the following modules:

[0042] The test data set module is used to construct the gyrotron test data set and initialize the parameters of the feature extractor and Gaussian mixture model;

[0043] A warm-up training module is used to perform warm-up training, use a feature extractor to characterize and learn the gyrotron test data, obtain the potential features of the data, and continuously determine whether the warm-up training is completed. If not, continue to optimize the learning objectives. If completed, enter the distribution modeling module. The feature extractor is a deep neural network.

[0044] A distribution modeling module is used to perform distribution modeling on the potential features of the learned gyrotron test data using a Gaussian mixture model after the warm-up is completed;

[0045] A confidence calculation module, used to calculate the distribution confidence of gyrotron test data using a Gaussian mixture model, and identify abnormal gyrotron test data;

[0046] The target convergence module is used to converge the target training function.

[0047] As a further optimized technical solution, the distribution modeling module specifically includes:

[0048] In the distribution modeling process, the feature extractor Latent features of gyrotron data after learning With discrete latent variables Perform correlation and then optimize the Gaussian mixture model Parameters and :

[0049] ,

[0050] ,

[0051] in, represents the gyrotron operation state predicted by the model; represents L2 regularization; is the latent feature of the gyrotron test data after learning; are the parameters of the neural network; and Gaussian mixture model Learned data mean and variance; T represents the matrix transpose operation.

[0052] As a further optimized technical solution, the confidence calculation module includes:

[0053] Computational unit, used to complete the Gaussian mixture model Parameters and After optimization, the latent features of the gyrotron test data after learning are calculated. The posterior probability of , to obtain the confidence score, confidence The calculation formula is:

[0054]

[0055] in, and All have been L2 regularized; Indicates the gyrotron parameters related to possible gyrotron failure at a given In the case of Gaussian mixture model predict Belong to The probability of a class, here there are only two categories: abnormal and normal;

[0056] Update unit, used to update the gyrotron test dataset according to the calculated confidence score Update and move the gyrotron anomaly test data from the gyrotron test dataset Filtered and excluded from the training step.

[0057] The advantages of the present invention are:

[0058] The present invention uses a deep neural network to perform feature characterization learning on gyrotron test data for the task of gyrotron fault identification, thereby effectively extracting potential features. This process can capture complex patterns and important relationships in the data through multi-layer nonlinear transformations, providing a solid foundation for subsequent anomaly detection. After the potential features are extracted, these potential features are distributed modeled using a Gaussian mixture model to obtain a robust data distribution. As a powerful probabilistic model, the Gaussian mixture model can accurately describe the inherent connection of the data through a weighted combination of multiple Gaussian distributions. This modeling process not only provides effective theoretical support for the identification of abnormal data, but also enhances the understanding of the relationship between normal and abnormal data. Based on the data distribution generated by the Gaussian mixture model, the confidence is further calculated to achieve accurate identification of gyrotron abnormal test data. By setting an appropriate confidence threshold, normal data and abnormal data can be effectively distinguished, thereby improving the accuracy of the gyrotron fault identification task.

[0059] This method not only significantly improves the accuracy of gyrotron fault identification tasks, but also can identify abnormal data points in gyrotron test data, providing a strong guarantee for ensuring the stable operation of the ECRH system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flow chart of a method for identifying abnormal gyrotron test data in an embodiment of the present invention;

[0061] Figure 2 2 is a diagram of the architecture of a feature extractor in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] 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.

[0063] Define the gyrotron test data set as ,in Indicates gyrotron parameters related to possible gyrotron failure, such as anode voltage, cathode voltage, etc.; Indicates the working status of the gyrotron, such as ignition, overcurrent, etc.

[0064] Define the feature extractor as , which is a deep neural network composed of multi-layer perceptrons, activation functions, etc.

[0065] The Gaussian mixture model is defined as , the distribution of its fitting Depend on and Two parameters determine:

[0066] ,

[0067] in, is a discrete latent variable that determines the assignment of observations to Gaussian mixture components; The observation data belongs to The probability of a sub-model; is the potential feature of the gyrotron test data after neural network learning (hereinafter referred to as is the latent features of the gyrotron test data after learning); and They are the mean and standard deviation of the data learned by the Gaussian mixture model; is a given parameter and Post-prediction The probability distribution expression of .

[0068] See also Figure 1 The gyrotron abnormal test data identification method of the present invention is carried out according to the following steps:

[0069] Step 1: Construct a gyrotron test dataset and initialize the parameters of the feature extractor and Gaussian mixture model.

[0070] Step 1.1 When using the gyrotron anomaly test data identification method, first build a gyrotron test data set , including gyrotron operating parameters such as anode voltage, cathode voltage, water load, and gyrotron operating status. These data provide a basis for subsequent feature characterization learning and anomaly detection, and help to deeply understand the performance characteristics of the gyrotron and its potential abnormal behavior. Through the comprehensive analysis of these parameters, gyrotron faults can be more effectively identified and diagnosed to ensure its stable operation.

[0071] Step 1.2 Initialize the feature extractor The parameters of the deep neural network and the parameters of the Gaussian mixture model.

[0072] Step 2: Preheating training: Use the feature extractor to perform characterization learning on the gyrotron test data, obtain the potential features of the data, and continuously determine whether the preheating training is completed. If not, continue to optimize the learning objectives. If completed, proceed to step 3.

[0073] Step 2.1: Perform warm-up training at the beginning of training. Assume that the gyrotron test data set There is no abnormal data in the feature extractor. By assuming a gyrotron test data set The feature extractor is trained without abnormal data. The basic patterns and features of the gyrotron test data can be learned, thus laying a good foundation for the subsequent detection of abnormal data points. This process not only helps to improve the convergence speed of the model, but also enhances its sensitivity to normal data features, providing reliable initial parameter settings for subsequent fault identification tasks. The warm-up training is set to end after 30% of the training rounds are completed. Of course, the preset value of the completion can be set as needed. Before the end of the warm-up training, only the deep neural network is optimized for parameter learning; after the warm-up training is completed, in the following step 3, the parameters of the deep neural network and the Gaussian mixture model are optimized at the same time.

[0074] Feature Extractor It is composed of a multi-layer linear neural network, such as Figure 2 As shown in the figure, it mainly includes three core parts, namely the shallow feature extraction layer, the deep feature extraction layer and the prediction classification layer.

[0075] The shallow feature extraction layer includes a first linear layer, a first ReLU activation function, and a second linear layer connected in sequence, the input end of the first linear layer inputs the gyrotron test data, and the output end of the second linear layer is connected to the deep feature extraction layer, and the shallow feature extraction layer is intended to map the gyrotron test data to a high-dimensional potential feature space. This process not only improves the expressibility of the data, but also provides a solid foundation for subsequent deep feature extraction.

[0076] The deep feature extraction layer includes two important units: a weighted quantization module and a self-gating module.

[0077] The reweighting module includes a third linear layer, a pooling layer, a first multiplication unit, a second ReLU activation function, and a first addition unit connected in sequence, wherein the output end of the second linear layer is respectively connected to the input ends of the third linear layer, the first addition unit, and the first multiplication unit, the output end of the first multiplication unit is connected to the second ReLU activation function, and the output end of the first addition unit is connected to the self-gating module element. The reweighting module adaptively allocates weights in the channel dimension to the features learned by the shallow feature extraction layer, so that the network can focus more on those feature channels that have a significant impact on the final result. Through this dynamic weight adjustment mechanism, the expression ability of important features can be effectively improved, while suppressing the interference of noise and redundant information, thereby optimizing the selection and representation of features.

[0078] The self-gating module includes a fourth linear layer, a third ReLU activation function, a second multiplication unit, a fourth ReLU activation function, and a second addition unit connected in sequence, wherein the output end of the first addition unit is simultaneously connected to the fourth linear layer, the second multiplication unit, and the second addition unit, the output end of the second multiplication unit is connected to the fourth ReLU activation function, and the output end of the second addition unit is connected to the prediction classification layer. By introducing a gating strategy, the learned features are self-suppressed to reduce the flow of information that is not important to the final result during the network learning process. The implementation of self-gating enables the network to more flexibly adjust the information flow during feature learning, effectively filter out unnecessary interference, and ensure that only the most relevant information is passed to subsequent layers.

[0079] Through the above two methods, the deep feature extraction layer can more accurately learn the potential feature representation of the gyrotron test data, enhancing the model's ability to capture complex data patterns.

[0080] The prediction classification layer includes a fifth linear layer, which uses the high-dimensional gyrotron potential feature representation learned through the deep feature extraction layer to make the final prediction and output the result of gyrotron fault identification. In this layer, the model comprehensively analyzes the results of feature extraction, combines the knowledge accumulated in the deep learning process, generates a reliable judgment on the gyrotron fault state, and obtains the potential features of the gyrotron test data after learning. : .

[0081] Step 2.2 Supervised Optimization of Feature Extractor The parameters of the deep neural network, the optimization objective loss function in this process for:

[0082] ,

[0083] in, is the hyperparameter weight factor; is the prediction classification loss defined; is the contrastive learning loss defined; It is the result of fault prediction by multi-layer perceptron on the potential features of gyrotron test data; It is the working state of the gyrotron, which is the real data label of the gyrotron test data here; is the latent feature of the gyrotron test data after learning; is the hyperparameter scaling factor used in contrastive learning; and For the positive and negative samples set using contrastive learning, only the sample itself is set as the positive sample, and the rest of the samples are negative samples.

[0084] This process automatically identifies complex patterns and nonlinear relationships in the data through a multi-layer network structure, thereby effectively capturing the key factors related to gyrotron operation failures.

[0085] Step 3: After the warm-up, the Gaussian mixture model is used to analyze the latent features of the learned gyrotron test data. Perform distribution modeling.

[0086] Latent features of gyrotron test data after learning Then, using the Gaussian mixture model This process is done by building the latent features of the learned gyrotron test data. The probability distribution of the data can effectively capture the intrinsic structure and characteristic distribution of the data. By judging the difference between the observed value and the target distribution, the abnormal test data of the gyrotron can be identified. Ideally, the real data label, that is, the working state of the gyrotron, is usually , directly assigned to the discrete latent variable , and then optimize the Gaussian mixture model Parameters and However, due to the gyrotron test data set The collection process of the gyrotron test data set was affected by human factors and environmental lighting, resulting in There is noisy data in , so it is impossible to obtain the true data label. To this end, the feature extractor Latent features of gyrotron data after learning With discrete latent variables To further optimize the Gaussian mixture model Parameters and :

[0087] ,

[0088] ,

[0089] Here, use represents the gyrotron operation state predicted by the model; represents L2 regularization; is the latent feature of the gyrotron test data after learning; are the parameters of the neural network; and Gaussian mixture model The learned data mean and variance; T represents the matrix transpose operation. For this, the Gaussian mixture model is used The representation distribution of gyrotron test data is learned.

[0090] Using Gaussian mixture model The latent features of the extracted gyrotron test data after learning Distribution modeling can establish the probability distribution of data features. This modeling process can not only describe the inherent structure of the data, but also identify the relationship between data points.

[0091] Step 4: Use Gaussian mixture model Calculate the distribution confidence of gyrotron test data and identify abnormal gyrotron test data.

[0092] Step 4.1 Complete the Gaussian mixture model Parameters and After optimization, the latent features of the gyrotron test data after learning are calculated. The posterior probability of is used to obtain a confidence score to identify the abnormal test data of the gyrotron. The calculation formula is:

[0093]

[0094] in, and All have been L2 regularized; Indicates the gyrotron parameters related to possible gyrotron failure at a given In the case of Gaussian mixture model predict Belong to The probability of a class, here there are only two categories: abnormal and normal; is the latent feature of the gyrotron test data after learning; and They are the mean and variance of the data learned by the Gaussian mixture model; T represents the matrix transpose operation.

[0095] Step 4.2 Based on the calculated confidence score, Update to effectively filter out abnormal gyrotron test data from the gyrotron test data set Filtered and excluded from the training step.

[0096] Step 5: Repeat steps 2-4 until the target training function converges.

[0097] Determine whether the target training function converges. If not, repeat steps 2 to 4. During the parameter optimization process, the identification of abnormal test data of the gyrotron is completed. If yes, the final confidence is output and the abnormal data is filtered based on it. The whole process follows the "EM" algorithm, so the convergence of the identification algorithm can be guaranteed.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying abnormal gyrotron test data, characterized in that: include: Step 1: Construct a gyrotron test data set and initialize the parameters of the feature extractor and Gaussian mixture model; Step 2: Perform warm-up training, use the feature extractor to characterize and learn the gyrotron test data, obtain the potential features of the data, and continue to determine whether the warm-up training is completed. If not, continue to optimize the learning objectives. If completed, proceed to step 3. The goal of this stage is to find a better solution for the optimization direction of the feature extractor. To this end, an optimization target loss function is constructed. The feature extractor is a deep neural network, including: A shallow feature extraction layer includes a first linear layer, a first ReLU activation function, and a second linear layer connected in sequence, wherein the gyrotron test data is input into an input end of the first linear layer; A deep feature extraction layer, comprising: a third linear layer, a pooling layer, a first multiplication unit, a second ReLU activation function, and a first addition unit connected in sequence, a fourth linear layer, a third ReLU activation function, a second multiplication unit, a fourth ReLU activation function, and a second addition unit connected in sequence, the output end of the second linear layer is respectively connected to the input ends of the third linear layer, the first addition unit, and the first multiplication unit, the output end of the first multiplication unit is connected to the second ReLU activation function, the output end of the first addition unit is simultaneously connected to the fourth linear layer, the second multiplication unit, and the second addition unit, the output end of the second multiplication unit is connected to the fourth ReLU activation function, and the output end of the second addition unit is connected to the prediction classification layer; The prediction classification layer, including the fifth linear layer, uses the high-dimensional gyrotron potential feature representation learned through the deep feature extraction layer to make a final prediction, output the result of gyrotron fault identification, and obtain the potential features of the learned gyrotron test data; Step 3: After the warm-up is completed, the Gaussian mixture model is used to perform distribution modeling on the potential features of the learned gyrotron test data; Step 4: Use the Gaussian mixture model to calculate the distribution confidence of the gyrotron test data, identify the abnormal gyrotron test data, and filter it from the gyrotron test data set; Step 5. Repeat steps 2-4 until the target loss function converges.

2. A gyrotron abnormal test data identification method as claimed in claim 1, characterized in that: The gyrotron test data set is ,in Indicates the gyrotron parameters related to possible gyrotron failure. represents the working state of the gyrotron, and the Gaussian mixture model is , the distribution of its fitting Depend on and Two parameters determine: ， in, is a discrete latent variable that determines the assignment of observations to Gaussian mixture components; The observation data belongs to The probability of a sub-model; is the latent feature of the gyrotron test data after neural network learning; and They are the mean and standard deviation of the data learned by the Gaussian mixture model; Is to give parameters and Post-prediction The probability distribution expression of .

3. A gyrotron abnormal test data identification method as claimed in claim 1, characterized in that: The step 1 specifically includes: Step 1.1: First, build a gyrotron test dataset ,in Indicates the gyrotron parameters related to possible gyrotron failure. Indicates the working status of the gyrotron. Step 1.2 Initialize the feature extractor The parameters of the deep neural network and the parameters of the Gaussian mixture model.

4. A gyrotron abnormal test data identification method as claimed in claim 3, characterized in that: The step 2 specifically includes: Step 2.1: Assume the gyrotron test data set There is no abnormal data in the feature extractor. The optimization direction is used to find a better solution. After the preset training rounds are completed, the warm-up training ends. Before the warm-up training ends, only the parameter optimization learning of the deep neural network is performed. Step 2.2: Optimize feature extractor The parameters of the deep neural network are optimized as follows: ， in, is the hyperparameter weight factor; is the prediction classification loss defined; is the contrastive learning loss defined; It is the result of fault prediction by multi-layer perceptron on the potential features of gyrotron test data; Indicates the working status of the gyrotron; is the latent feature of the gyrotron test data after learning; is the hyperparameter scaling factor used in contrastive learning; and For the positive and negative samples set using contrastive learning, only the sample itself is set as the positive sample, and the rest of the samples are negative samples.

5. The method for identifying abnormal gyrotron test data according to claim 1, characterized in that: The step 3 specifically includes: In the distribution modeling process, the feature extractor Latent features of gyrotron data after learning With discrete latent variables Perform correlation and then optimize the Gaussian mixture model Parameters and : ， ， in, represents the gyrotron operation state predicted by the model; represents L2 regularization; is the latent feature of the gyrotron test data after learning; are the parameters of the neural network; and Gaussian mixture model Learned data mean and variance; T represents the matrix transpose operation.

6. A gyrotron abnormal test data identification method as claimed in claim 5, characterized in that: The step 4 specifically includes: Step 4.1: Complete the Gaussian mixture model Parameters and After optimization, the latent features of the gyrotron test data after learning are calculated. The posterior probability of , to obtain the confidence score, confidence The calculation formula is: in, and All have been L2 regularized; Indicates the gyrotron parameters related to possible gyrotron failure at a given In the case of Gaussian mixture model predict Belong to The probability of a class, here there are only two categories: abnormal and normal; Step 4.2: Based on the calculated confidence score, Update and remove the gyrotron anomaly test data from the gyrotron test dataset Filtered and excluded from the training step.

7. A gyrotron abnormal test data identification system, characterized in that: Includes the following modules: The test data set module is used to construct the gyrotron test data set and initialize the parameters of the feature extractor and Gaussian mixture model; The warm-up training module is used for warm-up training. It uses the feature extractor to characterize and learn the gyrotron test data, obtains the potential features of the data, and continuously determines whether the warm-up training is completed. If not, it continuously optimizes the learning objectives. If completed, it enters the distribution modeling module. The goal of this stage is to find a better solution for the optimization direction of the feature extractor. To this end, an optimization target loss function is constructed. The feature extractor is a deep neural network, including: A shallow feature extraction layer includes a first linear layer, a first ReLU activation function, and a second linear layer connected in sequence, wherein the gyrotron test data is input into an input end of the first linear layer; A deep feature extraction layer, comprising: a third linear layer, a pooling layer, a first multiplication unit, a second ReLU activation function, and a first addition unit connected in sequence, a fourth linear layer, a third ReLU activation function, a second multiplication unit, a fourth ReLU activation function, and a second addition unit connected in sequence, the output end of the second linear layer is respectively connected to the input ends of the third linear layer, the first addition unit, and the first multiplication unit, the output end of the first multiplication unit is connected to the second ReLU activation function, the output end of the first addition unit is simultaneously connected to the fourth linear layer, the second multiplication unit, and the second addition unit, the output end of the second multiplication unit is connected to the fourth ReLU activation function, and the output end of the second addition unit is connected to the prediction classification layer; The prediction classification layer, including the fifth linear layer, uses the high-dimensional gyrotron potential feature representation learned through the deep feature extraction layer to make a final prediction, output the result of gyrotron fault identification, and obtain the potential features of the learned gyrotron test data; A distribution modeling module is used to perform distribution modeling on the potential features of the learned gyrotron test data using a Gaussian mixture model after the warm-up is completed; A confidence calculation module, used to calculate the distribution confidence of the gyrotron test data using a Gaussian mixture model, identify abnormal gyrotron test data, and filter it from the gyrotron test data set; The target convergence module is used to converge the target loss function.

8. A gyrotron abnormality test data identification system as claimed in claim 7, characterized in that: The distribution modeling module specifically includes: In the distribution modeling process, the feature extractor Latent features of gyrotron data after learning With discrete latent variables Perform correlation and then optimize the Gaussian mixture model Parameters and : ， ， in, represents the gyrotron operation state predicted by the model; represents L2 regularization; is the latent feature of the gyrotron test data after learning; are the parameters of the neural network; and Gaussian mixture model Learned data mean and variance; T represents the matrix transpose operation.

9. A gyrotron abnormality test data identification system as claimed in claim 8, characterized in that: The confidence calculation module includes: Computational unit, used to complete the Gaussian mixture model Parameters and After optimization, the latent features of the gyrotron test data after learning are calculated. The posterior probability of , to obtain the confidence score, confidence The calculation formula is: in, and All have been L2 regularized; Indicates the gyrotron parameters related to possible gyrotron failure at a given In the case of Gaussian mixture model predict Belong to The probability of a class, here there are only two categories: abnormal and normal; Update unit, used to update the gyrotron test dataset according to the calculated confidence score Update to include gyrotron abnormality test number.

Images