Accident prediction model generation method and nuclear power plant accident diagnosis method

Abstract

The invention provides an accident prediction model generation method and a nuclear power plant accident diagnosis method, and the method comprises the steps: obtaining a plurality of groups of sample training data, the sample training data comprises sample operation data of the nuclear power plant in different historical periods and corresponding sample accident types under the sample operation data; training a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models used for predicting target accident types corresponding to the target operation data of the nuclear power plant in any time period; and knowledge distillation is carried out based on the plurality of different intermediate sample models to obtain a student model, and the student model is used as the accident prediction model. And the nuclear power plant accident type can be quickly identified under the condition of limited computing resources through the accident prediction model, so that the accuracy and the computing efficiency of accident diagnosis are improved.

Description

Technical Field

[0001] The present disclosure relates to the field of nuclear power plant data processing, and particularly to a method for generating an accident prediction model and a method for diagnosing nuclear power plant accidents. Background Art

[0002] In recent years, artificial intelligence (AI) technology has been increasingly widely applied in various industries, especially achieving remarkable results in improving efficiency, productivity, and decision-making capabilities. The nuclear power industry has also gradually focused on using AI technology to enhance the safety and reliability of nuclear power plants.

[0003] However, due to the highly complex operating environment of nuclear power plants and the need to comply with strict safety regulations, obtaining sufficient real data for training AI models has become a major challenge; at the same time, nuclear power plant accidents are relatively rare, and data collection is difficult, which makes most AI research mainly rely on simulated data; moreover, there is a certain gap between simulated data and actual operating conditions, resulting in the performance of the trained AI models in real scenarios may not fully guarantee the safety and reliability of nuclear power plants. Summary of the Invention

[0004] The technical problem to be solved by the present disclosure is to overcome the defect in the prior art that the nuclear power plant cannot be accurately diagnosed, and to provide a method for generating an accident prediction model and a method for diagnosing nuclear power plant accidents.

[0005] The present disclosure solves the above technical problem through the following technical solutions:

[0006] According to a first aspect of the present disclosure, there is provided a method for generating an accident prediction model, where the accident prediction model is used to predict accident prediction information corresponding to actual operation data of a nuclear power plant within a preset time period;

[0007] The generating method includes:

[0008] Obtain a plurality of groups of sample training data, where the sample training data includes sample operation data of the nuclear power plant in different historical periods and corresponding sample accident types under the sample operation data;

[0009] Train a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting target accident types corresponding to the nuclear power plant under target operation data at any time period;

[0010] Wherein, different teacher models process the sample training data by adopting different processing methods to obtain different intermediate sample models;

[0011] Perform knowledge distillation based on several different intermediate sample models to obtain a student model, and the student model serves as the accident prediction model.

[0012] Optionally, the teacher model includes at least one of a basic teacher model, a noise processing teacher model, and a time trend teacher model;

[0013] When the teacher model is the basic teacher model, the steps of obtaining the intermediate sample model include:

[0014] Train a preset training model based on the sample training data to obtain the basic teacher model for predicting the corresponding target accident type under any target operation data;

[0015] When the teacher model is the noise processing teacher model, the steps of obtaining the intermediate sample model include:

[0016] Add Gaussian noise with different standards to the sample training data to obtain first sample training data;

[0017] Train the preset training model based on the first sample training data to obtain the noise processing teacher model for predicting the corresponding one under any target operation data;

[0018] When the teacher model is the time trend teacher model, the steps of obtaining the intermediate sample model include:

[0019] Perform time window smoothing processing on the sample training data based on different time window requirements to obtain second sample training data;

[0020] Train the preset training model based on the second sample training data to obtain the time trend teacher model for predicting the corresponding one under any target operation data.

[0021] Optionally, the steps of performing knowledge distillation based on several different intermediate sample models to obtain a student model specifically include:

[0022] Generate soft label losses corresponding to different teacher models based on a preset function;

[0023] Preprocess the obtained sample training data to generate hard label losses corresponding to different teacher models;

[0024] Generate a distillation loss function based on the soft label loss and the hard label loss;

[0025] Perform knowledge distillation on different intermediate sample models based on the distillation loss function to obtain the student model;

[0026] and / or

[0027] After the step of obtaining a plurality of sets of sample training data, before the step of training a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting the target accident type corresponding to the target operation data of the nuclear power plant at any time period, the generation method further includes:

[0028] Processing the sample training data by using a preset data processing method;

[0029] Wherein, the preset data processing method includes a normalization method and / or a sliding window technique.

[0030] According to a second aspect of the present disclosure, there is provided a method for detecting accidents in a nuclear power plant, the detection method including:

[0031] Obtaining the actual operation data of the nuclear power plant within a preset time period;

[0032] Inputting the actual operation data into the accident prediction model generated by the generation method of the accident prediction model according to the first aspect of the present disclosure to obtain accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

[0033] Optionally, after the step of obtaining the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period, the detection method further includes:

[0034] Updating the accident prediction model based on the actual operation data and the corresponding accident prediction information;

[0035] and / or

[0036] The accident prediction information at least includes accident type information.

[0037] According to a third aspect of the present disclosure, there is provided a generation system for an accident prediction model, the generation system including:

[0038] A first acquisition module, configured to acquire a plurality of sets of sample training data, where the sample training data includes the sample operation data of the nuclear power plant in different historical time periods and the corresponding sample accident types under the sample operation data;

[0039] A first training module, configured to train a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting the target accident type corresponding to the target operation data of the nuclear power plant at any time period;

[0040] Among them, different ones of the teacher models process the sample training data by adopting different processing methods to obtain different intermediate sample models;

[0041] A knowledge distillation module, which is used to perform knowledge distillation based on a number of different intermediate sample models to obtain a student model, and the student model serves as the accident prediction model.

[0042] Optionally, the teacher model includes at least one of a basic teacher model, a noise processing teacher model, and a time trend teacher model;

[0043] When the teacher model is the basic teacher model, the first training module is used to train a preset training model based on the sample training data to obtain the basic teacher model for predicting the corresponding target accident type under any target operation data;

[0044] When the teacher model is the noise processing teacher model, the first training module is used to add Gaussian noise with different standards to the sample training data to obtain first sample training data;

[0045] Based on the first sample training data, the preset training model is trained to obtain the noise processing teacher model for predicting the corresponding one under any target operation data;

[0046] When the teacher model is the time trend teacher model, the first training module is used to perform time window smoothing processing on the sample training data based on different time window requirements to obtain second sample training data;

[0047] Based on the second sample training data, the preset training model is trained to obtain the time trend teacher model for predicting the corresponding one under any target operation data.

[0048] Optionally, the knowledge distillation module is used to generate soft label losses corresponding to different teacher models based on a preset function;

[0049] Preprocess the obtained sample training data to generate hard label losses corresponding to different teacher models;

[0050] Based on the soft label loss and the hard label loss, a distillation loss function is generated;

[0051] Based on the distillation loss function, knowledge distillation is performed on different intermediate sample models to obtain the student model;

[0052] and / or,

[0053] The generation system of the accident prediction model further includes a data preprocessing module, which is configured to, after the step of obtaining a plurality of groups of sample training data, process the sample training data by using a preset data processing method before the step of training a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting the target accident type corresponding to the nuclear power plant under the target operation data at any time period;

[0054] Wherein, the preset data processing method includes a normalization method and / or a sliding window technique.

[0055] According to a fourth aspect of the present disclosure, there is provided a detection system for nuclear power plant accidents, the detection system including:

[0056] A second acquisition module, configured to acquire the actual operation data of the nuclear power plant within a preset time period;

[0057] A detection result acquisition module, configured to input the actual operation data into the accident prediction model generated by the accident prediction model generation method according to the first aspect of the present disclosure to obtain accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

[0058] Optionally, the detection system for nuclear power plant accidents includes a model update module, which is configured to, after the step of obtaining the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period, update the accident prediction model based on the actual operation data and the corresponding accident prediction information;

[0059] And / or,

[0060] The accident prediction information at least includes accident type information.

[0061] According to a fifth aspect of the present disclosure, there is provided an electronic device, including a memory, a processor, and a computer program stored on the memory and configured to run on the processor, where when the processor executes the computer program, it implements the accident prediction model generation method according to the first aspect of the present disclosure, and / or the nuclear power plant accident detection method according to the second aspect of the present disclosure.

[0062] According to a sixth aspect of the present disclosure, there is provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the accident prediction model generation method according to the first aspect of the present disclosure, and / or the nuclear power plant accident detection method according to the second aspect of the present disclosure.

[0063] According to a seventh aspect of the present disclosure, there is provided a computer program product including a computer program which, when executed by a processor, implements the method for generating an accident prediction model described in the first aspect of the present disclosure, and / or the method for detecting a nuclear power plant accident described in the second aspect of the present disclosure.

[0064] On the basis of conforming to common general knowledge in the art, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred examples of the present disclosure.

[0065] The positive and progressive effects of the present disclosure are as follows:

[0066] The method for generating an accident prediction model provided by the present disclosure is based on the knowledge distillation technology, transfers the knowledge of multiple teacher models to the student model, and thus obtains an accident prediction model capable of quickly identifying the types of nuclear power plant accidents. By transferring the knowledge of a large and complex model (teacher model) to a smaller and more lightweight student model through knowledge distillation, the computing efficiency and adaptability of the student model are improved without significantly losing performance. In addition, in the method for generating an accident prediction model provided by the present disclosure, not only can the gap between simulation data and real-scene data be bridged, but also the response speed and stability of the model can be improved in an environment with limited computing resources, which is helpful for the actual deployment and online update of the nuclear power plant accident diagnosis system;

[0067] In the method for diagnosing a nuclear power plant accident provided by the present disclosure, real-time and efficient diagnosis of nuclear power plant accidents is realized based on the accident prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic flowchart of the method for generating an accident prediction model provided in Embodiment 1 of the present disclosure;

[0069] Figure 2 It is a schematic flowchart of the process for performing knowledge distillation provided in Embodiment 1 of the present disclosure;

[0070] Figure 3 It is a schematic flowchart of the method for diagnosing a nuclear power plant accident provided in Embodiment 1 of the present disclosure during the real-time process;

[0071] Figure 4 It is a schematic flowchart of the process for constructing a teacher model provided in Embodiment 1 of the present disclosure;

[0072] Figure 5 It is a schematic flowchart of the knowledge distillation process provided in Embodiment 1 of the present disclosure;

[0073] Figure 6 It is a schematic flowchart of the method for diagnosing a nuclear power plant accident provided in Embodiment 2 of the present disclosure;

[0074] Figure 7Schematic diagram of the system for generating the accident prediction model provided in Embodiment 3 of the present disclosure;

[0075] Figure 8 Schematic diagram of the diagnostic system for nuclear power plant accidents provided in Embodiment 4 of the present disclosure;

[0076] Figure 9 Schematic diagram of the electronic device provided in Embodiment 5 of the present disclosure. Detailed implementation manners

[0077] The present disclosure will be further described below by way of embodiments, but the present disclosure is not limited thereto within the scope of the described embodiments.

[0078] In the embodiments of the present disclosure, prefix words such as "first" and "second" are only used to distinguish different described objects, and have no limiting effect on the position, order, priority, quantity, content, etc. of the described objects. The use of ordinal numbers and other prefix words for distinguishing described objects in the embodiments of the present disclosure does not constitute a limitation on the described objects. For the description of the described objects, reference may be made to the claims or the description in the context of the embodiments. It should not constitute an unnecessary limitation due to the use of such prefix words. In addition, in the description of this embodiment, unless otherwise specified, the meaning of "a plurality" is two or more.

[0079] Embodiment 1

[0080] In this embodiment, a method for generating an accident prediction model is provided. The accident prediction model is used to predict accident prediction information corresponding to the actual operation data of a nuclear power plant within a preset time period;

[0081] As Figure 1 shown, the generation method includes:

[0082] S11: Obtain a plurality of groups of sample training data, where the sample training data includes sample operation data of the nuclear power plant in different historical time periods and corresponding sample accident types under the sample operation data;

[0083] S12: Train a number of different teacher models based on the sample training data to obtain a number of intermediate sample models for predicting the target accident type corresponding to the nuclear power plant under the target operation data at any time period;

[0084] Among them, different teacher models are based on different processing methods for processing the sample training data to obtain different intermediate sample models;

[0085] S13: Perform knowledge distillation based on a number of different intermediate sample models to obtain a student model, and the student model is used as the accident prediction model.

[0086] The accident prediction model in this embodiment also generates corresponding probability distributions for common accident types in nuclear power plants (such as Loss of Coolant Accident (LOCA), Steam Generator Tube Rupture (SGTR), Main Steam Line Break (MSLB), etc.). During the diagnosis process, if the probability value of the Loss of Coolant Accident (LOCA) is the highest, the model determines that accident as LOCA; similarly, if the probability of Steam Generator Tube Rupture (SGTR) is the highest, the current situation is determined as SGTR.

[0087] In one implementation, during the accident diagnosis process, the probability distribution inferred and generated by the accident prediction model is used to determine the specific type of the accident. The accident prediction model calculates the probability of each accident category and selects the category with the highest probability as the final prediction result. The process formula for accident type judgment is:

[0088]

[0089] Where, is the predicted accident type, that is, the accident category finally output by the model. is the predicted probability of the student model for accident category c, and S is the probability distribution or sub-vector output by the "teacher model". That is, S is a vector composed of the predicted values of the teacher model for each category, representing the predicted probability of the model for each category. is to solve the category with the highest probability, that is, to select the accident category that the student model believes is most likely to occur.

[0090] In a specific example, assume that the operation data A corresponds to the output of accident categories a1, a2, and a3. Among them, the predicted probability corresponding to accident category a1 is 20%, the predicted probability corresponding to accident category a2 is 30%, and the predicted probability corresponding to accident category a3 is 50%. Then the accident category corresponding to the operation data A is a3.

[0091] In this embodiment, after the accident prediction model identifies the corresponding accident type of the nuclear power plant, it will generate a specific diagnosis result and output it.

[0092] The output content includes the predicted accident type, the probability distribution values of various accidents, the possible accident development trends, etc. This result will be transmitted to the control center of the nuclear power plant in real time and displayed to the operator through a visualization system so that they can quickly understand the current accident situation.

[0093] In addition, outputting the specific accident type can also help the operator quickly identify the nature of the accident, such as whether it is a Loss of Coolant Accident or a Main Steam Line Break, etc.

[0094] In this embodiment, the probability distribution values of the accident prediction model for different accident types will also be displayed for the operator to judge the reliability and severity of the accident type. A high probability value usually means a high confidence of the model in this accident type. Further, if the accident shows continuous changes or a certain trend, the diagnosis system will further provide trend prediction to help the operator formulate preventive measures in advance. For example, if the pressure data rises rapidly within a short period of time, the system will prompt that there may be a potential risk of coolant loss.

[0095] The decision support information will provide corresponding countermeasure suggestions according to the type and severity of the accident. The operator can make decisions quickly based on this to reduce the impact of the accident on the nuclear power plant. At the same time, the diagnosis process and results will be recorded for subsequent analysis and optimization of the diagnosis model.

[0096] Further, in the initial stage of a nuclear power plant accident, the data may be incomplete or vary greatly. A dynamic update mechanism is also set in the accident type judgment process. That is, the input data is sampled multiple times in the time dimension, and real-time judgment is performed at each time step. If the predicted probability of a certain type of accident is continuously at the highest for multiple times, the diagnosis system will determine this result as the final judgment result, thereby reducing the impact of instantaneous data fluctuations on the diagnosis result.

[0097] The method for generating the accident prediction model provided by the present disclosure is based on the knowledge distillation technology, which transfers the knowledge of multiple teacher models to the student model, so as to obtain an accident prediction model that can quickly identify the accident types of nuclear power plants. By transferring the knowledge of a large and complex model (teacher model) to a smaller and more lightweight student model through knowledge distillation, the computational efficiency and adaptability of the student model are improved without significant loss of performance. In addition, the method for generating the accident prediction model provided by the present disclosure can not only bridge the gap between simulation data and real-scene data, but also improve the response speed and stability of the model in an environment with limited computing resources, thereby contributing to the actual deployment and online update of the nuclear power plant accident diagnosis system.

[0098] The teacher model in this embodiment includes at least one of a basic teacher model, a noise processing teacher model, and a time trend teacher model;

[0099] When the teacher model is a basic teacher model, the steps of obtaining the intermediate sample model include:

[0100] Training a preset training model based on sample training data to obtain a basic teacher model for predicting the corresponding target accident type under any target operation data;

[0101] When the teacher model is a noise processing teacher model, the steps of obtaining the intermediate sample model include:

[0102] Add Gaussian noise with different standards to the sample training data to obtain the first sample training data;

[0103] Based on the first sample training data, train the preset training model to obtain a noise processing teacher model for predicting the corresponding data under any target operating data;

[0104] When the teacher model is a time trend teacher model, the steps to obtain the intermediate sample model include:

[0105] Based on different time window requirements, perform time window smoothing processing on the sample training data to obtain the second sample training data;

[0106] Based on the second sample training data, train the preset training model to obtain a time trend teacher model for predicting the corresponding data under any target operating data.

[0107] In a specific implementation, the teacher model is used to learn different characteristics in the data of the full - scope simulator of the nuclear power plant. The present disclosure constructs multiple teacher models to capture the data characteristics under different operating states of the nuclear power plant. Among them, the teacher models include a basic teacher model, a noise processing teacher model, and a time trend teacher model.

[0108] (1) Basic teacher model (Teacher 1): Adopt a bidirectional long - short - term memory network (Bi - LSTM) to learn the basic features of the nuclear power plant simulation data. The training objective of this model is to minimize the classification loss function:

[0109]

[0110] Among them, represents the loss function of the basic teacher model, which is used to evaluate the classification accuracy of the model on the training data set. N represents the number of samples, indicating the total number of samples used in the training process. represents the true label (hard label) of the i - th sample, usually represented by 0 or 1. represents the model prediction probability of the i - th sample, indicating the probability that the model believes the sample belongs to a certain category.

[0111] (2) Noise processing teacher model (Teacher 2): Based on the basic teacher model, by adding Gaussian noise with different standard deviations to the training data, the robustness of the model to noise is enhanced. Its training objective is to minimize the following loss function:

[0112]

[0113] Among them, Represents the loss function of the noise processing teacher model, which is used to improve the robustness of the model in a noisy environment.

[0114] Represents the regularization parameter, which is used to control the weight of the noise term in the loss function.

[0115] Represents the Gaussian noise added to the training data to simulate the uncertainty in the real environment.

[0116] M represents the number of noise samples, that is, the number of samples with noise added during the training process.

[0117] (3) Time Trend Teacher Model (Teacher 3): Preprocesses the data using moving averages with different time window sizes (such as 60s, 120s, 180s) so that the model can better capture the time-dependent characteristics. The loss function of this model is:

[0118]

[0119] Among them, is the loss function of the time trend teacher model, which is used to capture the trends and dependencies in the time series data. is the time smoothing coefficient, which is used to adjust the influence of the time trend term on the loss function. T is the total number of steps in the time series, representing the span of the data in the time dimension. is the model prediction value at the t-th moment, representing the prediction result of the model for the current moment. is the true value at the (t - 1)-th moment, representing the data of the previous moment.

[0120] This disclosure shows strong robustness in noisy data and time-dependent data by processing the input data and combining the knowledge distilled from multiple teacher models, and can identify the early features of accident types and generate corresponding probability distributions. By using multiple teacher models to learn the nuclear power plant simulation data and transferring the knowledge of complex models to lightweight student models, the student models can still effectively perform accident diagnosis under limited computing resources.

[0121] Such as Figure 2 shown, the steps of obtaining the student model by knowledge distillation based on several different intermediate sample models specifically include:

[0122] S21: Generate soft label losses corresponding to different teacher models based on a preset function;

[0123] S22: Preprocess the obtained sample training data to generate hard label losses corresponding to different teacher models;

[0124] S23: Generate a distillation loss function based on the soft label loss and the hard label loss;

[0125] S24: Perform knowledge distillation on different intermediate sample models based on the distillation loss function to obtain the student model.

[0126] In a specific implementation, the output of the teacher model for the input data is usually a probability distribution. A smoother soft label is generated by increasing the distillation temperature T:

[0127]

[0128] where, is the softmax function after temperature adjustment (a function that converts a real number vector into a probability distribution), which is used to generate soft labels.

[0129] is the predicted logit value (logistic regression value) of the model for class i, representing the score before the softmax transformation. T is the distillation temperature, which is used to control the smoothness of the generated probability distribution. The higher the temperature, the smoother the generated probability distribution.

[0130] The student model is trained by combining two loss functions, namely the hard loss and the soft loss. These two loss functions come from the distilled knowledge. The hard target loss is calculated through the true labels that are consistent with the basic teacher model. The student model directly learns to predict the correct class labels through the hard loss.

[0131] Taking the soft labels of the teacher model as the learning target of the student model, and combining the hard labels to form the distillation loss function:

[0132]

[0133] where, represents the distillation loss function, which is used to guide the student model to learn the prediction distribution of the teacher model. is the weight coefficient, which is used to balance the influence of the soft label loss and the hard label loss. is the Kullback-Leibler divergence (also known as relative entropy, or information divergence, which is an asymmetric measure of the difference between two probability distributions), which is used to measure the difference between two probability distributions. is the soft label probability distribution generated by the teacher model, is the soft label probability distribution generated by the student model. is the predicted probability of the student model for the i-th sample.

[0134] In this embodiment, the student model is trained by a distillation loss function so that it can learn the rich knowledge in the teacher model and improve its performance in the actual environment.

[0135] Furthermore, the student model realizes lightweight design through knowledge distillation and can quickly respond in the accident diagnosis of nuclear power plants. Its specific optimization steps include architecture simplification, temperature adjustment, and model deployment.

[0136] (1) Architecture simplification

[0137] The student model adopts a three-layer bidirectional long short-term memory network (Bi-LSTM), reducing the number of layers and hidden units by 50% compared to the teacher model. Its loss function is:

[0138]

[0139] where is the total loss function of the student model, which is used to optimize the performance of the student model during training. is the loss balance coefficient, which is used to adjust the weights of the distillation loss and the task loss. is the task loss of the student model, also known as the cross-entropy loss, which is used to ensure that the student model can correctly classify.

[0140] (2) Adjust the distillation temperature

[0141] To improve the imitation effect of the student model on the prediction results of the teacher model, by optimizing the distillation temperature T, the smoothness of the soft labels and the learning ability of the student model reach the optimal balance.

[0142] (3) Model deployment

[0143] Deploy the trained student model to the nuclear power plant accident diagnosis system. The student model can quickly identify the accident type under limited computing resources and achieve online update.

[0144] Knowledge distillation simplifies the architecture of the student model. Its parameter quantity is significantly lower than that of the teacher model, and the time required for training and inference is also significantly reduced. With this lightweight design, the student model can stably operate in an environment with limited computing resources such as a nuclear power plant, supporting online update and rapid deployment. Through the process of knowledge distillation, the student model not only learns the output information of the teacher model but also can extract the decision-making basis of the teacher model.

[0145] Through knowledge distillation, the student model in the present disclosure can better adapt to the data distribution in real scenarios, improve the diagnostic accuracy under real conditions, and through the distillation technology, successfully compress the knowledge of the teacher model into a lightweight student model, significantly reducing the computational resource requirements while maintaining high performance. In addition, in the teacher model, different types of teacher models are trained for the noise and time variation trends in the data, and through knowledge distillation, the student model is enabled to handle these changes, thereby enhancing the robustness and stability of the model.

[0146] In this embodiment, after the step of obtaining a plurality of groups of sample training data, before the step of training a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting the target accident type corresponding to the target operation data of the nuclear power plant at any time period, the generation method further includes:

[0147] Processing the sample training data by using a preset data processing method;

[0148] Wherein, the preset data processing method includes a normalization method and / or a sliding window technique.

[0149] In a specific implementation manner, real-time operation data is obtained from a nuclear power plant, including various state variables such as temperature, pressure, flow rate, and radiation level. These data are collected in real time through a sensor system, reflecting the characteristics of the nuclear power plant in different operating states and serving as the basis for accurate analysis by the diagnostic system.

[0150] Since the dimensions and data ranges of different sensors used to collect nuclear power plant data vary, direct input may cause the model to be overly sensitive to features with relatively large specific values. Therefore, data preprocessing ensures consistent and standardized data input, further improving the accuracy of the model.

[0151] The present disclosure adopts a normalization method from 0 to 1 to map the data into a standardized range to eliminate the scale differences between feature values and make the influence of the model on each feature more balanced. The formula for normalization processing is:

[0152]

[0153] In the preprocessing stage, the time series data can also be further processed by using the sliding window technique. The sliding window technique enables the model to capture the time-dependent relationships in the data during the diagnosis process, such as monitoring the pressure change trend or the temperature rise rate over a period of time to infer whether there are potential accidents. For time series data, a sliding window of size is used for preprocessing to generate input data containing multiple time steps to better capture the dynamic characteristics of accident evolution.

[0154] By preprocessing the acquired data, the accuracy of nuclear power plant accident judgment is further improved.

[0155] The preprocessed data is input into an optimized student model, and through the forward propagation process of the neural network, the probability distribution of various accidents is output. Among them, the model inference formula is as follows:

[0156]

[0157] is the prediction result of the student model for the input data x, representing the classification probability distribution of the student model for the input samples. is the function mapping of the student model, representing the prediction process of the student model. is the parameter set of the student model, representing the weights and biases learned during the training process.

[0158] Next, in combination with an example, the implementation principle of the accident prediction model generation method in this embodiment will be specifically described:

[0159] As Figure 3 shown, the present disclosure proposes a method for generating an accident prediction model, which realizes the efficient diagnosis of nuclear power plant accidents by combining the designs of a teacher model and a student model.

[0160] The architecture for obtaining the accident prediction model provided by the present disclosure includes the following modules: a data processing module, a teacher model module, a knowledge distillation module, a student model module, and a result output module.

[0161] Among them, the data processing module is responsible for receiving real-time status data from nuclear power plant sensors, including temperature, pressure, flow rate, radiation level, etc. The real-time status data can be obtained through nuclear power plant sensors.

[0162] The data processing module preprocesses the input data, mainly including normalization and sliding window processing, to ensure the standardization and consistency of data input. Normalization maps data with different dimensions to the interval from 0 to 1 to eliminate the scale differences between data. The sliding window is used to capture the time series characteristics, enabling the model to better understand the changes in status data over time.

[0163] The teacher model module is used to learn nuclear power plant simulation data through different teacher models, including a basic teacher model, a noise processing teacher model, and a time trend teacher model. Each teacher model captures different characteristics of the data. The basic teacher model learns the main accident characteristics, the noise processing model enhances the system's robustness to noise data, and the time trend model is used to capture the time-dependent relationship of the data.

[0164] In this embodiment, the teacher module at least includes a basic teacher model, a noise processing teacher model, and a time trend teacher model; and different models in the teacher module are trained using the data simulated by the full-scope simulator of the nuclear power plant and / or the data obtained by the data processing module to obtain the target model.

[0165] The knowledge distillation module is used to distill the knowledge learned in the teacher model into a lightweight student model. The specific process includes generating soft labels and calculating the distillation loss. The soft labels are output by the teacher model and generated through the temperature-adjusted softmax function, serving as the learning target for the student model. The distillation loss combines the soft label loss and the hard label loss to achieve knowledge transfer.

[0166] The student model is a lightweight bidirectional long short-term memory network (Bi-LSTM). By learning the knowledge of the teacher model through knowledge distillation, it has efficient real-time diagnosis capabilities. The student model receives the processed data and quickly identifies the accident type.

[0167] The diagnostic result of the student model is transmitted to the result output module. The result output module is used to feedback the judged accident type and its probability distribution value to the control center of the nuclear power plant and display the diagnostic result and development trend through a visualization interface to provide decision-making support for the operators.

[0168] The construction and training process of the teacher model is as Figure 4 shown. The teacher model captures diverse characteristics in the nuclear power plant data through different structures. The specific training steps of each model are as follows:

[0169] 1. Training of the basic teacher model

[0170] The training process of the basic teacher model is as follows: After obtaining the sample training data, no processing is performed on the sample training data. The simulation data is input, feature learning is carried out, the model is optimized, and the model output is performed.

[0171] The basic teacher model is based on the bidirectional long short-term memory network (Bi-LSTM) structure and is used to learn the main accident characteristics from the nuclear power plant simulation data. The training data of the model comes from the nuclear power plant simulation platform (such as the CNS (Central Nervous System) dataset), covering various accident types and their corresponding operating parameters. The training objective of the basic teacher model is to minimize the cross-entropy loss, and the loss function is:

[0172]

[0173] 2. Training of the noise processing teacher model

[0174] The training process of the noise processing teacher model is as follows: After obtaining the sample training data, Gaussian noise is added to the sample training data, the simulation data is input, feature learning is performed, the model is optimized, and the model output is carried out.

[0175] To enhance the stability of the diagnostic system for noisy data, the noise processing teacher model adds Gaussian noise with different standard deviations (such as 0.1, 0.2, 0.3, etc.) to the training data to simulate the fluctuations and anomalies in the actual operation data. The loss function of the noise processing model contains a regularization term to limit the impact of noise on the model:

[0176]

[0177] 3. Training of the time trend teacher model

[0178] The training process of the time trend teacher model is as follows: After obtaining the sample training data, time window smoothing is added to the sample training data, the simulation data is input, feature learning is performed, the model is optimized, and the model output is carried out.

[0179] The time trend teacher model is used to capture the time dependence of the data. By performing moving average processing through different time windows (such as 60 seconds, 120 seconds, and 180 seconds), the model can perceive the time trend. The loss function of this model includes a time smoothing term:

[0180]

[0181] As Figure 5 shown, the knowledge distillation process transfers the knowledge of the teacher model to the student model through the distillation loss function. The specific steps are as follows:

[0182] 1. Soft label generation

[0183] The soft labels in the knowledge distillation process are output by the teacher models (including the basic teacher model, the noise teacher model, and the time trend teacher model) and generated by the softmax function adjusted by the temperature parameter. The specific formula is:

[0184]

[0185] 2. Distillation loss calculation

[0186] The distillation loss function includes soft label loss and hard label loss, and its definition is:

[0187]

[0188] IV. Diagnostic process of the student model

[0189] After the student model undergoes knowledge distillation training, it can achieve real-time accident diagnosis in a nuclear power plant. The specific diagnosis steps are as follows:

[0190] 1. Data collection and preprocessing

[0191] Obtain real-time data from nuclear power plant sensors, perform normalization processing to eliminate the differences in feature scales. In addition, use a sliding window process to convert the data into a time series input, enhancing the student model's understanding of time-dependent characteristics.

[0192] 2. Model inference

[0193] The preprocessed data is input into the student model. Through forward propagation, the student model makes probability predictions for different accident types, generating the probability distributions of various accidents:

[0194]

[0195] 3. Accident type judgment

[0196] The student model determines the accident type with the highest probability based on the probability distribution as the prediction result:

[0197]

[0198] 4. Result output and decision support

[0199] The diagnosis result of the student model is output to the nuclear power plant control center and displayed on a visualization interface through the result output module, so that the operator can timely grasp the accident situation. The result output includes the following information:

[0200] Accident type: Displays the accident type predicted by the model, such as Loss of Coolant Accident (LOCA), Steam Generator Tube Rupture (SGTR), or Main Steam Line Break (MSLB).

[0201] Probability distribution: Displays the predicted probability values of the model for various accidents, helping the operator understand the severity of the accident and the diagnostic confidence of the system.

[0202] Time trend analysis: During the diagnosis process, the system analyzes the time trend information of the accident type, such as the change rate of key parameters, to judge whether the accident is likely to develop further.

[0203] This information is transmitted to the nuclear power plant control center through the result output module to help the operator quickly respond to the accident situation, take appropriate protective measures, and ensure the safety of the nuclear power plant.

[0204] To cope with changes in the operating environment of nuclear power plants and equipment aging, etc., the student model needs to be updated and optimized regularly. This disclosure also adopts an incremental learning strategy to ensure the accuracy and adaptability of the diagnostic system during long-term operation.

[0205] 1. Knowledge Update

[0206] During the operation of the system, the latest simulation data and historical accident data of nuclear power plants are regularly collected for retraining the teacher model. The new teacher model then transfers knowledge to the student model through the knowledge distillation process to ensure that the student model can make judgments following the latest accident characteristics.

[0207] 2. Incremental Learning

[0208] To reduce the impact of model updates on system operation, this disclosure adopts an incremental learning approach, which only requires fine-tuning with a small amount of new data. The advantage of incremental learning is that there is no need to completely retrain the model, only adjust the characteristics of the new data, ensuring the real-time performance of the model.

[0209] 3. Performance Evaluation and Optimization

[0210] After updating the model, the system will use an offline dataset to evaluate the performance of the updated model to verify whether the student model meets the accuracy and robustness criteria. If the expected diagnostic effect is achieved, the optimized model will be redeployed to the accident diagnostic system of the nuclear power plant to achieve seamless switching.

[0211] Through the above update and optimization mechanism, this disclosure can adapt to changes in the dynamic operating environment of nuclear power plants, further enhancing the long-term stability and reliability of the system.

[0212] Based on knowledge distillation technology, this disclosure transfers the knowledge of multiple teacher models to the student model, combined with the bidirectional long short-term memory network (Bi-LSTM) structure, which can achieve real-time and efficient diagnosis of nuclear power plant accidents.

[0213] The specific technical advantages of this disclosure are as follows:

[0214] 1. High Diagnostic Accuracy

[0215] By introducing multiple teacher models, the student model shows high accuracy in different accident scenarios. Especially when dealing with noisy data and time series data, it can fully understand the characteristics of accident data and improve the diagnostic accuracy.

[0216] 2. High Computational Efficiency

[0217] After knowledge distillation, the student model realizes a lightweight design, reducing the number of network layers and parameters, and can operate efficiently in an environment with limited computing resources, meeting the real-time accident diagnosis requirements of nuclear power plants.

[0218] 3. Strong adaptability and robustness

[0219] The student model of the present disclosure continuously adapts to environmental changes through incremental learning, and with the support of the noise processing teacher model and the time trend teacher model, it has strong robustness and can maintain stable performance in complex environments.

[0220] 4. Visual decision support

[0221] Through the visual display of the result output module, the diagnostic results are friendly and intuitive to operators, which helps to make quick and reasonable response decisions during accidents and improves the safety guarantee ability of nuclear power plants.

[0222] The implementation effect of the present disclosure is remarkable. It provides a practical and reliable diagnostic method for nuclear power plant accidents, which can effectively bridge the gap between simulation data and real data and provides strong technical support for the safe operation of nuclear power plants.

[0223] Embodiment 2

[0224] As Figure 6 shown, in this embodiment, a detection method for nuclear power plant accidents is provided, characterized in that the detection method includes:

[0225] S61: Obtain the actual operation data of the nuclear power plant within a preset time period;

[0226] S62: Input the actual operation data into the accident prediction model generated by the generation method based on the accident prediction model to obtain the accident prediction information corresponding to the actual operation data of the nuclear power plant within the preset time period.

[0227] In the diagnostic method for nuclear power plant accidents provided by the present disclosure, real-time and efficient diagnosis of nuclear power plant accidents is realized based on the accident prediction model.

[0228] In this embodiment, after the step of obtaining the accident prediction information corresponding to the actual operation data of the nuclear power plant within the preset time period, the detection method further includes:

[0229] Update the accident prediction model based on the actual operation data and the corresponding accident prediction information; the accident prediction information in this embodiment at least includes accident type information.

[0230] In this embodiment, after the step of inputting the real-time operation data into the target prediction model to output the real-time accident information of the nuclear power plant, the diagnostic method further includes:

[0231] Update the data of the target prediction model based on the real-time operation data and the real-time accident information.

[0232] In a specific embodiment, due to the complex operating environment of nuclear power plants and the fact that data characteristics change with equipment aging and external environment changes, the student model of the accident diagnosis system needs to be updated and optimized regularly to ensure its adaptability to real-time data. The present disclosure updates and optimizes the model in the following aspects:

[0233] Knowledge update: Retrain the teacher model with newly collected simulation data and historical accident data to obtain the latest knowledge. Subsequently, distill this knowledge back into the student model to ensure that the student model can learn the latest accident characteristics and environmental change information.

[0234] Parameter adjustment: The distillation temperature parameter and loss function weight in the student model may need to be dynamically adjusted over time to adapt to the characteristics of new operating data. For example, in a new round of training, the distillation temperature can be experimentally adjusted to optimize the smoothness of the soft labels, thereby enhancing the generalization ability of the student model.

[0235] Incremental learning: To reduce the impact of model updates on the normal operation of nuclear power plants, the accident diagnosis system uses incremental learning to make minor updates to the model. During the incremental learning process, only a small amount of new data is added to fine-tune the model, rather than retraining the entire model completely, thus ensuring the real-time performance and stability of the diagnosis system.

[0236] Model performance evaluation and optimization: After the model is updated, perform performance evaluation using offline data to ensure that the accuracy and robustness of the updated model meet the set standards. If the performance meets the requirements, deploy the updated student model to the diagnosis system to continuously optimize the real-time accident monitoring and judgment of nuclear power plants.

[0237] By updating and optimizing the target prediction model, the accuracy of the target prediction model for nuclear power plant accident prediction is further improved.

[0238] Embodiment 3

[0239] As Figure 7 shown, in this embodiment, a generation system for an accident prediction model is provided. The generation system includes:

[0240] A first acquisition module 701, configured to acquire a plurality of groups of sample training data, where the sample training data includes sample operation data of a nuclear power plant in different historical periods and corresponding sample accident types under the sample operation data;

[0241] A first training module 702, configured to train a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting the target accident type corresponding to the target operation data of the nuclear power plant at any time period;

[0242] Among them, different teacher models process the sample training data using different processing methods to obtain different intermediate sample models;

[0243] A knowledge distillation module 703 is used to perform knowledge distillation based on a number of different intermediate sample models to obtain a student model, and the student model serves as an accident prediction model.

[0244] The teacher model in this embodiment includes at least one of a basic teacher model, a noise processing teacher model, and a time trend teacher model;

[0245] When the teacher model is a basic teacher model, the first training module is used to train a preset training model based on the sample training data to obtain a basic teacher model for predicting the corresponding target accident type under any target operation data;

[0246] When the teacher model is a noise processing teacher model, the first training module is used to add Gaussian noise with different standards to the sample training data to obtain first sample training data;

[0247] Based on the first sample training data, train a preset training model to obtain a noise processing teacher model for predicting the corresponding target accident type under any target operation data;

[0248] When the teacher model is a time trend teacher model, the first training module is used to perform time window smoothing processing on the sample training data based on different time window requirements to obtain second sample training data;

[0249] Based on the second sample training data, train a preset training model to obtain a time trend teacher model for predicting the corresponding target accident type under any target operation data.

[0250] The knowledge distillation module in this embodiment is used to generate soft label losses corresponding to different teacher models based on a preset function;

[0251] Preprocess the obtained sample training data to generate hard label losses corresponding to different teacher models;

[0252] Based on the soft label loss and the hard label loss, generate a distillation loss function;

[0253] Based on the distillation loss function, perform knowledge distillation on different intermediate sample models to obtain a student model;

[0254] The generation system of the accident prediction model further includes a data preprocessing module 704, which is used to process the sample training data by a preset data processing method before the step of training a number of different teacher models based on the sample training data to obtain a number of intermediate sample models for predicting the target accident type corresponding to the nuclear power plant under the target operation data at any time period;

[0255] Wherein, the preset data processing method includes a normalization method and / or a sliding window technique.

[0256] In the accident prediction model generation system provided by the present disclosure, based on the knowledge distillation technology, the knowledge of multiple teacher models is transferred to the student model, so as to obtain an accident prediction model that can quickly identify the accident types of nuclear power plants. By transferring the knowledge of a large and complex model (teacher model) to a smaller and more lightweight student model through knowledge distillation, the computing efficiency and adaptability of the student model can be improved without significant loss of performance. In addition, in the accident prediction model generation method provided by the present disclosure, it can not only bridge the gap between simulation data and real scenario data, but also improve the response speed and stability of the model in an environment with limited computing resources, thus contributing to the actual deployment and online update of the nuclear power plant accident diagnosis system.

[0257] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present disclosure.

[0258] Embodiment 4

[0259] As Figure 8 shown, in this embodiment, a detection system for nuclear power plant accidents is provided. The detection system includes:

[0260] A second acquisition module 801, configured to acquire the actual operation data of the nuclear power plant within a preset time period;

[0261] A detection result acquisition module 802, configured to input the actual operation data into an accident prediction model generated by the accident prediction model generation method according to the first aspect of the present disclosure, so as to obtain accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

[0262] The detection system for nuclear power plant accidents in this embodiment includes a model update module 803. The model update module 803 is used to update the accident prediction model based on the actual operation data and the corresponding accident prediction information after the step of obtaining the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

[0263] The accident prediction information at least includes accident type information.

[0264] In the nuclear power plant accident diagnosis system provided by the present disclosure, real-time and efficient diagnosis of nuclear power plant accidents is realized based on the accident prediction model.

[0265] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present disclosure.

[0266] Embodiment 5

[0267] Figure 9 FIG. is a schematic structural diagram of an electronic device provided for Embodiment 5 of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method in the above embodiment is implemented. Figure 9 The displayed electronic device 30 is merely an example and should not impose any limitation on the functions and usage scope of the embodiments of the present disclosure.

[0268] As Figure 9 shown, the electronic device 30 can be presented in the form of a general-purpose computing device. For example, it can be a server device. The components of the electronic device 30 may include, but are not limited to: the above-mentioned at least one processor 31, the above-mentioned at least one memory 32, and a bus 33 connecting different system components (including the memory 32 and the processor 31).

[0269] The bus 33 includes a data bus, an address bus, and a control bus.

[0270] The memory 32 may include volatile memory, such as a random access memory (RAM) 321 and / or a cache memory 322, and may further include a read-only memory (ROM) 323.

[0271] The memory 32 may also include a program / utilities 325 having a set (at least one) of program modules 324. Such program modules 324 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include the implementation of a network environment.

[0272] The processor 31 executes various functional applications and data processing by running computer programs stored in the memory 32, such as the methods in the above embodiments of the present disclosure.

[0273] The electronic device 30 may also communicate with one or more external devices 34 (such as a keyboard, a pointing device, etc.). Such communication may be carried out through an input / output (I / O) interface 35. And, the model generation device 30 may also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through a network adapter 36. As Figure 9 shown, the network adapter 36 communicates with other modules of the model generation device 30 through the bus 33. It should be understood that although not shown in the figure, other hardware and / or software modules may be used in combination with the model generation device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (Redundant Array of Independent Disks) systems, tape drives, and data backup storage systems, etc.

[0274] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more of the above-described units / modules may be embodied in one unit / modules. Conversely, the features and functions of one unit / modules described above may be further divided and embodied by multiple units / modules.

[0275] Embodiment 6

[0276] The embodiments of the present disclosure also provide a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the accident prediction model generation method provided in the above Embodiment 1 or the nuclear power plant accident diagnosis method provided in Embodiment 2.

[0277] Among them, the readable storage medium may more specifically include, but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0278] Embodiment 7

[0279] Embodiments of the present disclosure also provide a computer program product, including a computer program which, when executed by a processor, implements the method for generating an accident prediction model provided in Embodiment 1 or the method for diagnosing a nuclear power plant accident provided in Embodiment 2.

[0280] Among them, the program code for executing the computer program product of the present disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on the user device, partially on the user device, executed as an independent software package, partially on the user device and partially on a remote device, or entirely on a remote device.

[0281] Although the specific embodiments of the present disclosure have been described above, those skilled in the art should understand that this is only an example, and the protection scope of the present disclosure is defined by the appended claims. Without departing from the principles and essence of the present disclosure, those skilled in the art can make various changes or modifications to these embodiments, but these changes and modifications all fall within the protection scope of the present disclosure.

Claims

1. A method for generating an accident prediction model, characterized in that: The accident prediction model is used to predict accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period; The generation method comprises: Acquire several groups of sample training data, wherein the sample training data include sample operation data of the nuclear power plant in different historical periods and sample accident types corresponding to the sample operation data; Training a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting target accident types corresponding to target operating data of a nuclear power plant in any time period; Wherein, different teacher models process the sample training data based on different processing methods to obtain different intermediate sample models; Knowledge distillation is performed based on several different intermediate sample models to obtain a student model, and the student model is used as the accident prediction model.

2. The method for generating an accident prediction model according to claim 1, characterized in that: The teacher model includes at least one of a basic teacher model, a noise processing teacher model, and a time trend teacher model; When the teacher model is the basic teacher model, the step of obtaining the intermediate sample model includes: Training a preset training model based on the sample training data to obtain the basic teacher model for predicting the target accident type corresponding to any target operation data; When the teacher model is the noise processing teacher model, the step of obtaining the intermediate sample model includes: Adding Gaussian noise of different standards to the sample training data to obtain first sample training data; Training the preset training model based on the first sample training data to obtain the noise processing teacher model corresponding to any target operating data for prediction; When the teacher model is the time trend teacher model, the step of obtaining the intermediate sample model includes: Based on different time window requirements, the sample training data is subjected to time window smoothing processing to obtain second sample training data; The preset training model is trained based on the second sample training data to obtain the time trend teacher model corresponding to any target operating data for prediction.

3. The method for generating an accident prediction model according to claim 2, characterized in that: The step of performing knowledge distillation based on several different intermediate sample models to obtain a student model specifically includes: Based on a preset function, generating soft label losses corresponding to different teacher models; Preprocessing the obtained sample training data to generate hard label losses corresponding to different teacher models; Generating a distillation loss function based on the soft label loss and the hard label loss; Performing knowledge distillation on different intermediate sample models based on the distillation loss function to obtain the student model; and / or, After the step of obtaining several groups of sample training data, and before the step of training several different teacher models based on the sample training data to obtain several intermediate sample models for predicting the target accident type corresponding to the target operation data of the nuclear power plant in any time period, the generation method further includes: Processing the sample training data using a preset data processing method; Wherein, the preset data processing method includes a normalization method and / or a sliding window technology.

4. A method for detecting a nuclear power plant accident, characterized in that: The detection method comprises: Obtain actual operating data of nuclear power plants within a preset time period; The actual operation data is input into the accident prediction model row generated by the accident prediction model generation method according to any one of claims 1 to 3 to obtain the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

5. The method for detecting a nuclear power plant accident according to claim 4, characterized in that: After the step of obtaining the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period, the detection method further includes: Based on the actual operation data and the corresponding accident prediction information, updating the accident prediction model; and / or, The accident prediction information at least includes accident type information.

6. A system for generating an accident prediction model, characterized in that: The generation system comprises: A first acquisition module is used to acquire several groups of sample training data, wherein the sample training data includes sample operation data of the nuclear power plant in different historical periods and sample accident types corresponding to the sample operation data; A first training module is used to train a plurality of different teacher models based on the sample training data to obtain a plurality of intermediate sample models for predicting a target accident type corresponding to the target operation data of the nuclear power plant in any time period; Wherein, different teacher models process the sample training data based on different processing methods to obtain different intermediate sample models; The knowledge distillation module is used to perform knowledge distillation based on several different intermediate sample models to obtain a student model, and the student model is used as the accident prediction model.

7. A nuclear power plant accident detection system, characterized in that: The detection system comprises: The second acquisition module is used to acquire the actual operation data of the nuclear power plant within a preset time period; A detection result acquisition module is used to input the actual operation data into the accident prediction model row generated by the accident prediction model generation method according to any one of claims 1 to 3, so as to obtain the accident prediction information corresponding to the actual operation data of the nuclear power plant within a preset time period.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, characterized in that: When the processor executes the computer program, the method for generating an accident prediction model as described in any one of claims 1 to 3 and / or the method for detecting a nuclear power plant accident as described in claim 4 or 5 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method for generating an accident prediction model as described in any one of claims 1 to 3, and / or the method for detecting a nuclear power plant accident as described in claim 4 or 5.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the method for generating an accident prediction model as described in any one of claims 1 to 3, and / or the method for detecting a nuclear power plant accident as described in claim 4 or 5.

Images