CN-122020016-A - Nuclear power plant water supply system fault diagnosis method and device, electronic equipment and storage medium

CN122020016ACN 122020016 ACN122020016 ACN 122020016ACN-122020016-A

Abstract

The invention provides a method and a device for diagnosing faults of a water supply system of a nuclear power plant, electronic equipment and a storage medium, and belongs to the technical field of operation safety and state monitoring of the nuclear power plant. The method solves the problem that the traditional diagnosis method is difficult to consider multiple fault categories and unknown fault identification. A fault diagnosis method for a water supply system of a nuclear power plant comprises the steps of collecting multivariable operation parameters of the water supply system of the nuclear power plant, preprocessing the multivariable operation parameters to obtain a parameter sequence, training a multivariable time sequence prediction model based on historical operation parameters only including normal working conditions, inputting the preprocessed parameter sequence under target working conditions into the time sequence prediction model to obtain a predicted parameter value, calculating a difference value and constructing a residual characteristic vector, training a fault diagnosis model by taking the residual characteristic vector as input, and outputting operation state types of the water supply system through the fault diagnosis model. The method is mainly used for fault diagnosis of the water supply system of the nuclear power plant.

Inventors

ZHANG BOWEN
LAN NING
Gou jinlan
Xiao Ken
CHENG SHOUYU

Assignees

哈尔滨工程大学
中国船舶集团有限公司第七一九研究所

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (10)

1. The fault diagnosis method for the water supply system of the nuclear power plant is characterized by comprising the following steps of: Collecting multivariable operation parameters of a water supply system of the nuclear power plant under a target working condition and preprocessing the multivariable operation parameters to obtain a preprocessed parameter sequence; training a multivariate time series prediction model for predicting future time parameter values from a sequence of historical parameters based on historical operating parameters comprising only normal operating conditions; inputting the preprocessed parameter sequence under the target working condition into the time sequence prediction model to obtain a predicted parameter value, calculating the difference value between the actual parameter value and the predicted parameter value, and constructing a multi-variable multi-time-step residual characteristic vector; And training a fault diagnosis model by taking the residual characteristic vector as input, and outputting the operation state type of the water supply system through the fault diagnosis model, wherein the operation state type comprises a normal working condition, a known fault working condition and an unknown fault working condition.
2. The method for diagnosing a fault in a feedwater system of a nuclear power plant according to claim 1, wherein said multivariate operating parameters include nuclear power, electrical power, feedwater pressure, feedwater flow, steam generator level and temperature, stabilizer pressure and level, reactor coolant inlet and outlet temperatures and average temperature, and said preprocessing includes time alignment, outlier rejection, denoising and normalization, and employs a sliding time window to construct an input sample sequence.
3. The method for diagnosing faults of a water supply system of a nuclear power plant according to claim 1, wherein the multivariate time sequence prediction model is a deep learning model based on a recurrent neural network and comprises a long short term memory network (LSTM) or a variant thereof, and in the model training process, the preprocessed normal working condition parameters are divided into a training set and a verification set, and iterative training is conducted based on a mean square error loss function until the prediction error on the verification set converges to a preset threshold.
4. The method for diagnosing faults of a water supply system of a nuclear power plant according to claim 1, wherein the residual feature vector is constructed in a mode that differences of actual parameter values and predicted parameter values are calculated according to parameter dimensions and time dimensions to obtain a multi-variable residual sequence, the residual sequences of a plurality of continuous time steps are spliced according to preset rules to form feature vectors with fixed lengths, and the differences are original differences or absolute differences or normalized differences.
5. The method for diagnosing faults of a water supply system of a nuclear power plant according to claim 1, wherein the fault diagnosis model is obtained by training a supervised learning algorithm based on a tree model, the supervised learning algorithm comprises a decision tree algorithm, a random forest algorithm or a gradient lifting tree algorithm, and the unknown fault condition is obtained by marking an abnormal sample caused by other system faults as an unknown fault category in a training stage.
6. The method for diagnosing faults of a water supply system of a nuclear power plant according to claim 1, wherein the target working conditions comprise a steady-state working condition and a transient working condition, independent multivariable time sequence prediction models and fault diagnosis models are respectively built for the two working conditions, and a framework of the fault diagnosis models is shared, wherein the fault diagnosis model corresponding to the steady-state working condition adopts a decision tree algorithm, and the fault diagnosis model corresponding to the transient working condition adopts an integrated learning algorithm.
7. The method for diagnosing faults of a water supply system of a nuclear power plant according to claim 1, wherein the multivariate time sequence prediction model corresponding to a steady-state working condition is a one-dimensional time sequence prediction network based on a long-short-term memory network (LSTM), and the multivariate time sequence prediction model corresponding to a transient working condition is a deep time sequence network adapting to transient characteristics of variable loads and is used for adapting to strong nonlinearity and strong time-varying characteristics of parameters under the transient working condition.
8. A nuclear power plant water supply system fault diagnosis apparatus, comprising: the data acquisition module is used for acquiring multivariable operation parameters of the water supply system of the nuclear power plant under a target working condition and preprocessing the multivariable operation parameters to obtain a preprocessed parameter sequence; the prediction model module is used for training a multivariable time sequence prediction model for predicting future time parameter values according to a historical parameter sequence based on the historical operation parameters only including normal working conditions; the residual construction module is used for inputting the preprocessed parameter sequence under the target working condition into the time sequence prediction model to obtain a predicted parameter value, calculating the difference value between the actual parameter value and the predicted parameter value, and constructing a multi-variable multi-time-step residual characteristic vector; The fault diagnosis module is used for taking the residual characteristic vector as input, training a fault diagnosis model and outputting the operation state type of the water supply system through the fault diagnosis model, wherein the operation state type comprises a normal working condition, a known fault working condition and an unknown fault working condition.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the method of claims 1-7.
10. A computer readable storage medium, characterized in that it has stored thereon a computer program for causing the computer to perform the method according to claims 1-7.

Description

Nuclear power plant water supply system fault diagnosis method and device, electronic equipment and storage medium Technical Field The invention belongs to the technical field of operation safety and state monitoring of nuclear power plants, and particularly relates to a method and a device for diagnosing faults of a water supply system of a nuclear power plant, electronic equipment and a storage medium. Background The water supply system of the nuclear power plant is an important component for ensuring the stable water level of the steam generator and maintaining the safe operation of the reactor coolant system. Along with the development of the nuclear power unit to the high-parameter and high-capacity direction, the coupling relation of the water supply system is complex, the dynamic characteristics are obvious, and once the faults of water loss, water supply pump failure, blockage of the regulating valve and the like occur, the water level control of the steam generator and the cooling safety of the reactor core are directly affected. The prior diagnosis method of the water supply system of the nuclear power plant mainly comprises a diagnosis method based on an analytical model, a knowledge reasoning method based on expert experience and an intelligent diagnosis method based on data driving. The analysis model relies on accurate mechanism modeling and parameter identification, various working conditions are difficult to fully cover in a complex coupling water supply system, the expert system has high requirements on the integrity and maintenance cost of a knowledge base and is difficult to adapt to a novel unit and a novel working condition, the data driving method is used for carrying out fault classification by directly utilizing operation parameters or signal characteristics, and abnormal confusion caused by normal working condition change and faults is easy to occur under the scene of obvious working condition change and strong parameter time variability. In the prior art, one type of method predicts future short-term parameters by establishing a trend model or a linear residual autoregressive model of equipment operation parameters, and judges abnormality by judging whether a predicted value exceeds a preset threshold or a predicted interval. The method is usually aimed at the out-of-limit early warning of single variables or few variables, focuses on 'whether abnormal or not', is difficult to realize multi-fault type distinction and unknown fault identification, and does not fully consider nonlinear coupling relation and time sequence dependence among multiple variables. Another type of method directly inputs the original sensor signals (such as vibration signals) into end-to-end models such as a depth residual error network, automatically extracts features through a deep convolution structure and outputs fault types. Although the feature extraction capability can be improved, prediction and diagnosis are coupled in the same network, it is difficult to explicitly describe the difference between the normal behavior and abnormal deviation of the system, and the adaptability to the change of working conditions and the interpretability of the diagnosis result still have room for improvement. Under the water supply system scene of the nuclear power plant, especially under transient working conditions such as power rising, falling and the like, the normal variation range of the operation parameters is large, the time scale is various, and the traditional method is difficult to simultaneously consider the following requirements: 1. fully learning the dynamic evolution rule of the multivariable parameters under the normal working condition; 2. Effectively distinguishing abnormal deviation caused by faults from normal working condition changes; 3. on the basis, the accurate identification of multiple fault types and unknown faults is realized, and the method is suitable for different power classes and load change processes. Therefore, a new fault diagnosis method is urgently needed to decouple 'normal behavior modeling' and 'fault mode identification', namely, a depth time sequence model is utilized to learn the dynamic response of a water supply system under normal working conditions, and then residual errors between actual operation and predicted behaviors are utilized to construct diagnosis features, so that high-robustness multi-fault diagnosis is realized under steady-state and transient working conditions. Disclosure of Invention In view of the above, the invention aims to provide a patent name to solve the problems that the diagnosis precision is reduced when the working condition changes are obvious, the normal working condition changes and the abnormality caused by the faults cannot be effectively distinguished, and the multi-fault type and the unknown fault identification are difficult to consider in the traditional diagnosis mode. In order to achieve the above object, the present invention adopt