Research On The Method Of Nuclear Power Plant Operation Status Monitoring And Intelligent Alarm Analysis

Nuclear energy is becoming more and more widely used,and nuclear power systems need to focus on and research nuclear safety issues.Currently,the operation control and decision support of nuclear power plants are mainly implemented by operators,but once the operator makes a mistake in judgment,it may lead to serious consequences.With the rapid development of China’s Industry 4.0 concept,"smart nuclear power" is becoming a hot topic.Operator support technology is one of the core technologies of "smart nuclear power".The operator support technology can provide operators with more detailed information about the operation of nuclear power plants,thus reducing the probability of misoperation and ensuring the safe operation of nuclear power plants,which is one of the core technologies of "smart nuclear power".At present,the main problem facing the operator support technology is the lack of advanced and detailed condition monitoring and alarm analysis methods,which leads to a lot of time and effort required by the operator for decision analysis.In previous researches,the monitoring and alarming methods are based on fixed thresholds.Such fixed thresholds or models are difficult to adapt to fluctuations during operation and cannot effectively analyze unknown faults.With the development of artificial intelligence technology,the lifelong learning method provides an important basis for dynamic updating of the model.This thesis transforms the static monitoring,alarm and diagnosis model into a dynamically updated evaluation model,integrates the advantages of data-driven and bio-memory systems into the computational process,and proposes an intelligent analysis method for condition monitoring and alarm based on lifelong learning to provide more detailed and accurate nuclear power plant status information.This thesis focuses on the following aspects:The false alarm generated in the long-cycle operation of nuclear power plants is one of the problems to be solved.Besides the noise and disturbances,the false alarm is also affected by the slow time-varying effects of operating parameters.In nuclear power plant operation,the slow time-varying parameter effect is a phenomenon in which the operating parameters are affected by core fuel loss,power regulation,and cyclic changes in seawater measurement temperature,resulting in changes in the steady-state benchmark of the system.However,fixed thresholds can be affected by the parameter slow time-varying effect to generate a large number of false alarms.To solve this problem,this thesis combines the multivariate statistical analysis with the complementary learning system to propose a design method of adaptive threshold.The threshold value can be adaptively adjusted according to the slow time-varying effect of the parameter,so as to suppress false alarms.Finally,the accuracy of the described method is verified by online data interaction with the full-scale simulator of a nuclear power plant.The ambiguous alarm hierarchy in nuclear power plants is the second problem to be solved.Faults in nuclear power plants can occur at any system level.At this stage,the condition monitoring and alarm analysis model is unable to distinguish between signal failure,system anomalies,and control loop malfunctions,and often divides several of these faults into one alarm level.This problem causes the operator to spend a lot of time analyzing the fault hierarchy in order to make a reasonable control decision.To solve that problem,this thesis establishes an alarm hierarchy classification model based on monitoring data reconstruction space and cyclic monitoring,and analyzes the maximum number of cyclic monitoring required for fault reconstruction in different hierarchies,to classify faults into alarm hierarchies and give accurate information about the hierarchy where the faults are located.Finally,the accuracy of the method is verified by simulation examples.The method of grouping alarm information in nuclear power plants is the third of the problems to be solved.Currently,advanced alarm messages are often group-based.However,influenced by the structure and data characteristics of nuclear power plant systems,the traditional grouping indexes and grouping schemes cannot effectively group nuclear power plant monitoring parameters.This thesis proposes a hierarchical grouping scheme for nuclear power plant monitoring and alarm analysis,which can give the overall status information of nuclear power plant and the key alarm information of each subsystem.Firstly,the scheme classifies the monitoring parameters into multiple levels according to the nuclear power plant system structure and generalized correlation coefficients,and uses adaptive clustering to regroup the data in each level to reduce the influence of human factors on the data grouping process.Then,the alarm information within each monitoring level and grouping is determined by the weighted contribution rate.Finally,Bayesian fusion is used to integrate the alarm information of each grouping and each level to obtain the overall system status information.Fault identification in nuclear power plants is the fourth problem to be solved.The fault data accumulation in nuclear power plants now is insufficient to establish a diagnostic model with relatively wide coverage.Moreover,at the beginning of the fault,the parameters do not all contain fault characteristics;in the early stage of fault occurrence,the fault characteristics contained in the parameters are also limited.All these factors can affect the effectiveness of fault diagnosis.Therefore,this thesis proposes a dynamic weighting method to increase the difference between fault features and non-fault features by assigning dynamic weights to the monitoring data,and thus improve the fault diagnosis accuracy.Finally,this thesis proposes a just-in-time-learning library update strategy for the fault library to implement the update of the diagnostic model to improve the performance of the model in practical applications.