Study On The Reliability Of Operators’ Behavior Of Digital Nuclear Power Plant Based On Statistical Learning

A systematic study of the reliability of operator’s behavior in digital nuclear power plants(NPPs)to improve the safety and economy of NPPs is a hot and difficult problem in engineering and academic today.Especially in the era of rapid development of artificial intelligence(AI)in the context of big data,the Human Reliability Analysis(HRA)method also needs to be improved.In this paper,we identify and quantify the main performance shaping factors(PSF)that affect the reliability of operators in the control room of digital NPPs through statistical learning methods,and then establish an operators’ behavior reliability model based on the main PSF,and finally discuss strategies and algorithms to reduce and prevent the operator’s behavior errors based on the Human Error Probability(HEP)obtained from the model,mainly as follows.(1)The applicability of typical HRA methods in the new contextual of AI diagnosis of faults in the main control room is studied.By analyzing the operator’s task characteristics in this context,establishing a hierarchical structure of HRA method applicability evaluation criteria,conducting a fuzzy comprehensive evaluation of typical HRA methods from four factor indicators of qualitative analysis,quantitative analysis,reliability and usability,and conducting a discussion on the determination of factor weights of evaluation indicators according to the characteristics of acquired data.The experimental results show that one HRA method alone cannot be well applied to human reliability analysis in the context of AI diagnostic faults,and it is necessary to improve the original method or establish a new method.(2)A new model for evaluating PSF in digital control rooms of NPPs is developed.The number of PSF for quantifying HEP in the HRA method is large,and some PSF are correlated with each other,bringing erroneous evaluation to the quantification of HEP.To overcome this deficiency,this paper proposes an unsupervised learning algorithm that uses a combination of correlation coefficient array,distance classification of graphs and principal component analysis method to construct an evaluation model of PSF of operators in digital control rooms of NPPs according to the characteristics of human-caused event statistics.The model is able to identify PSF in different types of human factors events that primarily affect human performance for decision making to reduce human-caused errors.Experiments with 179 human-caused incident reports from a NPP are conducted,and the results show that the model can effectively evaluate the PSF of operators in digital control room of NPPs.(3)The models and algorithms for quantitative prediction of PSF based on experimental data types using statistical learning methods are discussed,in which the linear regression algorithm is studied for optimization.In order to identify the mechanism of operator behavioral errors,establishing the function relationship between PSF and HEP and accurately quantifying PSF is an urgent problem to be solved.In this paper,typical statistical learning methods are selected to study the quantification of PSF based on the types of values taken by PSF and the types of values taken by their characteristic factors.Experiments were conducted with the important PSF of human-machine interface interactivity level,and the related indexes of context awareness and mental load were taken as its quantified characteristic factors.According to the type of data obtained from the experiments,a BP neural network model was selected for statistical learning,and the results showed that the training error and testing error were small.(4)A mathematical model for the reliability of operator operational behavior in digital NPPs is developed.Firstly,the operational behavior is decomposed into basic action units,i.e.,meta-operations.Secondly we micro-analyze the characteristics of operator’s behavioral errors to establish differential difference equation about the number of errors and operating time.Finally we get the human error probability density that obeys exponential distribution,in which the human factor parameter is determined by using statistical learning method of controlled parameter optimization,then get the mathematical model to predict the reliability of operator’s operational behavior.An experiment with DCS+SOP as an example is conducted to verify the rationality of the model.(5)An intelligent fuzzy diagnosis model for nuclear power accidents based on gray correlation degree is constructed.The ultimate goal of the research on HRA methods for digital NPPs is to improve the reliability of the operational behavior of operators in NPPs,in which diagnostic behavior is crucial.when the operator of the main control room can not handle the accident in time or make wrong judgment under accident condition in nuclear power plant,which may lead to the occurrence of major accidents or even more accidents concurrently.An intelligent fuzzy diagnosis model based on Grey Relational degree is constructed in this paper,by establishing accident-state correlation matrix and accident-state probability matrix,defining the vector of information weights for accident characteristic vector and monitoring result vector with linear transform,using Grey Relation to measure the relevance degree between monitoring result vector and accident characteristic vector,Searching for suspected accidents approaching to real solutions in iterative recursive algorithms.LOCA accidents as well as Monte Carlo numerical simulation experiments show that the model and algorithm can accurately and quickly diagnose various common accidents in complex nuclear power systems and provide strategies for operator’s diagnostic decisions.(6)A human-computer interaction-based early warning model and algorithm for the cognitive reliability of operators in NPPs are studied.In order to improve the reliability of operators’ behavior in NPP,in addition to the cognitive behavior of diagnosing accidents,there is also the cognitive behavior of operators guided by protocols to eventually transition the nuclear power system to a safe state.Thus,we establish an early warning model of operator behavior reliability and its algorithm based on the relevant theory of graph theory,then conduct an optimization study of the algorithm to realize intelligent monitoring and correction of nuclear power management to make positive contributions in such fields as human-caused accident prevention.The reliability of operators’ behavior in digital NPPs studied in this paper based on statistical learning method can complement and improve the existing human-caused reliability analysis system,which provide strategies and theoretical support for predicting the reliability of operator behavior in the main control room under accident conditions in NPPs and preventing human-caused accidents,thus further helping to improve the safety level of NPPs.