Research On Fault Warning And Diagnosis Technology Based On Process Variables Of Nuclear Power Equipment

Nuclear energy is a stable,clean energy source that occupies an important place in our energy structure.To vigorously develop nuclear power,we need to put forward higher requirements on the timeliness,accuracy and intelligence level of fault warning and diagnosis.Pressurized water reactor nuclear power plants are the main type of nuclear power plants in the world,and once a failure occurs,it may turn into a catastrophic nuclear leakage accident.In order to reduce the possible economic losses or safety accidents caused by failure shutdown of pressurized water reactor nuclear power plants,accurate nuclear power failure warning and diagnosis technologies are urgently needed.A large amount of historical operational data has been accumulated over the decades since the development of nuclear power.Various simulators also provide data support for research related to datadriven nuclear power fault diagnosis.The rapid development of artificial intelligence and various branches of theory has laid the theoretical foundation for the research of data-driven nuclear power fault warning and diagnosis.In this paper,the following work is carried out on the main system of the first circuit of a nuclear power plant,which is the reactor coolant system:(1)The structure,function and operation of pressurized water reactor nuclear power plants are first explained,focusing on the four core components of the reactor coolant system.On this basis,the causes,effects and characteristics of steam generator tube rupture(SGTR)and loss of reactor coolant(LOCA)were analyzed and FMEA tables were created to select the relevant parameters that characterize the failure states.(2)A multivariate state estimation technique(MSET)based early warning method for nuclear power equipment failures is proposed.Based on the operation data of Fuqing 2nd simulator,the memory matrix is constructed by the method of equidistant sampling.A similarity function is defined to reflect the degree of deviation between the estimated and observed values.Based on the idea of interval estimation in statistical principles,an adaptive threshold calculation method for similarity is proposed.The MSET model is validated with the normal operation data set,the fixed degree fault set inserted by the simulator,and the asymptotic fault set with the introduction of linear deviation,respectively,and the results show that the proposed model can provide accurate and timely warning of abnormal conditions.(3)In order to integrate the advantages of various artificial intelligence algorithms,a stacking fault diagnosis model based on integrated learning is proposed,which consists of a layer of base learner and a layer of meta-learner.The fault data segments after MSET model warning are used as the original data set for the stacking diagnostic model.First,to avoid the occurrence of overfitting,a K-fold cross-validation method is used to divide the dataset and train the base learner.The fault diagnosis models based on different algorithms are diagnosed individually and their results are subjected to Spearman correlation analysis,and the algorithm with correlation greater than 0.9 is selected as the first layer of base learners.An improved MSAWCNN algorithm is proposed as a meta-learner,and the proposed MSAWCNN has a multiscale feature extraction layer and an adaptive weighting layer with stronger generalization compared to a normal CNN.The validation results show that the stacking fault diagnosis model proposed in this paper improves the diagnosis accuracy by 8.71%on average over other single algorithms.