Study On Fault Diagnosis Method For Nuclear Power Plant Based On Data Mining

Fault diagnosis system plays an important role for ensuring the safety and economy ofthe operation of nuclear reactor. At present threshold monitoring method is the main methodemployed in our country. This method can not provide the information of the root cause andthe deterioration trend. So studying the intelligent fault diagnosis technology of nuclearpower plant (NPP) has great significance to improve the safety and economy of nuclearpower plant. Knowledge acquisition has become the bottleneck problem of establishingintelligent fault diagnosis system. Data mining is the main way to solve problem ofknowledge acquisition. The comprehensibility, the efficiency and generalization ability ofdiagnostic model evaluation are all important indicators to evaluate the performance of thefault diagnosis model, when we establish the fault diagnosis model by data mining. Andthese aspects are the difficulties to establish fault diagnosis systom.In this paper, with typical faults of primary loop and secondary loop system in NPP forobject, the research and exploration of fault diagnosis technology and methods according tothe above difficulties and problems are listed as follows:Firstly，the typical faults of NPP are analyzed, and fault samples are collectedaccording to the analysis results. Knowledge acquisition is the bottleneck problem ofestablishing expert system. And the model established by black-box modeling method suchas artificial neural networks is not comprehensibile. According to these problems, in thispaper fault diagnosis model is established by decision tree ID3and C4.5respectively. Andthe classification rules which are obtained by decision tree algorithm are compared with therules which are obtained by theoretical analysis. The effectiveness of knowledge acquisitionof decision tree algorithm is verified.Secondly, NPP is a very complex system, which need to collect and monitor vastparameters. So the dimension of fault samples is very high, leading to the executionefficiency of fault diagnosis model is low and the training time is long. According to theseproblems, in this paper the important degree of the parameters are evaluated byneighborhood rough set. And parameters reduction algorithm is constructed base on greedysearch strategy. The problem which Pawlak rough set can not process continuous data isovercomed. The simulation result shows that the execution efficiency of fault diagnosis model is greatly improved by parameters reduction.Parameters reduction may be lose some useful information, leading to the thegeneralization ability of fault diagnosis model decreases. In this paper ensemble learningmethod was proposed according to the problem. Firstly, multiple base classifiers is trained.And the final results is obtained through fusing the base classifier. Ensemble learningmethod can increase the generalization ability of fault diagnosis model on condition ofensure the execution efficiency of fault diagnosis model. At the same time, this paperstudies the problem of invalid and absent parameters which may happen in the actualoperation of NPP. The Simulation results show that this method can get a good result on thecondition of invalid and absent parameters. So this medthod shows very good faulttolerance.At the same time, the time-varying characteristic of NPP is studied in this paper. Theoccurrence of NPP failure is a gradual process of evolution through the analysis of theautocorrelation coefficient curve of parameters. So it is necessary to deal with faultdiagnosis of NPP as a sequential supervised learning problem. The dependency ofparameters and their order lags is computed according to the important degree of theparameters base on neighborhood rough set. The classification capability of parameters isimproved as adding the order lags of parameters. The size of sliding-window is chosenaccording to the dependency curve. The feature extracting algorithm is studied and chosenaccording to the time-variant characteristics of parameters. Simulation results verify that thesequential supervised learning algorithm can get more information from the parameters.Some problems which can’t be solved by classical algorithm can be diagnosed by sequentialsupervised learning method.