Recognition And Prediction Of Fault State For Reactor Coolant Pump In Nuclear Power Plants

The reactor coolant pump(RCP)is the only high-speed rotating machinery in the reactor coolant system in nuclear power plants(NPPs).As part of the pressure boundary,the RCP usually operates in a harsh environment of high temperature,high pressure and high radiation,and its health state is directly related to the safe and economic operation of NPPs.At present,preventive maintenance is widely adopted in NPPs,that is,the equipment is periodically inspected and maintained according to its security level and average service life.Although this maintenance strategy reduces the probability of fault to a certain extent,unreasonable maintenance cycles are prone to problems of insufficient maintenance and excessive maintenance,thereby increasing the operation risk and maintenance cost of NPPs.Whereas,predictive maintenance can reasonably design the maintenance plan based on the state assessment and prediction of the equipment,and this effectively reduces the probability of unplanned downtime while improving the economics of NPPs.State recognition and prediction are the basis of predictive maintenance.State recognition is to identify the operating status of the system or component by using fault detection and diagnosis technologies,and the trend of the fault can be predicted after sufficient fault information is obtained.In order to promote the transformation of RCP maintenance from preventive maintenance to predictive maintenance,key technologies of the state recognition and prediction for the RCP have been investigated in this article.(1)To improve the sensitivity of the model to the fault,a distributed fault detection strategy for the RCP based on the correlation principle is proposed.In addition to this,there is a strong nonlinearity and time-correlation between signals of the RCP.Hence,a fault detection model,dynamic kernel principal component analysis(DKPCA),is adopted in this article,and in order to improve the performance of the model,local outlier factor and moving average filtering are used to detect outliers in the training data and eliminate false alarms,respectively.The performance of the improved DKPCA is validated by using the Tennessee-Eastman process and real faults of the RCP in a NPP,and the results highlight the advantages of the model in processing dynamic data.To solve the problem of the performance degradation of the DKPCA model in the long-term detection,a model update strategy based on the improved moving window is proposed in this article.The improved strategy adds a model update judgment mechanism based on the K-means clustering to the traditional model update strategy based on the moving window.The improved moving window DKPCA(MWDKPCA)is applied to the prolonged detection of the RCP under normal and fault conditions,and the results show that the improved MWDKPCA can effectively solve the time-varying problem of signals in the fault detection of the RCP,and it can effectively eliminate unnecessary updates of models while ensuring the performance of fault detection,thereby reducing the calculation cost of fault detection.(2)To improve the time-frequency analysis of the traditional Hilbert-Huang transform(HHT),a new time-frequency analysis method based on the variational mode decomposition(VMD)and the Hilbert transform(HT)is proposed in this article.In order to automatically tune the parameter of the mode number in the VMD,a parameter optimization method based on correlation coefficient is also proposed,which effectively solves the problems of insufficient decomposition and excessive decomposition of signals caused by the unreasonable setting of the mode number of the VMD.The time-frequency analysis performance of the improved VMD-HT is validated by using numerical simulation and rotor fault signals,and the results highlight the advantages of this method in processing non-stationary signals.In order to improve the intelligent fault diagnosis of the RCP,a fault diagnosis model based on the multi-sensor information and residual neural network(Res Net)is proposed in this article.Compared with the traditional convolutional neural network,the Res Net has stronger abilities of feature extraction and feature learning,and can effectively solve the degradation problem of deep networks.In addition,the proposed model can also make full use of information from sensors at different locations,which improves the accuracy and robustness of fault diagnosis.Moreover,in order to obtain more comprehensive feature information,vibration signals are converted from the time domain to the time-frequency domain by using the short-time Fourier transform,thereby making up for the shortcomings of feature extraction in a single domain.The diagnosis performance of the proposed model is validated by using the experiments of rotor and motor faults,and the results highlight the advantages of the proposed model in terms of convergence speed,accuracy and noise robustness.(3)To solve the problem of large errors in the time series prediction of the RCP by single models,a hybrid prediction model based on the EMD,sample entropy(Samp En)and long short-term memory(LSTM)neural network is proposed in this article.The proposed model reduces the complexity of the original signal by using EMD decomposition,and the Samp En-based component aggregation strategy effectively reduces the amount of calculation of the hybrid model under the premise of ensuring the prediction accuracy.In addition,the global optimization algorithm,Bayesian optimization,is used to automatically tune the hyperparameters of LSTMs.The proposed hybrid model is applied to the time series prediction of the RCP in a NPP,and the results validate the effectiveness of the hybrid model,and comparisons with other benchmark models highlight the advantages of the hybrid prediction model.In summary,key technologies,such as fault detection,fault diagnosis and fault prediction,for the RCP maintenance have been investigated in this article.The proposed methods have been fully validated by using simulations,experiments and real data from NPPs,and the research results have important value of engineering application in terms of improving the intelligent operation and maintenance of the RCP.