Monitoring Signal Abnormal Identification Method Based On LSTM-Bayes Model For Nuclear Power Turbine

The safe operation of nuclear power plant is the most concerned problem for nuclear power industry.Intelligent fault warning and predictive maintenance for mechanical equipment such as steam turbine can effectively prevent safety accidents of nuclear power plant and reduce maintenance cost.At present,most of the early warning systems for machinery and equipment used for nuclear power are not ideal,and their functions are also imperfect.The fault early warning algorithms they carry are difficult to effectively deal with the data’s non benign,redundancy and complexity,and the accuracy needs to be improved;The threshold method is still used for anomaly recognition,but there is no unified standard for the selection of threshold.When the residual error approaches the threshold and fluctuates nearby,it is difficult to make decisions and quantify the degree of failure,which is likely to cause misdiagnoses and missed diagnoses.In view of the above problems,a anomaly detection method for nuclear power steam turbine monitoring signals based on LSTM-Bayes model and a predictive maintenance system based on this method are proposed.Integrating a variety of data preprocessing technologies,including outlier filling,wavelet packet de-noising,standardization and principal component dimension reduction,can effectively remove the interference of data defects on subsequent prediction and reduce the computational complexity;The principal component generated by dimension reduction is used as the characteristic signal to represent the overall state of the equipment.The fault signal reverse seeking method is derived from the solution of contribution rate,and in case of failure,it can not only locate abnormal signal quickly,but also reduce the calculation amount.For time series signal prediction model,Long Short-Term Memory(LSTM)is used to build a model body,the phase space reconstruction theory is used to determine the input and output structure of the model,and a new model evaluation index is proposed based on Bayes hypothesis test.The results show that LSTM is the most accurate prediction model compared with other models,the model structure can be determined quickly by phase space reconstruction,and the new index can not only quantify the reliability of the model,but also show the local prediction accuracy,which makes up for the shortcomings of the existing evaluation indexes in quantitative and local evaluationAiming at the prediction residual,an anomaly recognition strategy based on Bayesian inference is proposed,which considers the prior information,estimates the mean and variance of the prior residual set,and realizes the identification and quantification of equipment faults by calculating the Bayesian factor and fault confidence,so as to capture the anomaly more sensitively.The results show that the strategy can identify the abnormal conditions of steam turbine 10 hours in advance,predict the six alarms of air pump blade crack 100%,and compared with other models,has longer warning time and lower misdiagnosis rate.The application of the method to steam turbine fault cases shows that the data preprocessing process can effectively improve the bad data in Turbine Supervisory Instruments(TSI);the combination of probabilistic principal component method and equipment mechanical factors can significantly improve the dimensionality reduction effect,and the contribution rate of principal component is more than 60%;Compared with the residual threshold method,the Bayesian anomaly recognition strategy can give an early warning 36 hours earlier and has higher accuracy.The predictive maintenance system of nuclear turbine developed is equipped with LSTMBayes fault early warning algorithm.Compared with foreign systems,it has improved on many indicators,and added early warning parameters such as failure probability,remaining service life and characteristic signal to help maintenance,making itself more intelligent.It is verified that the input of the system can reduce the the equipment accident rate,save the maintenance cost,effectively ensure the operation safety,improve the productivity and reduce the cost.