Fault Prediction Method Based On Recurrent Neural Network And Evidence Reasoning

As the modernization process of industrial production keeps getting faster,the demand for production safety will become increasingly high.With the continuous high-intensity operation of industrial equipment,different types of fault will appear.Fault prediction technology can predict the type of faults in the future of the equipment in advance according to the current state of the equipment,so as to obtain more sufficient time for maintenance and effectively avoid the occurrence of serious accidents.Fault prediction models need to analyze and predict the changing trends of the fault features in order to infer the future fault status of the equipment,and the information of fault features is typically viewed as a time series.Due to the dynamic change of equipment working conditions and various disturbances in the monitoring environment,there is a strong uncertainty within time series data of fault features.Evidence Reasoning(ER)rule and Belief Rule-Based(BRB)inference methodology have obvious advantages in modeling various uncertain data,and are widely used to establish nonlinear uncertainty relationships between multi-dimensional inputs and output.In addition,Long Short-Term Memory(LSTM)can effectively use the historical fault feature sequence to realize the prediction of fault feature.In this thesis,the problem of fault prediction is studied based on recurrent neural network and evidence reasoning methodology,and the main contents are as follows:(1)Fault feature prediction method based on LSTM.Illustrate the specific tasks of fault feature prediction in fault cases of electromechanical equipment such as motor rotor and railway turnout.Establish the prediction model of fault feature based on LSTM,then input the historical and current fault feature data into the model to predict the changes of the future feature.The adjustable factor η of the input data is introduced to adjust the amount of the input data to improve the prediction accuracy of the model.Finally,through the fault feature prediction experiments of motor rotor and railway turnout,it is shown that different prediction effects can be obtained by adjusting the value of η.(2)BRB-LSTM-based fault feature prediction method with adjustable input data.Due to the uncertainty of the feature series,it is necessary to dynamically adjust the value of η to achieve better prediction accuracy.Therefore,the BRB model is introduced to dynamically infer the η value according to the real-time statistical features of the LSTM input series,and then established the BRB-LSTM fault feature prediction model,which the better prediction accuracy of model can be obtained by adaptive adjustment of η value.The effectiveness of the adaptive adjustment strategy is verified in the fault feature prediction experiments of motor rotor and railway turnout.(3)Fault prediction fusion decision-making method based on ER rule.A reference evidence matrix(REM)of multi-source fault features is constructed to describe the uncertainty mapping relationship between the feature vectors and the types of fault.After the prediction vector of the fault features is obtained online through the method in(2),it is input into the REM to activate the corresponding predicted evidence,and the multiple pieces of activated evidence are fused by ER rule,then the decision is made based on the fusion results to obtain the types of fault in the future.Finally,the effectiveness of the proposed method is verified in the fault prediction cases of motor rotor and railway turnout,respectively.