Based On Data-driven Compressor Vibration Signal Prediction

The fault prediction and diagnosis of compressor is important to the smooth operation of industrial production environment.Vibration signal is the sensitive characteristic parameter of compressor fault.Therefore,it is great significance to establish an accurate and stable model of vibration signal prediction for fault prediction and diagnosis.However,the complicated compressor structure and harsh production environment cause the vibration signal to have obvious nonlinearity and nonstationarity,which makes it difficult to accurately predict the vibration signal.Meanwhile,with the passage of time,the state of the equipment changes,resulting in the normal operation range of the vibration signal changes.At this time,the prediction accuracy of the previously established model may decline.To ensure the continuous validity of the model,it is necessary to introduce online learning algorithm into the prediction model.Therefore,accurate prediction of vibration signal and online updating of prediction model are deeply studied in thesis.The main research is as follows:To reduce the complexity of modeling,firstly,the Local Characteristic scale Decomposition(LCD)algorithm is used to decompose the original vibration signal into several Intrinsic scale components(ISC),so as to smooth the change of vibration signal sequence.To solve the problem of end effect in the decomposition of vibration signal by LCD algorithm.A Boundary Modification(BM)algorithm is proposed in this thesis on the basis of the traditional modification methods,which is introduced into the process of LCD decomposition(BMLCD)to further suppress the end effect.The experiment results show that BMLCD algorithm can effectively process real industrial signals.Secondly,due to the strong nonlinear and non-stationary characteristics of the vibration signal,the single prediction model is often one-sidedness,resulting in the limited prediction accuracy of the model.A hybrid prediction model based on BMLCDPSOLSTM is proposed.The original signal is decomposed into a number of ISC components using BMLCD algorithm,the PSOLSTM prediction model is established for each component,and the component results are added to obtain the final prediction result.Since the selection of hyper-parameters directly affects the model accuracy,a multi-objective optimization model is established that weighs the model prediction accuracy(RMSE)and model complexity(hidden layer neurons and time-lag number),and Particle Swarm Optimization(PSO)algorithm is used to obtain the optimal parameter combination to ensure the best performance of the model.Finally,to ensure the long-term validity of the model when data distribution changes,the idea of online updating is introduced into the prediction model.Due to the low calculation efficiency of the LSTM model,it is difficult to adapt to the online update scenario.An online update model based on Error-LSTM(E-LSTM)is proposed in this thesis.The efficiency of the model is improved by updating the hidden layer neurons in blocks.Moreover,the E-LSTM model can adjust the training mode adaptively according to the data distribution changes,so as to improve the accuracy of the model.The LSTM model is used for comparison experiment,and the results show that the updating model based on E-LSTM is superior to the updating model based on LSTM in both accuracy and efficiency.