A Hybrid Vibration Signal Prediction Model Using Autocorrelation Local Characteristic-Scale Decomposition and Improved Long Short Term Memory

Abstract

Accurate state monitoring and the fault prediction model is very important for the smooth running of a reciprocating compressor. Vibration signal is a sensitive characteristic parameter for fault prediction of a reciprocating compressor. Thus, it is necessary to develop an accurate and stable vibration signal prediction model. However, it is difficult to predict using a simple model for its nonlinear and nonstationary characteristics. Aiming at the characteristics of the vibration signal, a hybrid prediction modeling strategy called ACLCD-PSOLSTM is proposed by combining autocorrelation local characteristic-scale decomposition (ACLCD) and improved long short term memory (LSTM) neural network. To reduce the complexity of modeling, the original vibration signal is decomposed into many intrinsic scale components (ISCs) and a residue item by ACLCD. Then, each of the ISCs is predicted using the particle swarm optimization LSTM (PSOLSTM) model. And all the predicted results are accumulated as the final predicted result of the vibration signal, where the autocorrelation characteristics of the signal are considered to overcome the end effect of traditional LCD. For better performance of the LSTM prediction model, a multiobjective optimization model is established that balanced the prediction ability of the LSTM (RMSE) with the model complexity (hidden neurons and time lags). And the model is solved by the PSO algorithm. To validate the predicting capacity of the proposed hybrid ACLCD-PSOLSTM model, four different predicting models are implemented on the vibration signal series. The results of experiments show the superiority of the hybrid model over other models in improving the predictive performance.