Health Assessment for Piston Pump Using LSTM Neural Network

A piston pump is one of the key components in hydraulic system, which once fails will severely hurt the reliability of the hydraulic system and cause great loss. Therefore, accurate, reliable and effective health assessment must be performed. Currently, effective representation of diagnostic results and efficient health state assessment of dynamical pump systems have remained challenging. In this paper, a novel health assessment method is proposed using Long Short-Term Memory (LSTM) neural network. Firstly, 256 time-frequency features from wavelet packet decomposition, combined with 8 classical time-domain and frequency-domain features, are extracted to form an original feature set. Then, based on Fisher's Criterion, the most sensitive features are selected from the original feature set. Finally, these selected features are fed into a LSTM neural network to establish a relationship between feature set and health states, which achieves the aim of health state classification and health assessment. The effectiveness of this method is experimentally validated in a piston pump test. Performances of other classification methods, such as LR, SVM, MLP, RNN and direct LSTM, are tested and contrasted. Experiment results show that the proposed approach achieves the highest classification accuracy.