Lightweight bidirectional long short-term memory based on automated model pruning with application to bearing remaining useful life prediction

Abstract

Rolling bearings are key components in industrial machinery, and their remaining useful life (RUL) prediction plays a prominent part in machine safety and maintenance. Bidirectional long short-term memory (Bi-LSTM) shows great performance in this field, but in most cases, the methods do not consider the model redundancy that may cause negative effects. To address this issue, this paper proposes a novel lightweight Bi-LSTM based on automated model pruning for bearing RUL prediction. The method carries out automatic model pruning by directly using the original Bi-LSTM model as input and intelligently identifying the redundant elements in the model based on norm information and reinforcement learning. It avoids the complex manual process of searching and selecting the optimal pruning architecture. The method can achieve an efficient adaptive allocation of computational resources and obtain a lightweight Bi-LSTM model with better performance. The application on the bearing RUL prediction shows that the lightweight Bi-LSTM can effectively predict the bearing degradation curve with the root mean square as health index. In comparison with several popular methods, the lightweight Bi-LSTM shows excellent learning ability in predicting long and complex time series. The lightweight Bi-LSTM achieves a 36% model pruning rate while improving the prediction accuracy by 3% when compared with the original Bi-LSTM. This study is of significance to the bearing fault state prediction and sub-health stage detection.