Multisensor bearing fault diagnosis based on one-dimensional convolutional long short-term memory networks

Bearings are the key components of various rotating machinery, and their fault diagnosis is very important for improving production safety and economic efficiency. In this paper, an end-to-end solution with one-dimensional convolutional long short-term memory (LSTM) networks is presented, where both the spatial and temporal features of multisensor measured vibration signals are extracted and then jointed for better bearing fault diagnosis. In addition, the number of time steps in the LSTM layers for the long-term temporal feature extraction is much smaller than the length of the input segments, which can highly reduce the computational complexity of the LSTM layers. The experimental results demonstrate the presented solution has better performance than other methods for bearing fault diagnosis, meanwhile, its adaption to different loads and low signal-to-noise ratios is also verified. (C) 2020 Elsevier Ltd. All rights reserved.