Prediction Method for Remaining Useful Life of Rolling Bearings Based on Bidirectional Temporal Convolutional Network and Long Short-Term Memory Network.

Due to the insufficient sensing field of the temporal convolutional networks (TCN), the key degradation information of the bearing is often ignored, which results in poor prediction of the remaitning useful life (RUL) of bearings. Moreover, the long-term dependence problem of long short-term memory (LSTM) may not be well solved with the increase of data volume and sequence length. Therefore a new prediction method based on Bidirectional temporal convolutional network and Long short-term memory (Bi-TCN-LSTM) was proposed. Firstly, the multi-sensor data was normalized and fused, and then the Bi-TCN-LSTM was used for data feature extraction and deep learning, in which the convolutional attention mechanism (CAM) was introduced into the TCN module, and the three gates of the LSTM were simplified into one gate. It effectively accelerated the learning speed of the prediction model and improved the accuracy of the prediction model. The IEEE PHM 2012 bearing dataset was used to carry out the RUL prediction experiments. The results show that compared with other advanced prediction models, the Bi-TCN-LSTM method has relatively lower prediction error and better performance. [ABSTRACT FROM AUTHOR]