Knowledge-Informed Attention for Remaining Useful Life Prediction for Rolling Bearings

This research introduces a novel Knowledge-Informed Attention Module (KIAM) within a deep learning framework to enhance the remaining useful life (RUL) prediction accuracy for rolling bearings. The traditional squeeze-and-excitation attention mechanism has achieved great success in various applications; however, directly applying this mechanism without tailoring the design for rolling bearing applications may result in suboptimal performance (e.g., it tends to emphasize the most distinguished information rather than the degradation-relevant information that may be embedded in the signal). Therefore, in this research, we propose incorporating human knowledge into the attention mechanism to emphasize degradation-sensitive information and thereby improve RUL prediction accuracy. More specifically, the proposed KIAM redesigns the spatial attention by introducing the health indicator (HI, fused from hand-crafted features) to replace the original spatial filter, and redesigns the channel attention by incorporating the information of the current timestamp. It is noted that the entire deep learning framework accepts the raw vibration signal as input (with human knowledge also provided as input to the attention), which effectively extracts health-representative patterns without losing information. Experiments are conducted with the publicly available PRONOSTIA and IMS datasets to validate the effectiveness of the proposed method. The results show that the proposed method achieves a promising RUL prediction accuracy (e.g., 0.13 in terms of RMSE for the average prediction error with PRONOSTIA and 0.119 with IMS data) and outperforms state-of-the-art methods. In addition, the results also indicate that the quality of the features (measured in terms of the Cori Index) is significantly improved after being recalibrated by the proposed attention mechanism (improved by up to 176.4%).