Remaining useful life prediction of milling cutters based on long-term data sequence and parallel fully convolutional feature learning

Multi-sensor time-series signals demonstrate long-term dependence along the temporal dimension, and exhibit significant data distribution discrepancies between sensor channels. However, existing remaining useful life (RUL) prediction methods do not fully utilize this information. To solve this problem, a new deep feature learning network named parallel fully convolutional feature learning network (PFCFLN) is proposed in this paper for RUL prediction of milling cutters. In the proposed PFCFLN, multiple parallel fully convolutional subnets with different scales and layers are first employed to learn local features from long-term multi-sensor data sequences. Secondly, long-term dependence between local features is captured through the long-term attention mechanism, and local feature expression is enhanced through short-term attention and channel attention. Finally, the RUL of the milling cutter is predicted in the regression layer using attention-weighted features. In addition, a training method and an architecture search method are also presented to improve the practicality and prediction accuracy of the PFCFLN. To evaluate the proposed PFCFLN and related methods, multi-sensor data was collected from full-life testing of the milling cutter. Evaluation experiments were conducted on these data, and also compared with five advanced prediction methods. Experimental results proved the superiority of the proposed PFCFLN in improving milling cutter RUL prediction accuracy.