A Novel Bearing RUL Prediction Method Using Temporal Convolutional Attention Calibration Network with Multi-Resolution Feature Extraction

Remaining useful life (RUL) prediction is essential to ensure safe and economic operation for mechanical equipment. During the whole service life, the bearing degradation is diverse and nonlinear, and the data are characterized by strong noise and strong time-varying, which increases the challenge of bearing RUL prediction. Hence, a novel bearing RUL prediction method is proposed using temporal convolutional attention calibration network (TCACN) with multi-resolution feature extraction in this paper. First, the bearing raw degenerate signal is decomposed by improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) to generate a range of intrinsic mode functions (IMFs). Next, the signal complexity of IMFs is calculated using refined composite hierarchical fuzzy dispersion entropy (RCHFDE) to construct the raw high dimensional feature space (RHDFS). Then, the Introduce Local Tangent Space Alignment (LTSA) is used to fuse the RHDFS to obtain the low dimensional fused feature space (LDFFS) for eliminating the influence of redundant information and noise. Second, TCACN is built for mining deep feature information in the LDFS. Meanwhile, the network could focus on the useful feature information, suppress the influence of redundant information, and calibrate the LDFS. Finally, the method proposed is verified using experiments. The results demonstrate that a considerable benefit is offered by the suggested methodology.