Temporal convolution-based sorting feature repeat-explore network combining with multi-band information for remaining useful life estimation of equipment

Remaining useful life (RUL) estimation of key components is a particularly important link in the reliability evaluation of overall unit. Due to complex nonlinearity and uncertainty in degradation process of mechanical systems, conventional methods are difficult to fulfill the accurate medium & long-term predictive maintenance tasks. To address this issue, this paper proposes a novel concurrent residual temporal convolution network in repeat-explore mode (CRTCN-RE Mode), which utilizes multi-branch structure to extract abstract sorting features from multi-band information and filters high-discrimination degraded features in dual modes. First, multi-band envelope spectrums of raw signals are calculated as initial inputs. Dilated causal convolution with multi-branch structure extracts high-level degraded representations from multi-band information by virtue of its ability to overcome long-term dependence. Then, the extracted sorting features are fed into dual modes, that is, repeat mode combines historical information to obtain the recurring general degradation features of each branch, while explore mode calculates the feature contributions of different degradation moments in various branches. Finally, the end-to-end RUL estimation is implemented relied on the CRTCN-RE Mode. The effectiveness of proposed method is validated by two life-cycle datasets. Comparative study indicates that proposed method has better accuracy and rationality than other state-of-art methods in RUL estimation and prognostic analysis.