A Novel Approach for Remaining Useful Life Prediction With Anomaly-Driven Detection and Deep Feature Transfer

Abstract Remaining useful life (RUL) prediction is crucial for the predictive maintenance and health management of mechanical transmission systems. Although data-driven RUL prediction methods have advanced significantly, three key challenges remain: (i) constructing a physically interpretable health indicator (HI) that is invariant to operating conditions yet sensitive to incipient degradation; (ii) achieving adaptive time to start prediction (TSP) detection that accommodates nonstationarity and controls false and missed alarms; and (iii) accurate and generalizable RUL prediction under cross-domain scenarios. To address these issues, a method combining unsupervised anomaly detection and deep feature transfer for TSP determination and RUL prediction is proposed. A residual health indicator is constructed from an expected trajectory learned by Gaussian process regression (GPR) on the healthy segment and a Mahalanobis distance (MD) deviation measure. For the first time, extreme value theory (EVT) is used to create dynamic thresholds for performance degradation analysis and TSP determination. During the RUL prediction phase, transfer component analysis (TCA) mitigates feature distribution discrepancies between source and target domains. A dual long short-term memory (dual-LSTM) network with L2 regularization is then employed for RUL prediction across two transfer tasks. Experimental validation on a public benchmark dataset and a laboratory run-to-failure bearing dataset demonstrates the superior predictive accuracy and generalization capability of the proposed approach compared to existing methods.