Data-mechanism fusion-based threshold-free operational reliability assessment integrating stochastic degradation process theory with deep learning

Operational reliability assessment is an important issue for predictive maintenance and health management of complex electromechanical equipment. However, traditional operational reliability assessment methods are trapped in the dilemma of the harsh reality of scarce failure samples, complex degradation mechanisms, and difficulty in calibrating failure thresholds. Thus, a data-mechanism fusion-based threshold-free operational reliability assessment method integrating stochastic degradation process theory with deep learning is proposed in this paper. On the one hand, a Reliability-oriented LSTM-Attention-Mixture Network (RLAM-Net) is developed to embed prior reliability degradation knowledge into a deep learning model, thus the model’s generalization and prediction accuracy are enhanced. On the other hand, to address the difficulty of the lack of calibrated failure thresholds in engineering practice, a threshold-free dynamic operational reliability assessment method is proposed by incorporating design-life prior information to implicitly characterize the degradation boundary. Finally, two failure probability assessment applications based on a simulated non-homogeneous Gamma stochastic degradation process and a real offshore wind turbine gearbox’s reliability degradation example demonstrate the feasibility and effectiveness of the proposed method. This method can be extended to other equipment for reliability assessment and prediction, supporting operational safety assurance.