Life-cycle modeling driven by coupling competition degradation for remaining useful life prediction

Estimating latent degradation states of mechanical systems from observation data provide the basis for their prognostic and health management (PHM). Recently, deep learning models have been employed to extract latent degradation features from observation signals. However, most of the existing methods using DL in PHM ignore the temporal causal dependencies throughout the entire life-cycle degradation process due to the slice training manner. To address this issue, this work proposes a novel state space model (SSM) named Coupling Competition Degradation based Deep Markov Model (C2D2M2). C2D2M2 utilizes deep neural networks to parameterize emission function and transition function in SSM, enhancing the latent feature representations. To describe the strong nonlinear degradation process of mechanical systems, coupling competition degradation mechanism (CCDM) is embedded into the transition function as prior degradation assumption. Specifically, we establish the transition equations according to three degradation mechanisms (linear, power rate, exponential degradation) and employ attention mechanism to realize competition among them. To predict remaining useful life (RUL), degradation indicator (DI) is estimated from the latent degradation state and two similarity-instance based learning (SBL) frameworks are designed for bearings and turbofan engines. Experimental results demonstrate that SBL frameworks based on C2D2M2 obtain excellent prognostic performance and attention heat map interprets competition process of three degradation mechanisms.