A Dynamic Health Index Construction Framework Based on Deep Reinforcement Learning for Prognostic Analysis

Intelligent prognostics based on condition monitoring face significant technical challenges, particularly in accurately identifying equipment degradation from multisensor data. Developing a health index (HI) from monitoring data has emerged as a promising solution. However, existing approaches commonly construct the HI using a set of fixed weights, which may not capture the varying information provided by different sensors. Moreover, existing weight optimization approaches, including solvers and heuristic-based optimization, have limited efficacy in solving dynamic weight-based HI fusion. To address these limitations, this study proposes a novel deep reinforcement learning (DRL)-based framework for HI construction, which integrates dynamic data-level fusion with prognostic evaluation. Specifically, in the proposed DRL environment, the agent’s actions are defined as fusion weights. To enhance the agent’s ability to capture long-term degradation patterns from monitoring data, the temporal convolutional network (TCN) is employed for feature extraction. The agent is trained through rewards based on desirable prognostic properties and remaining useful life (RUL) prediction, thereby discovering an optimal data fusion policy for HI construction. Finally, a degradation dataset of the aircraft gas turbine engine is used as a case study to evaluate the proposed method.