Health Indicator Construction Based on Multisensors for Intelligent Remaining Useful Life Prediction: A Reinforcement Learning Approach
Zhaoqin Peng, Xucong Huang, Diyin Tang, Quan Quan
Abstract
Health indicator (HI) representing the latent degradation pattern of engineering systems plays an irreplaceable role in system remaining useful life (RUL) prediction tasks. The HI is often constructed by fusing multiple sensors of the analyzed system and further applied to RUL prediction tasks. However, most existing HI construction methods combine signals without directly considering the following RUL prediction performance, resulting in a limited prediction accuracy based on the constructed HI. Therefore, this article proposes a reinforcement learning (RL)-based approach to construct HI based on multisensors, to directly link HI construction and the RUL prediction task. The HI construction problem is then transformed into leading an RL agent to automatically learn to find a combination rule of sensors with the most accurate predicted RUL result. Moreover, by setting different rewards for the RL agent, unique requirements for intelligent RUL prediction, such as HI being sensitive to a specific life stage, can also be fulfilled, which cannot be achieved by any other HI construction counterparts. Comparison with benchmark HI construction methods is conducted using two different datasets, and the advantages of our proposed approach are revealed.