Remaining useful life prediction of multi-sensor monitored degradation systems with health indicator

Rolling bearings are kernel components of machinery devices. To evaluate degradation processes of rolling bears in real-time devices, health indicators (HIs) are required to be built. Due to the constraints of sensors, the degradation pattern cannot be denoted by commonly used signals such as vibration data. Moreover, the practical requirements of HI for prognostics are always ignored, such as monotonicity and trendability. Therefore, a novel HI construction method based on reinforcement learning (RL) is proposed. Firstly, the HI construction process is regarded a data fusion process. Observed multi-sensor data is used to abstract degradation information. In this way, the determination of HI is changed into the optimization of fusion coefficient vector. Secondly, a RL agent is established to automatically learn the strategy though the intergradation with environment. Then the trained strategy is directly used for real-time HI construction. The effectiveness of proposed approach is verified though a real bearing dataset.