Remaining Useful Life Prediction Method Based on Multisensor Fusion Under Time-Varying Operating Conditions

Accurate prediction of remaining useful life (RUL) serves as the foundation for predictive maintenance of industrial equipment. In recent years, the fusion of multi-source information has achieved remarkable advancements for the development and application of RUL prediction. However, under time-varying operating conditions, the distribution of monitoring data exhibit time-varying characteristics, posing two challenges for RUL prediction in this scenario. One is adaptive decoupling of operating condition data and monitoring data, and the other is adaptive weighting of multi-source information. To address these challenges, a novel method for RUL prediction is proposed in this paper driven by the fusion of multi-source information under time-varying operating conditions. The proposed approach is designed to track the degradation process of equipment in scenarios involving cyclic variation and multiple levels in operating conditions. An optimization function is constructed to comprehensively characterize the frequency domain distribution of current signals and the continuity of health index over time. Then, a time-varying observation matrix for the degradation state space model is derived, which aims to eliminate the influence of operating condition data on degradation information. Two Kalman filter models are developed based on linear degradation model and double exponential degradation model focused on different stages of equipment degradation, which can calculate time-varying weights for vibration and sound information at different time coordinates. In this way, a multidimensional data mapping from multi-source information to the degradation curve is established under time-varying operating conditions. In order to verify the superiority of the proposed method in RUL prediction, two sets of run-to-failure experimental dataset are studied and analyzed. The result demonstrates that the proposed method achieves superior performance compared with single-source information methods.