Remaining useful life estimation with multiple local similarities

Abstract In prognostics and health management (PHM), remaining useful life (RUL) estimation has become a major focus for guaranteeing the safety and reliability of systems. Similarity-based RUL estimation methods, which predict a system’s RUL based on the RULs of other systems with similar degradation behaviors, have been proven effective when no or limited mechanism knowledge is available. Global similarity-based approaches apply the whole-life history to find similar degradation patterns and may lead to few or even no candidates. In contrast, local similarity-based methods only utilize the data close to the prediction time, and then, false positives are inevitable. A given system may have experienced several events before being tested for RUL, and each event may impact its RUL. Systems that have undergone similar events will probably degrade similarly in the future. Hence, the past events must be effectively identified and fully utilized. This paper proposes estimating a system’s RUL by using multiple impacts from its past. The system’s history is transformed into a set of local segments by which the degradation events are represented. Then, a coarse-to-fine strategy is introduced to efficiently locate the events similar to the test. The most similar segments are regarded as references, and their corresponding RULs are a natural data basis for RUL estimation. Since segments may correspond to different features, we adopt two adjustment strategies to make reference RULs more applicable. A self-adaptive weight allocation method is also proposed to further improve the prediction performance. The experimental results show the effectiveness and advantages of our proposed method.