Just-in-time learning based probabilistic gradient boosting tree for valve failure prognostics

— Historical failure instances of a system with diversified degradation patterns will pose great challenge for prognostics. Consequently, it is challenging to accurately predict the remaining useful life (RUL) using a prognostic model trained from such data. To solve this problem, this paper proposes a just-in-time learningbased data-driven prognostic method for reciprocating compressors with diverse degradation patterns and operating modes. The proposed framework employs a just-in-time learning (JITL) scheme to deal with the stochastic nature of fault evolution and the diversity of degradation patterns. Moreover, a data-driven forecasting model that features a randomized and smoothed gradient boosting decision tree (RS-GBDT) is developed for RUL and uncertainty predictions. The effectiveness of the proposed approach was validated on temperature measurements collected from 13 valve failure cases of an industrial reciprocating compressor.