RUSHAP: A Unified approach to interpret Deep Learning model for Remaining Useful Life Estimation

Abstract

The maintenance decision models based on Prognostic and Health Management (PHM) technology have significantly improved complex equipment submission reliability and economy. One of the essential techniques of PHM is predicting the remaining useful life (RUL) of the system or the system components. Compared to other RUL prediction methods, deep learning has become a research hotspot due to its automatic feature extraction capability, big data process efficiency, powerful representation of complex mappings, and “end-to-end” learning process. However, deep learning (DL) models are with high complexity, huge parameter quantity, and low interpretability, namely black box models. Lack of interpretability limits their application and development in “high-risk” fields such as aviation maintenance decision-making. To solve this problem, we propose a universal RUL interpretation method for DL named as RUL Shapley Additive explanation (RUSHAP). RUSHAP uses the input and output of the DL model to calculate the Shapley value and then obtain the interpretation from three different hierarchies, i.e., time level, feature level, and component level. With RUSHAP, it is possible to go from only knowing the RUL of the system to locating fault state points, observing the declining trend of sensor data, and evaluating the health status of subsystem, achieving partial white-boxing of the RUL prediction DL model. RUSHAP can also compare the advantages and disadvantages between different DL models, giving references for model debugging and ideas for model design.