Sensor-aware CapsNet: Towards trustworthy multisensory fusion for remaining useful life prediction

Abstract

Multisensory data-driven remaining useful life (RUL) prediction based on deep learning techniques is gaining increasing popularity as it captures the degradation process of mechanical equipment more comprehensively, the prediction performance of which is generally higher. However, less emphasis has been placed on enhancing the trustworthiness of the model, especially in terms of reasonability and interpretability in uncertainty-aware prediction. Aiming to solve such a problem, a novel framework called a sensor-aware capsule neural network (SACN) is proposed for multisensory fusion in RUL prediction. Then a novel Monte Carlo-based dropout method based on the activation level of capsules was developed for uncertainty quantification. Furthermore, the proposed method allows for multisensory fusion interpretation to attribute prediction results and uncertainties to the relevant features across sensors. A case study on C-MAPSS dataset shows that the proposed method outperforms the popular deep learning-based methods. In particular, the SACN based on capsule dropout achieves both greater accuracy and better reasonability than Gaussian dropout, as evaluated by conventional accuracy metrics and the proposed uncertainty coverage score, and the relationship between input features and final results can be revealed through multisensory fusion interpretation, which evaluates the contributions of each sensor for the given input.