LAUSR

Abstract Recent advances in sensor technology have made it possible to monitor the degradation of a system using multiple sensors simultaneously. Accordingly, many neural network-based prognostic models have been proposed to use observed multiple sensor signals as inputs and estimate the degradation status or failure time of the system. Although these models have achieved promising prognostic performance, it is still difficult to interpret the extracted features, and the models are often used in a black-box manner providing only the final results. In this study, a novel sensor selection framework is proposed to address this challenge by adaptively deciding which sensors to use at the moment to enhance remaining useful life prediction. The contributions of this work are summarized as follows: (1) being generic and can be attached to a variety of existing neural network-based prognostic models; (2) being trained in a unified manner to optimize both the sensor selection and prognostic accuracies simultaneously; (3) improving the interpretability of the model by explaining how different sensors contribute to the final remaining useful life prediction of individual systems over time; and (4) introducing several regularization techniques to ensure the stability of the training process. We validate the proposed framework using a series of numerical studies on the degradation of aircraft gas turbine engines.