LSTM and Transformers based methods for Remaining Useful Life Prediction considering Censored Data

Predictive maintenance deals with the timely replacement of industrial components relatively to their failure. It allows to prevent shutdowns as in reactive maintenance and reduces the costs compared to preventive maintenance. As a consequence, Remaining Useful Life (RUL) prediction of industrial components has become a key challenge for condition based monitoring. In many applications, in particular those for which preventive maintenance is the general rule, the prediction problem ismade harder by the rarity of failing instances. Indeed, the interruption of data acquisition before the occurrence of the event of interest leads to right censored data. There are few articles in the literature that take that phenomenon into account for RUL prediction, even though it is common in the industrial environment to have a high rate of censored data. The present article proposes a deep-learning approach based on multi-sensor time series which allows to consider censored data during the training of the neural networks. Two methods are proposed, respectively based on the Dual Aspect Self Attention based on Transformer proposed by (Z. Zhang, Song, & Li, 2022) for non-censored data and on a recurrent neural network. Their evaluation on the C-MAPSS benchmark dataset shows, compared to the state-of-the-art RUL prediction methods, no loss in the absence of censoring, and outperformance on censored data.