A neural turing machine-based approach to remaining useful life estimation

Estimating the Remaining Useful Life of a mechanical device is one of the most important problems in the Prognostics and Health Management field. Being able to reliably estimate such value can lead to an improvement of the maintenance scheduling and a reduction of the costs associated with it. Given the availability of high quality sensors able to measure several aspects of the components, it is possible to gather a huge amount of data which can be used to tune precise data-driven models. Deep learning approaches, especially those based on Long-Short Term Memory networks, achieved great results recently and thus seem to be capable of effectively dealing with the problem. A recent advancement in neural network architectures, which yielded noticeable improvements in several different fields, consists in the usage of an external memory which allows the model to store inferred fragments of knowledge that can be later accessed and manipulated. To further improve the precision obtained thus far, in this paper we propose a novel way to address the Remaining Useful Life estimation problem by giving an LSTM-based model the ability to interact with a content-based memory addressing system. To demonstrate the improvements obtainable by this model, we successfully used it to estimate the remaining useful life of a turbofan engine using a benchmark dataset published by NASA. Finally, we present an exhaustive comparison to several approaches in the literature.