The prediction of the Remaining Useful Life of a machine component represents a strategic problem in predictive maintenance, which may have important consequences for a company. Recent approaches to this problem leverage data-driven methodologies based on deep learning, achieving impressive results. In particular, due to the temporal nature of the sensor measurements detailing the life of a component, neural sequence models are often chosen to automate the feature extraction process. In this paper, we investigate several of these models on a particle filtration system. The experiments performed present the good prediction capabilities of these models, highlighting some of them for their accuracy. Nonetheless, the qualitative analysis shows that when the fault is farther away, most of these models tend to have unstable predictions. These results motivate some future research directions which are discussed in the conclusions.
