Interaction models for remaining useful lifetime estimation

Abstract

The paper deals with the problem of controlling the state of industrial devices according to the readings of their sensors. The current methods are based on an approach to feature extraction in which the prediction occurs. We propose an interaction method of multiple blocks of different complexity, which aggregate information differently over time, to create a common latent space for RUL prediction, and train the resulting architecture in a single pass with a new loss function aimed at heterogeneous latent space. A new TFI model based on sensor reading-wise information aggregation and adapted hierarchical convolution achieved state-of-the-art results on the C-MAPSS dataset.