Remaining useful life prediction via a deep adaptive transformer framework enhanced by graph attention network

Abstract

Accurate monitoring of mechanical device conditions requires a large number of sensors working together. There are potential connections between sensors throughout the degradation monitoring process of mechanical devices. Conventional deep learning (DL) models suffer from the following shortcomings when dealing with this type of multi-sensor degraded data. To begin with, most existing methods based on DL mainly use CNN as the feature extractor, focusing too much on temporal correlations and ignoring spatial correlations of multiple sensors. Then, the most popular remaining useful life (RUL) model is based on recurrent neural network, which oftentimes suffer from the issue of gradient exploding and vanishing. Therefore, a bran-new end-to-end framework based on a deep adaptative transformer enhanced by graph attention network, named GAT-DAT, is proposed to tackle these weaknesses. First, the graph data is constructed by the correlation of sensors. Next, GAT submodules fuse node features to extract spatial correlation. Finally, the DAT submodule is used to efficiently abstract the temporal features of the data through a self-attention mechanism and adaptively implements RUL prediction for mechanical equipment. Two case studies are employed to attest the efficacy of our proposed GAT-DAT model and the analysis of the experimental data illustrates that the GAT-DAT framework outperforms the existing state-of-the-art methods.