Remaining useful life prediction for multi-sensor mechanical equipment based on self-attention mechanism network incorporating spatio-temporal convolution

Abstract

Driven by the limitations of spatial feature extraction in graph learning methods of multi-sensor mechanism equipment, this paper proposes a spatio-temporal self-attention mechanism network (STCAN) that integrates spatial relationships and time series information to predict the remaining useful life (RUL). Firstly, a graph convolutional network (GCN) is applied to extract the spatial correlation characteristics and fused with the self-attention mechanism network to obtain the global and local spatial features. Subsequently, a dilated convolutional network (DCN) is integrated into the self-attention mechanism network, to extract the global and multi-step temporal features and mitigate long-term dependency issues. Finally, the extracted spatio-temporal features are used to predict the equipment’s RUL through fully connected layers. The experimental results demonstrate that STCAN outperforms some existing methods in terms of RUL prediction.