MST-GAT: A multi-perspective spatial-temporal graph attention network for multi-sensor equipment remaining useful life prediction

Abstract

Graph neural networks (GNNs) have been applied as an emerging technique for remaining useful life (RUL) prediction due to their ability to effectively exploit the dependencies between multi-sensor data. Despite existing GNN-based RUL prediction methods have achieved promising results, they do not consider the spatial-temporal dependencies of multi-sensor data from different perspectives (global and local), making it difficult to fully utilize the useful information in multi-sensor data. Secondly, traditional point prediction method can only provide the possible future RUL prediction value, but cannot quantify the uncertainty of RUL prediction, making it difficult to provide more information for decision-making. To address the above challenges, we propose a multi-perspective spatial-temporal graph attention network (MST-GAT) for multi-sensor equipment RUL prediction. MST-GAT constructs a multi-perspective graph attention network (MP-GAT) and a temporal convolutional network (TCN) to capture complex spatial-temporal dependencies in multi-sensor data. Specifically, MP-GAT employs a multi-head self-attention mechanism and two multi-head explicit sparse self-attention mechanisms to explicitly model global and local spatial dependencies, and feeds multi-perspective spatial information into TCN to capture temporal dependencies. Furthermore, a joint optimization network is constructed for predicting RUL and its prediction interval to quantify uncertainty. Experiments on three benchmark datasets demonstrate that MST-GAT outperforms state-of-the-art baselines in RUL prediction and can accurately quantify uncertainty.