An adaptive multi-scale spatial-temporal graph attention ensemble network with physical guidance for remaining useful life prediction of multi-sensor equipment

With the arrival of Industry 4.0, advanced sensors are used to monitor real-time status information, facilitating contactless equipment health monitoring. However, the inherent spatial-temporal dependencies and scale differences of multi-sensor data pose a challenge in learning discriminative degradation features and fully utilizing effective information for remaining useful life (RUL) prediction. Secondly, most existing methods lack engineering physics guidance for model training, limiting their performance and application. Therefore, this work proposes an adaptive multi-scale spatial-temporal graph attention ensemble network with physical guidance for RUL prediction of multi-sensor equipment. Firstly, an adaptive multi-scale graph attention network is constructed, exploiting a self-attention mechanism and two explicit sparse (ExpSparse) self-attention mechanisms to model the spatial dependencies of multi-sensor data at various scales, and incorporating a graph scale attention mechanism to adaptively perceive the importance of various spatial scales. Then, multi-scale spatial information is fed into an adaptive bidirectional temporal convolutional network to capture temporal dependencies. Furthermore, a physics-informed loss function balancing economy and safety is designed to guide the training of the proposed model. Finally, experiments on three benchmark datasets, engine degradation, tool wear, and lithium battery, demonstrate that the proposed model outperforms state-of-the-art baselines and can effectively balance economy and safety.