Multi-view fully connected graph to fuse multi-sensor signals for mechanical equipment remaining useful life prediction

Abstract

Accurate remaining useful life prediction is essential for enhancing equipment reliability and optimizing maintenance strategies. However, existing methods struggle to effectively integrate multi-sensor data while quantifying uncertainty. To address these challenges, a multi-view fully connected graph neural network is proposed for multi-sensor mechanical equipment remaining useful life prediction. Firstly, local fully connected graphs and global graphs are constructed to comprehensively characterize the multi-view spatial correlations from global and local views. Meanwhile, the graph convolution operations are performed on local and global graphs to extract the intricate spatial dependencies within multi-sensor signals. Then, the learned multi-view spatial representations are fed into the temporal convolutional network to capture the temporal dependencies across sensor timestamps. Finally, a joint optimization network is developed to simultaneously predict the remaining useful life and its associated prediction interval, enabling uncertainty quantification. Extensive experiments on two multi-sensor monitoring degradation datasets demonstrate the superior performance of the proposed model, offering valuable technical support for predictive maintenance.