Attention-aware temporal-spatial graph neural network with multi-sensor information fusion for fault diagnosis

Intelligent fault diagnosis has attracted intensive efforts in machine predictive maintenance. However, the structural information from multi-sensor signals has not been fully investigated. In this study, a novel temporal-spatial graph neural network with an attention-aware module (A-TSGNN) is proposed to accomplish multi-source information fusion. First, the graph structure naturally organizes the diverse sensors. The graph convolution model realizes the feature representation in the spatial dimension. Then, time-dependent learning is applied in the temporal dimension, and a temporal-spatial learning framework is built. An additional attention module is designed to learn the flexible weights and model the importance of individual sensors and their correlations. Experiments on a wind turbine dataset achieves an accuracy of 0.9669 and an F1-score of 0.9649. For the gearbox dataset, the values are 0.9927 and 0.9920, respectively. The overall macro-average area under the curve metrics reach a perfect score of 1.00 for both datasets, indicating exceptional performance. The adaptive attention mechanism is also discussed to verify the superiority of the A-TSGNN. Furthermore, comparisons with the single-sensor scheme and other fusion models demonstrate the stable performance of the proposed method. The A-TSGNN provides a potential model for comprehensively utilizing multi-sensor data, showing a promising prospect.

Automated summary

This study proposes a novel temporal-spatial graph neural network with an attention-aware module (A-TSGNN) for multi-source information fusion, achieving exceptional performance on wind turbine and gearbox datasets.