Deep Graph Neural Network Fusing Multi-Sensor Physical Information in Aero-Engine Bearing Fault Diagnosis

As multi-sensor technology advances and data storage and processing capabilities improve, multi-sensor signals can provide a more comprehensive and multi-angle data perspective. This improves the reliability of fault diagnosis. The data of multiple sensors belongs to non-Euclidean data. Traditional neural network methods are very effective in processing Euclidean data. However, they have difficulties in processing non-Euclidean data and cannot fully utilize the structural information and similarity between sensors. On the other hand, graph neural network methods can handle nonEuclidean data, but most graph data inputs lack practical physical meaning. To address this issue, this paper presents a graph construction method based on space and similarity (SS method), which models multi-sensor data as graph data. This method can fully utilize the structural features between sensors, giving the graph data practical physical meaning. Then graph aggregation is performed on a batch of graph data samples. And the DeeperGCN model is applied to extract signal features, achieving recognition of the health status of aircraft engine bearings. Experiments have shown that this method can significantly improve the effectiveness of fault diagnosis.