Separable physical spatiotemporal graph message aggregation for fault diagnosis

Abstract

Spatiotemporal graph has become a research hotspot for it can excavate spatiotemporal information in multi-sensor fault diagnosis. However, the existing methods do not fully consider the physical attenuation characteristics in edge when the fault features are transmitted to the next sensor in the case of cross-sensor spatial temporal correlation. Besides, existing spatiotemporal convolutional networks pay much attention to the integration of all nodes for information update and the network structure design without realize the aggregation of edge information with different attributes. To address these issues, we propose Separable Physical Spatiotemporal Graph Message Aggregation (SPSGMA) for Fault Diagnosis. Firstly, a spatiotemporal graph of physical connection properties across sensors is proposed to assign different properties to different edges. Then, a novel wavelet frequency selection method is proposed for node feature extraction of different physical edge. Finally, a separable message aggregation network is designed to realize aggregation of frequency messages on different physical edges and classification rather than unified feature extraction. Three different datasets are used to verify the effectiveness of SPSGMA. Compared with other methods, SPSGMA achieves the best diagnostic performance in long chain sensor data diagnosis, and its average diagnosis accuracy in different diagnosis respectively are 99.99%, 98.59%, and 99.93%.