Spectral-domain Spatial-temporal Convolution Graph Neural Network for Industrial Fault Diagnosis

Abstract

Abstract In modern industrial production, once a fault occurs in the production process, it will affect the whole production and bring huge economic losses. There are a large number of device sensors in modern production, and there are complex interactions between them. But the existing methods can not efficiently and accurately mine the depth of information from mass production data. We proposed the Spectral-domain Spatial-Temporal Convolution Graph Neural Network model, which comprehensively considers the information of spatial and temporal dimensions in production data. The nodes and edges of the heterogeneous graph represent the sensor signals and their interactions. Each sensor data is converted to the spectral domain to extract the time domain features, and graph convolution is used to carry out the spatial relationship between input sequences. Finally, the proposed model is verified using the Three-Phase Flow Facility (TFF) dataset, and the overall accuracy rate reached 96.28%, showing better performance than other baseline models.