A Multiscale Graph Convolutional Neural Network Framework for Fault Diagnosis of Rolling Bearing.

Fault diagnosis for rolling bearings has been an important engineering problem for decades. To detect the damaged bearing surface, engineers analyze the features from the extracted vibration signals of the machine. As artificial intelligence rapidly develops and provides favorable effects in data analytics, using deep-learning technology to attack fault diagnosis problems has attracted increasing research interest in recent years. However, most existing methods do not provide satisfactory performance in mining the relationship between the signals. Even graph convolutional networks (GCNs) that can perform well in non-Euclidean spaces have limitations, since it only considers extracting features from a single scale, ignoring the potential relationship between signals at various scales. For salvation, we propose a multiscale graph convolutional network (MS-GCN) for this specific problem. In MS-GCN, we put forward a multiscale feature extraction module, which extracts the features of vibration signals two times under diverse receptive field ranges, ensuring that the regularities of signal features can be fully discovered. In addition, we devise a multiscale graph iteration module, which incorporates two single-scale graph iteration modules and a cross-scale graph iteration module, which can fully retain local features based on extensive mining of global information. We also propose a mutual fusion module based on the Bayesian method, which forcibly manipulates various features as a prior and achieves a convincing result. The horizontal visibility graph (HVG) method is used to construct graph models at multiple scales, which can better capture the hidden information of signal vibrations. In experiments, we verify the proposed model on the CWRU dataset to evaluate the method's performance. The results show that our model significantly improved accuracy compared with other state-of-the-art methods.