Temporal Heterogeneous Relationships Enhanced Deep Clustering Network for Fault Diagnosis Using Unlabeled Multisensor Data

The persistent advancement of graph convolutional networks (GCNs) based fault diagnosis methods presents innovative and remarkable concepts for leveraging multisensor data effectively. Nevertheless, practical implementation poses significant challenges due to the scarcity of accessible data labels. Moreover, the temporal heterogeneity in multisensor data has not been explicitly explored. To tackle the aforementioned challenges, a novel temporal heterogeneous relationships enhanced deep clustering network is proposed for fault diagnosis. First, the heterogeneous graph gate recurrent unit (HGGRU) has been specially designed. By incorporating heterogeneous graph convolution operations into the GRU, the HGGRU can capture temporal heterogeneous relationships effectively. Then, the graph generation decoder is further employed to reconstruct the heterogeneous graph structures and the original multisensor signals from deep fault representation. Finally, a novel relationship-enhanced unsupervised fault diagnosis strategy is proposed to guide the training of the overall framework. To verify the effectiveness of the proposed method, extensive experiments have been conducted on real industrial production process datasets as well as an experimental bench dataset. The experimental results unequivocally demonstrate the superiority of the proposed method in unsupervised fault diagnosis, surpassing other comparative methods.