A Semisupervised GCN Framework for Transfer Diagnosis Crossing Different Machines

When it comes to intelligent diagnosis, the transfer of knowledge and experiences across machines is crucial. By transferring the learned fault features and abstract representations from one machine to another, the intelligent diagnosis algorithm can adapt to new scenarios, thereby enhancing its versatility and robustness. However, for some transfer learning-based diagnosis methods, the accuracy of cross-working condition diagnosis is satisfactory, whereas the accuracy of cross-equipment diagnosis is disappointing. To address this issue, a semisupervised graph convolutional networks framework, named SSGCN, was proposed for transfer diagnosis across different machines. To align feature distributions of different domains and aggregate features within the same category, the multiobjective optimization loss function (MOOLF) was constructed. Taking advantage of graph structure data, a weighting method that defined the edge between nodes using Chebyshev distance, called WECD, was proposed for effective feature propagation and improved learning performance. Furthermore, a pseudo-label was added to unlabeled graph structure data in the target domain to enhance positive migration for transfer diagnosis. Experimental results across different machines demonstrate that the proposed SSGCN framework exhibits superior performance.