Graph Fusion and Propagation for Fault Diagnosis in Industrial Robots With Limited Labeled Data

Deep learning has propelled the advancement of fault diagnosis modeling for industrial robots. However, the scarcity of fault samples in industrial robot datasets constrains the improvement of fault diagnosis deep learning algorithm performance. Due to the graph-like nature of fault diagnosis data in industrial robots, semi-supervised learning (SSL) methods applied to industrial robots are primarily based on graph-based label propagation algorithms (LPAs). However, the fitting capability of LPAs is unstable, often resulting in less-than-ideal accuracy of the generated pseudo-labels. Therefore, this study proposes a novel fault diagnosis method for industrial robots, termed graph fusion and propagation (GFP). In GFP, multigranularity-based label propagation (MGLP) initializes and annotates the unlabeled dataset, generating two datasets of pseudo-labels: determinate and uncertain. These determinate pseudo-labels, combined with truly labeled data, form the training dataset for the spectral temporal graph neural network (StemGNN), which can extract both spatial and temporal features from the robotic monitoring data. Upon training completion, the uncertain pseudo-label set serves as the testing dataset. Subsequently, GFP refines the results by eliminating potential noise through comparison with the initial MGLP generated pseudo-labels. Experimental studies conducted using data collected from real-world industrial robots demonstrate that GFP achieves an accuracy of 85.4% in scenarios with constrained sample availability. This result underscores the method’s effectiveness in enhancing fault diagnosis when labeled data is insufficient. Moreover, GFP demonstrates decent performance compared to other state-of-the-art methods, further validating its efficacy in fault diagnosis for industrial robots.