Fault diagnosis of resistance variation in mine ventilation systems based on auxiliary sample-guided fusion graph convolutional network

Resistance variation faults in mine ventilation systems are formed by local resistance changes propagating along the ventilation network and generating correlated responses at multiple monitoring points. During long-term operation, mines accumulate a large amount of unlabeled monitoring data, but information related to resistance variation faults has not been effectively used. Response relationships among multiple monitoring points are insufficiently used, and it is difficult to simultaneously characterize the overall linkage changes among multiple monitoring points and the local differential responses related to fault locations during resistance variation fault propagation. To address these problems, a resistance variation fault diagnosis method for mine ventilation systems based on an Auxiliary Sample Guided Fusion Graph Convolutional Network (ASGF-GCN) was proposed. The Unlabeled Auxiliary Sample Mining Mechanism (UASMM) was used to select samples with response patterns similar to labeled fault samples from unlabeled operating samples as unlabeled auxiliary samples. The Similarity Weighted Fusion Mechanism (SWFM) was used to weight and fuse labeled fault samples and unlabeled auxiliary samples to construct fused graph representations, thereby enhancing the contribution of highly correlated unlabeled auxiliary samples. A Frequency Adaptive Graph Convolutional Network (FAGCN) was introduced to adjust the information propagation intensity of neighboring nodes through a self-gating mechanism, considering both the overall linkage characteristics of multiple monitoring points and the local differential responses related to fault locations. Experimental results showed that the accuracy and F1 score of ASGF-GCN were 88.89% and 88.87%, respectively, which were 4.17% and 4.28% higher than those of the Graph Convolutional Network (GCN). The method had stable diagnostic performance under limited labeled fault samples and monitoring noise disturbances. When deployed on a Jetson Xavier embedded device with Float 16 precision, the single-sample inference time was 19.62 ms, meeting the real-time requirements for online diagnosis of resistance variation faults in mine ventilation systems.