关键词: 电机轴承, 声纹识别, 故障诊断, 图卷积网络

Abstract: A Dynamic Spatial-Temporal Sub Graph Convolutional Network (DSTSGCN) was proposed to address limitations in existing voiceprint-based fault diagnosis methods,including poor adaptability and insufficient dynamic graph construction capabilities caused by difficulties in capturing spatial-temporal coupling relationships. Firstly,a edge-level dynamic graph convolutional network was designed,where spatial correlations between voiceprint signals were adaptively learned through optimization of edge weights,and the sensitivity of traditional k-nearest neighbor graphs to hyperparameters was effectively mitigated.Secondly,dilated causal convolution was integrated with cross self-attention mechanisms,and a temporal feature fusion module was constructed to capture critical long-term dependency information within signals,thereby highlighting temporal relationships between signals. Finally,discriminative representations of fault features were enhanced through multi-signal spatial-temporal information fusion,enabling graph-level fault diagnosis. Experimental validation was conducted on a three-phase asynchronous motor test platform. Results demonstrated that the proposed DSTSGCN achieved 99.72 % accuracy with only one training sample,outperforming seven comparative voiceprint-based diagnostic methods.

Key words: motor bearing, voiceprint recognition, fault diagnosis, graph convolutional network