Fault Voiceprint Blind Domain Diagnosis Technology Based on Multi-scale Spatio-Temporal Masks

The fault voiceprint diagnosis in the blind area scenario of municipal solid waste incineration power plants is restricted by problems such as complex environmental noise, differences in equipment models across plants, and scarcity of samples. This paper proposes a diagnostic technology based on multi-scale spatio-temporal masks. A 10ms-1 s time mask is generated through a parallel multi-branch encoder to extract and fuse multi-scale temporal series features. The spatial coherence of the microphone array is utilized to generate spatial masks to suppress environmental noise, and the spatio-temporal masks are fused in combination with the attention mechanism. Experiments show that compared with traditional deep learning models, the fault recognition rate of this technology in cross-factory scenarios has increased from 68% to 86.3%, the false alarm rate has decreased to 6.2%, the F1-score has increased by 19%, and the diagnostic accuracy for non-standard mixed fault samples exceeds 97%, significantly improving the diagnostic accuracy and reliability.