Fault Voiceprint Recognition Method Based on Cross-modal Distillation and Semantic Calibration

To address the issue of acoustic signals being easily interfered by noise and insufficient information in traditional fault acoustic recognition methods, this paper proposes a fault acoustic recognition method based on crossmodal distillation and semantic calibration. By introducing vibration modality as auxiliary information, a modal interference suppressor and a semantic calibration module are designed, combined with triplet loss and adaptive contrastive loss functions to enhance feature representation and transfer efficiency. Experimental results show that the proposed method achieves recognition accuracies of 98.57% and 95.13% on the UORED-VAFCLS and JUST datasets, respectively, significantly outperforming existing methods and effectively improving the accuracy and robustness of fault recognition.