Voiceprint Diagnosis Method for Urban Rail Power Transformers Based on Mel Spectrogram and Improved Vision Transformer

The fault diagnosis for urban rail power transformers is crucial for ensuring the safety and reliability of urban rail power systems. Voiceprint-based fault diagnosis methods have gradually become a research hotspot due to their non-contact and uninterrupted advantages, especially with significant progress in the application of deep learning methods in this field. However, challenges such as poor robustness against noise interference and scarcity of real fault samples still persist. Therefore, this paper proposes a voiceprint diagnosis method for urban rail power transformers based on Mel spectrogram and an improved Vision Transformer(ViT). The improved ViT model integrates a CNN encoder and adopts a column-wise patch strategy, building upon the original ViT architecture. Experimental results show that the proposed method, with only 0.47M parameters and 0.149GLOPs, achieves a mere 6.75% accuracy drop when the signal-to-noise ratio decreases from 10dB to -4dB, outperforming other comparative methods.