A Novel Fault Diagnosis Method for Power Transformers Based on Voiceprint Recognition Considering Multitype Noises

Power transformer fault diagnosis based on voiceprint signals has gained substantial attention in recent years. However, the presence of various types of noise hinders the existing methods from achieving better diagnosis performance. Therefore, this paper proposes a novel fault diagnosis method considering persistent and transient types of noise. First, a voiceprint feature representation method is proposed that combines the improved group delay function phase spectrogram with the Mel amplitude spectrogram to achieve a more comprehensive expression of the voiceprint signal. Second, a triplet attention feature fusion mechanism is designed to strengthen the cross-dimensional interaction between phase and amplitude features and to suppress noise-sensitive regions. Third, a RepViT network enhanced with wavelet transform convolution based on structural reparameterization is constructed to perform multi-band decomposition and achieve focused analysis of key fault-related frequency bands. Experimental results show that this method can quickly and accurately identify the operating status of the transformer, and is superior to existing models. The accuracy rate reaches 97.25% in various noise scenarios, and the accuracy rate exceeds 95% in both transient and continuous high noise environments of -5 dB.