Intelligent diagnosis model for abnormal noise of power transformer based on multi-channel wavelet convolution

Abnormal acoustic signals from transformers can reflect mechanical and electrical defects. This paper proposes an intelligent diagnosis model for abnormal transformer acoustic signals by combining wavelet convolution with MobileNetV2. First, a set of time-frequency plots is constructed using learnable multicategory wavelet convolution. Parameter training is used to find the optimal frequency band parameters and extract the characteristics of the acoustic signal's different frequency bands. Waveforms from different wavelet basis functions can also be used to extract different characteristics. Feature extraction and classification are then performed using the lightweight MobileNetV2 model, reducing hardware dependency and improving the model's lightweight deployment capabilities. Experimental training was conducted using abnormal noise data collected from normal operation, DC bias, discharge, and winding deformation. Results show that the proposed model achieved an average accuracy of 97.64% and a maximum accuracy of 99.35% for abnormal noise recognition after 30 training rounds. This model also achieved a 10.7% improvement in training efficiency compared to traditional CNN models. Combined with the advantages of MobileV2's lightweight deployment, it has significant application value in the field of intelligent diagnosis of abnormal transformer noise.