Abstract: To address the limitations of traditional contact-based monitoring methods, including complex installation, high cost, and limited noise resistance, an acoustic-based intelligent fault diagnosis method for power transformers is proposed using energy–phase joint modeling. A diagnostic network named TranSonicNet is constructed by integrating Mel amplitude spectra and modified group delay function (Modgdf) phase spectra. A Triple Attention Module (TAM) is introduced to achieve cross-dimensional adaptive feature fusion, and a lightweight Tiny Vision Transformer (TinyViT) architecture is employed for fault classification. Experiments are conducted on a 220 kV transformer acoustic dataset under multiple noise conditions. An accuracy of 96.2%, precision of 95.8%, recall of 95.3%, and F1-score of 95.6% are achieved, outperforming Gated Recurrent Unit (GRU), Probabilistic Neural Network (PNN), Time Delay Neural Network (TDNN), and Convolutional Neural Network (CNN)-Transformer models by up to 6.4%. Stable diagnostic performance is maintained under a 0 dB signal-to-noise ratio condition. Enhanced diagnostic accuracy and noise robustness are obtained for non-contact transformer acoustic monitoring in complex power grid environments.