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. Results show that, on the test set containing one normal condition and six typical fault categories, the proposed method achieves an accuracy of 96.2%, a precision of 95.8%, a recall of 95.3%, and an F1-score of 95.6%, outperforming the GRU (gated recurrent unit), PNN (probabilistic neural network), TDNN (time delay neural network), and CNN-Transformer (convolutional neural network-Transformer) baselines across all four evaluation metrics. It maintains stable recognition performance even under 0 dB low signal-to-noise ratio conditions. This method effectively improves the accuracy and anti-noise ability of non-contact acoustic fault diagnosis for transformers, and is applicable to online monitoring scenarios in complex power grid environments.