Aiming at the difficulties in data acquisition and low accuracy of online monitoring of power transformers, this paper proposes a method for identifying the state of transformer windings based on frequency domain decomposition and multi-field coupling to achieve online detection of windings. Firstly, the vibration winding model is analyzed by the method of multi-physics coupling, and the axial vibration distribution is obtained. Secondly, feature extraction and data analysis of vibration signals are carried out through frequency domain decomposition, and finally, the online identification of transformer winding faults is realized through probabilistic neural network learning and classification. The model was verified on a transformer model prototype with a capacity of SOOKVA. The calculated value is in good agreement with the experimental value, which verifies the accuracy and effectiveness of the proposed method.