Power transformers are the core equipment of the power system. The reliable assessment of the mechanical condition of their windings is of vital importance for the safety of the power grid. Currently, the detection of winding deformation mainly relies on offline methods, which require disassembling the transformer leads or applying additional excitation signals to conduct tests. This approach has problems such as complex detection procedures, long cycle, low efficiency, and high dependence on manual analysis, making it difficult to meet the engineering requirements for rapid response and real-time monitoring after an accident. This paper focuses on the research: Based on the principle of Lissajous figure method, the extraction and characterization mechanism of transformer winding fault characteristics is studied; Through the construction of a transformer field-path coupling simulation model, the simulation of winding faults is completed, and the simulation characteristics of different types and degrees of faults are analyzed; To verify the effectiveness of the method, physical experiments of winding deformation are systematically carried out, and a load standardization algorithm adapted to the actual operating conditions is proposed. Further, based on the characteristic graphs and characteristic parameters of different fault types, a winding fault classification diagnosis index system is constructed, and the Softmax+SVM artificial intelligence model is used for training to achieve intelligent diagnosis of winding faults. The results show that the accuracy of winding fault diagnosis reaches 97.3%, and through transformer fault examples for testing and verification, it can provide effective decision support for transformer intelligent operation and maintenance.