Winding failures are recognized as one of the primary causes of transformer accidents, making effective monitoring of winding conditions crucial. A study on autotransformer (AT) winding faults diagnosis is conducted through the following procedure. Firstly, an experimental platform is established to simulate typical single and combined winding faults in autotransformers, through which frequency responses under various fault conditions are tested. Subsequently, a fast vector matching method is employed to fit transfer functions of winding systems under normal and faulty states, from which zero point distribution diagrams in polar coordinates are derived. Then, the gray level difference statistical (GLDS) features and gray-gradient co-occurrence matrix (GGCM) features are extracted from the zero point distribution diagrams, and the particle swarm optimization (PSO)-random forest (RF) algorithm is combined to realize the classification of faulty windings and fault types. Finally, the proposed method is validated using actual autotransformer fault cases. The results show that the zero point distributions in polar coordinates obtained by fast vector fitting can capture the subtle differences in the original frequency response curves by combining amplitude-frequency and phase-frequency information. Compared with optimization algorithms such as cuckoo search and genetic algorithm, the PSO-RF algorithm maintains an accuracy rate consistently exceeding 93% in identifying winding faults and fault types of autotransformers. The analysis results of the proposed method are consistent with the tank lifting inspection results in real autotransformer fault cases.