Abstract: Autotransformer (AT) is one of the key devices of high-speed railway power supply system. Split winding structures are adopted for impedance requirements in high-speed railway AT, which are connected inside transformer. Frequency response analysis (FRA) curve of each split winding cannot be directly measured, thus it is difficult for the ex-isting FRA model and diagnostic method to distinguish the faults. In this paper, the method of FRA interpretation based on multi-decomposition and image features is proposed to obtain more information on the frequency response curves, which helps to analyze the small changes. Verification and analysis are carried out by conducting series capacitance variation (SCV) experiments on different windings and different positions in the split winding AT. The results show that the multi-level decomposition FRA curve has a strong correlation with original FRA curve. After being converted into a polar coordinate graph, the graph is different under different fault conditions. The features correlation degree (FCD) values of different fault windings are obviously different, and different faults have their corresponding laws. According to the 12 image features extracted from polar graph and corresponding features correlation after seven-level decomposition, the SCV at different fault windings, fault locations and fault degrees in autotransformer can be distinguished.