Due to the inherent defects of traditional spectrum estimation methods， it is difficult to accurately identify the abnormal wear signal of the wheel， especially for the identification of the initial wheel polygonal wear. To solve the above problem， a dynamic detection method of wheel polygon is proposed based on parametric power spectrum estimation. Firstly， the harmonic frequency recovery model is established according to the dynamic characteristics of the wheel polygon. Secondly， the order of the harmonic recovery model is determined in terms of the singular value decomposition method and normalized error analysis. Thirdly， the total least square method is used to calculate the parameters of the harmonic recovery model. Finally， the power spectrum of the abnormal wear signal is estimated according to Cadzow's estimation theory. Taking the measured signal of axle box vertical vibration acceleration of a subway vehicle as an analysis case， the feasibility and correctness of this method are verified. The results show that this method can realize high-precision frequency-domain estimation based on short-time sequence data， which is sensitive to harmonic signals， especially suitable for identifying initial wheel polygon abnormal wear.