Affected by the variable speed operation of vehicles， random irregularity of tracks and other factors， the operation process of rail vehicles is a typical non-stationary process. It is difficult to accurately identify the wheel polygon wear order and amplitude， because the traditional methods are not ideal for non-stationary processing signals. A dynamic detection framework based on iterative modified discrete Fourier transform （DFT） is proposed to improve the detection accuracy of wheel polygon wear state. The vertical vibration acceleration signal of axle box is used for analysis， when the vehicle is running at a relatively stable speed. Firstly， relatively stationary short-time signals are extracted from the sample signals by setting appropriate stability test conditions. Secondly， the frequency domain analysis and iterative calculation of the extracted short-time signal are carried out to obtain each order wheel polygon's vibration frequency and period. Then， according to the length of the vibration period， the extracted short-time signal is truncated twice to obtain a new short-time signal representing an integer multiple of the wheel polygon's vibration period. Finally， combined with the geometric and dynamic characteristics of the wheel polygon， the frequency domain analysis and wear parameters （order and amplitude） of the extracted new short-time signal are calculated again to realize the accurate identification of the wear state of the wheel polygon. The verification analysis shows that the identification framework effectively reduces the identification error caused by inherent defects such as the fence effect and spectrum leakage in traditional analysis methods， eliminating the interference of most non-stationary factors and provide theoretical support and method reference for rail vehicles' safe operation and maintenance.