As a common defect of heavy-haul railway wagons, the wheel diameter difference (WDD) will deteriorate wheel/rail dynamic interaction, which severely threatens the running stability and safety of the wagons. Accordingly, it is of great importance to detect the WDD forms of wagons and take appropriate measures in time. In this paper, a novel method for WDD form detection of the running wagons by analyzing axle box acceleration signals is proposed. To solve the difficulty of weak feature extraction for the vibration signals, a novel feature extraction method combining adaptive chirp mode decomposition (ACMD) with fractal box dimension (FBD) is proposed. Firstly, a 3D feature space is constructed by the FBDs, which is calculated from the extracted chirp modes by ACMD. Then, a multiple kernel extreme learning machine optimized by genetic mutation particle swarm optimization is developed for the classification of the feature vectors. Both simulation and field test results indicate that the proposed detection method is powerful for accurate identification of the wheelsets with standard diameter, in-phase WDD, and anti-phase WDD, and the algorithm efficiency shows practicability in onboard monitoring.