An exploration was conducted into the performance of wheel flat recognition driven by various structured data to achieve efficient and accurate detection of wheel flats, and a new quantitative identification method was proposed. Synthetic wheel flats data were input into the metro vehicle?rail rigid-flexible coupling dynamics model as wheel out-of-roundness excitation to obtain axle box vibration responses under different working conditions. The axle box vibration responses were regulated into sample sets of various structural forms. The sample sets were fused with speed signals and input into a multi-input convolutional neural network (MCNN) model for training, and the differences in performance of MCNN model under different data structure inputs were explored. Results show that compared with other input data structures in the setup, MCNN model recognition performance is best when the input data structure is a combination of time domain, frequency domain and time-frequency domain, with mean absolute percentage error and R-squared coefficient (R²) reaching 1.947% and 0.9978 respectively, and relatively low time-consumption, with 0.157 9 ms for a single sample. Results of the classical model comparison experiment, speed information ablation study, and real-data transfer learning experiment show that the MCNN model for engineering applications when the input data structure is a combination of time domain, frequency domain, and time-frequency domain.