Gearboxes are one of the essential components in the railway vehicle, and their fault diagnosis is critical to safe operation. Traditional deep learning is difficult to accurately identify the gear’s health status under variable conditions and small sample size. To tackle this problem, we propose an ensemble decision approach that combines the dislocated time–frequency representation and a pre-trained convolutional neural network (CNN) to evaluate the gear’s health status. The experimental results indicate that the continuous wavelet transform and the synchrosqueezed transform have better diagnostic performance than the time-domain signal and the short-time Fourier transform. Also, the dislocated operation helps the CNN learn the characteristics of continuous signals more profoundly and increases the sample size. Moreover, the ensemble decision can improve the accuracy and stability of diagnosis. Consequently, the proposed framework can effectively diagnose railway vehicle gearboxes and significantly enhance CNN’s robustness and generalization under a limited sample size.