A fault diagnosis method has been proposed for hydraulic dynamometer bearings based on speed tracking processing and convolutional neural network (CNN) for an accurate diagnosis of the bearing fault status caused by frequent bearing failures in complex environments with high speed and heavy load. Firstly, speed tracking preprocessing is performed on the vibration data in view of a solution of feature extraction under variable speed conditions. Next, the preprocessed signal is subjected to fast Fourier transform for an extraction of fault spectrum features, with the frequency domain signal used to generate a Hank matrix, which is subsequently used as the original feature input to train a convolutional neural network fault diagnosis model. Finally, the trained bearing fault diagnosis model is used to achieve an online diagnosis of hydraulic dynamometer bearing faults. The simulation results show that the proposed hydraulic dynamometer model based on speed tracking processing and convolutional neural network is characterized with an improved generalization ability, feasibility and effectiveness, which can accurately and effectively diagnose the fault status of bearings.