The two currently existing fault diagnosis methods for rolling bearings based on signal processing technology and big data processing technology have the disadvantages of over-reliance on signal processing, complicated model and weak interpretability. Aiming at the shortcomings of traditional fault diagnosis technologies, this paper introduces the time series classification method based on shapelets learning algorithm into the field of fault diagnosis, and establishes the unbalanced data set of the faults of the EMU axle box bearing through the EMU wheelset bench rolling vibration experiment. The diagnosis model is improved based on the idea of Dropout. Experiment results show that the method guarantees the accuracy of fault diagnosis while retaining the strong interpretability of the shapelets as “the most representative time series subsequences”. At the same time, the improvement of the model based on Dropout improves the generalization performance of the model. The diagnostic accuracy of 100% on the training set and test set of bearing fault data was achieved, which proves that the improved learning algorithm based on shapelets is a feasible method applied to the fault diagnosis of axle box bearing of electric multiple unit.