The condition monitoring data of emulsion pump follow the long-tail distribution. The amount of monitoring data for the normal condition is very large, while the amount of monitoring data for different fault conditions is very small, the problem of class-imbalance is prominent. The traditional intelligent fault diagnosis methods are proposed under the assumption of class balance, which the fault diagnosis model has the shortcoming of insufficient generalization ability when dealing with the class-imbalance problem. Thus, an end-to-end fault diagnosis method for emulsion pump with class-imbalance is proposed. conditional variational autoencoder is used to extract features and learn the state data distribution of emulsion pump, and the loss value of training samples is adjusted based on focal loss to balance the influence of different types of data on the model. Moreover, the end-to-end fault diagnosis model can be obtained based on the decoder model. Finally, the effectiveness of the proposed method is verified by simulation experiment data of emulsion pump faults. Compared with other methods under different types of imbalanced rates, the results show that the fault of emulsion pump can be accurately identified under the condition of only a small amount of fault data by the proposed method and the corresponding recognition accuracy is better than other methods.