An ensemble deep transfer learning method for fault diagnosis based on soft voting is proposed for diagnosing waterjet pump faults under variable working conditions. The source domain and few target domain data samples are normalized after FFT transformation and then fed into three deep transfer learning diagnosis models for training: the CORAL based deep transfer metric learning model, the MMD based deep transfer metric learning model, and the transfer component-based deep belief network. The target domain test samples are diagnosed and analyzed based on this approach. An ensemble deep transfer diagnosis model is subsequently established by combing the soft voting method to obtain the final diagnosis results. Through the diagnosis of three different types of faults in waterjet pumps under variable working conditions, the results show that the proposed ensemble deep transfer diagnosis model not only effectively addresses the high-precision fault diagnosis of waterjet pumps under variable working conditions, but also has better diagnostic accuracy than the single deep transfer fault diagnosis model.