Abstract: Intelligent fault diagnosis method based on data-driven is the research focus of modern mechanical system. However，due to practical limitations，it is impossible to obtain samples of all working conditions or fault types，which makes the data-driven model lack of specific training data，leading to unsatisfactory performance of the method. In view of the above challenges，a rolling bearing fault diagnosis model based on contrastive embedding feature generation is proposed. By learning the faults with sufficient samples, the common features of existing faults and missing faults are mined to realize the generation and diagnosis of specific types of missing faults. Firstly, the spectrogram of the corresponding fault characteristics is obtained by analyzing the time-frequency characteristics of the original vibration signal; Secondly, the feature extraction network which has been pre-trained and fine tuned is used to extract the fault features in the spectrogram; Then, the extracted features are input into the confrontation network based oncontrastive embedding feature generation and Wasserstein generative adversarial network with gradient penalty. Through the cross stage contrastive embedding method, and according to the pre-defined fine-grained fault description, the feature generation of missing faults is completed，and the actual and generated fault features are mapped into the embedding space; Finally, the classification of all fault types including known and unknown is completed in the embedded space. The proposed method is applied to three typical zero-shot fault diagnosis scenarios. The experimental results show that compared with other methods, the proposed method can diagnose the known and unknown fault types, has advantages in accuracy.

Key words: rolling bearing, fault diagnosis, generalized zero-shot learning, feature generation, contrastive embedding