A label description space embedded model for zero-shot intelligent diagnosis of mechanical compound faults

It has always been an issue of significance to diagnose compound faults of machines. Existing intelligent diagnosis methods have to be trained by sufficient data of each compound fault. However, both labeled and unlabeled data of mechanical compound faults are usually difficult to collect or even completely inaccessible for training in real scenarios. Therefore, compound faults are usually unseen fault patterns. Unseen fault patterns are those that have no labeled or unlabeled training data. Without training data of compound faults, the current intelligent diagnosis methods usually fail in recognizing compound faults. This paper proposes a zero-shot intelligent diagnosis method for unseen compound faults of machines. The proposed method contains three stages, i.e., the feature learning, pre-judgment and fault recognition. The key to this method is a label description space embedded model for intelligent fault diagnosis (LDS-IFD) in Stage 3. In LDS-IFD, a label description space (LDS) is built to construct the relationship among different fault patterns. LDS is embedded between the feature space (FS) and the health condition label space (HCLS). Then the projection between FS and LDS is constructed by a linear supervised autoencoder (LSAE). By similarity evaluation in LDS or FS, LDS-IFD is able to recognize mechanical compound faults when only the data of single faults are accessible for training. The proposed method is demonstrated on a bearing dataset and a planetary gearbox dataset. Results show that the proposed method is effective in diagnosing unseen compound faults of machines.