A zero-shot attribute-embedded model with a feature difference mapping sigmoid function for compound fault diagnosis of rotating machinery

The detection of machinery compound faults has always been a great challenge. Most of the current compound fault diagnosis methods require a large number of compound fault data to participate in training. However, in actual engineering, it is impractical to collect abundant fault samples, especially compound fault samples. To address the issue of lacking the compound fault data for training, this paper proposes a zero-shot attribute-embedded model for compound fault diagnosis (ZSAECFD). This model only uses the data of various single faults for training, but the trained model is able to diagnose the unseen compound faults. Using the data of single faults, the attribute prototypes for single and compound faults of bearings and gearbox are first constructed. By calculating the Euclidean distances between attributes and attribute prototypes, the compound fault types can be distinguished. Moreover, considering that the traditional sigmoid has the limited ability to map the difference of features in multi-label classification tasks, we propose a new activation function, feature difference mapping sigmoid (F-sigmoid). It can effectively amplify the differences between features, which is helpful for improving the accuracy of attribute recognition. It is also proven that F-sigmoid can effectively alleviate the problem of gradient vanishing compared to sigmoid. The performance of the proposed ZSAECFD is validated through the compound fault diagnosis experiments on bearings and gearboxes. Without using the compound fault data for training, the diagnostic accuracy of bearing faults reaches 81.82 %, and the diagnostic accuracy of gear faults is up to 88.17 %. The experimental results show that the proposed model can effectively diagnose the unseen compound faults, and has advantages over the classical and advanced zero-shot learning methods.