Abstract: In engineering practice, wind turbine fault diagnosis faces situations where the training fault category is different from the actual fault category. To achieve the diagnosis of unknown wind turbine faults, it is necessary to transfer the fault feature information learned during the training process to the unknown fault. Unlike traditional methods that directly establish the mapping relationship between fault samples and fault categories, a zero sample fault diagnosis method for wind turbines based on fault attributes is proposed to complete fault feature migration. The fault attribute matrix is established by describing the fault attributes of each fault, and it is embedded into the fault Sample space and fault category space; The fault attribute learner is established based on Convolutional neural network, and the fault classifier is established based on Euclidean distance, forming the diagnosis process of predicting fault attributes from fault samples and then classifying faults. Finally, the effectiveness and superiority of the proposed fault diagnosis method were verified through comparison with other zero-shot learning methods.       

Key words: fan fault diagnosis, zero-shot learning (ZSL), convolutional neural network (CNN), knowledge-data hybrid driven