Abstract: Big data-driven intelligent machinery fault diagnosis methods have achieved great success in the recent years. The high diagnosis accuracies mostly rely on large amounts of labeled condition monitoring data and centralized model training. However, in the real industries, it is usually difficult for a single user to collect sufficient labeled data, that makes the intelligent diagnosis methods less applicable in practice. It is noted that different industrial users may have similar machines and condition monitoring data. Therefore, collaborative model development is promising to address the data scarcity problem. However, data privacy is very important and different users are generally not comfortable sharing private data with others, that results in a challenging collaborative diagnosis problem. A privacy-preserving collaborative intelligent machine fault diagnosis method FedTL is proposed. The private data are used for training without leaving local storage. The high-level representations of shared data are communicated among different users. A soft label-based information transmission method is proposed. Through capturing the relationship between different fault modes of shared data, the diagnosis knowledge of private data can be well delivered. The federated transfer learning framework is formulated, considering different working conditions of different users. The experiments in bearing condition monitoring cases validate the proposed method. The results show the proposed method is a promising tool for privacy-preserving collaborative machine fault diagnosis.

Key words: machinery, fault diagnosis, data privacy, federated learning, transfer learning