Federated transfer fault diagnosis of electromechanical transmission system under Non-IID conditions

Abstract As a core component of industrial equipment, the safe and stable operation of an electromechanical transmission system is of great significance to equipment reliability. In federated learning, fault data often exhibit non-independent and identically distributed (Non-IID) characteristics arising from the combined effects of long-tailed distributions and label distribution skew, leading to difficulties in tail-class recognition and cross-client generalization. This paper proposes a federated transfer fault diagnosis method based on prototype anchor construction, a dual-branch classifier, and a class-structure regularization function. First, prototype anchors integrating orthogonal direction prototypes, aggregated feature prototypes, and distribution-statistics-based pseudo-features are constructed to explicitly share class-structure information. Second, a dual-branch classifier composed of an angle classifier and a feature-prototype classifier is constructed to enhance the model’s ability to recognize missing classes and tail classes. Finally, a class-structure regularization function is introduced to optimize the global feature-space structure, improve the overall diagnostic performance on a class-balanced test set under long-tailed training conditions, and alleviate recognition confusion among classes with limited samples. The proposed method is validated on three types of federated experimental tasks. The experimental results show that the proposed method achieves favorable results in terms of accuracy and recall, demonstrates good overall fault-recognition capability and class-recognition completeness, and effectively improves tail-class recognition and cross-client generalization under complex Non-IID conditions.