Bearing fault diagnosis based on improved federated learning algorithm

• Published in: Computing Vol. 104; no. 1; pp. 1 - 19
• Main Authors: Geng, DaoQu, He, HanWen, Lan, XingChuan, Liu, Chang
• Format: Journal Article
• Language: English
• Published: Vienna Springer Vienna 01.01.2022; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Computer Appl. in Administrative Data Processing; Computer Communication Networks; Computer Science; Fault detection; Fault diagnosis; Federated learning; Information Systems Applications (incl.Internet); Machine learning; Nodes; Regular Paper; Software Engineering; Bearing fault diagnosis; Federated averaging (FedAvg); Federated learning; Convolutional neural network (CNN); 68T05
• ISSN: 0010-485X; 1436-5057
• DOI: 10.1007/s00607-021-01019-4

Bearing fault diagnosis can be used to accurately and automatically identify the type and severity of faults. Federation learning can perform learning without transferring local data among multiple local nodes with the same data features. The traditional federated learning algorithm is difficult to identify high-quality local models among class-unbalanced local nodes, which leads to the poor quality of the training model and the slow training speed. To improve the quality and training speed of the model, this paper proposes the FA-FedAvg algorithm for fault diagnosis based on the traditional federated learning algorithm. Specifically, the weighting strategy of the model is optimized, which is conducive to increasing the weight of high-quality local models, thereby improving the quality of training models. Then, a model aggregation strategy based on precision difference is proposed to reduce the number of iterations and accelerate the convergence of the training model. Finally, the proposed algorithm is compared with FedAvg and FedProx algorithms under different data distribution conditions. The experimental results show that, compared with the comparison algorithm, the number of model training iterations of the FA-FedAvg algorithm is reduced by 52.5% on average, and the fault classification accuracy has an average increment of about 8.6%. Moreover, the FA-FedAvg fault diagnosis method is robust under different classes and data volumes.