Federation Aggregation Strategy Learning Method for Inter-client Data Imbalance

The problem of insufficient accuracy of traditional deep learning models relying on massive samples for fault diagnosis can be solved by achieving joint optimization of multiple clients through federated learning when a few samples of a single client are used. However, the traditional federated learning method may lead to negative transfer when there is an imbalance in the sample number for inter-clients. A novel federation aggregation strategy learning method is proposed to handle inter-client data imbalance. The method constructs a federation loss to automatically optimize the aggregation strategy by driving the joint update of model parameters and aggregation weights for each client. It effectively solves the negative transfer problem in federated learning, which arises when an overfitted model is uploaded to the federated center due to insufficient sample size within individual clients. The experimental results show that the proposed method improves the fault diagnosis accuracy of clients with few samples and imbalanced sample sizes for inter-clients while avoiding the negative transfer problem in the federal process for clients with larger sample sizes.