An Active Federated Method Driven by Inter-Client Informativeness Variability of Labeled Data

Federated learning does well in jointly training multiple deep learning models for fault diagnosis. The accuracy of local model may affect the effectiveness of federated learning. In the case when the local model is suffered from few labeled data with poor quality, it is significant to design an active federation strategy to screen informative unlabeled samples by which supervised model can be further optimized. However, traditional active learning method for federated learning is unable to guarantee the effectiveness of federation since it focuses on screening the samples helpful for a specific local model rather than global model. This paper proposes a method of combining the variability of informativeness within the client and the variability between clients to overcome this problem. Informativeness variability of labeled data between clients is used to guide the screen strategy using active federation model rather than local model, such that influence of unreliable clients can be decreased to avoid negative transferring due to samples with low quality. Experimental validation for benchmark dataset of rolling bearing shows that 25.04% improvement of fault diagnosis accuracy can be achieved in the case when only few labeled data are available and 13.6% improvement can be achieved in the case when only labeled data with low quality are available.