A personalized federated learning-based fault diagnosis method for data suffering from network attacks

Federated learning (FL) is an effective way to incorporate information provided by different clients when a single local client is unable to provide sufficient training samples for establishing a satisfactory deep learning model to diagnose rolling bearing faults, which plays an important role in ensuring the safe operations of motors. However, it is difficult to guarantee the effectiveness of FL when clients operating in different working conditions suffer from network attacks. This paper aims to study a new personalized FL (PFL) mechanism to secure each client's maximum benefit from the federation process such that the negative effects of condition variations or network attacks can be effectively prevented. By designing the inconsistency between the local model and the inherited global model, the information screening process in PFL is guided to ensure that each local client receives the maximum benefit. The model inconsistency derived from a certain round of federation is characterized by the output of an attention mechanism. Since personalized client information is emphasized, the proposed method can build reliable FL fault diagnosis models from unreliable samples in cases with attacked client-side sample data. The effectiveness of the proposed method is validated by using the benchmark dataset provided by the Rolling Bearing Center of Case Western Reserve University. In the case when a certain client suffers from a strong network attack, the proposed method can achieve a fault diagnosis accuracy improvement of 27.11% over the existing FL fault diagnosis methods.