Fault diagnosis of planetary roller screw mechanism with a lightweight model based on federated learning

The fault data for Planetary Roller Screw Mechanisms (PRSM) is challenging to collect in real industrial settings due to the complex nature of practical operations and the lengthy accumulation period. Consequently, there has been little research on PRSM fault diagnosis. Additionally, the high processing cost of PRSM means that institutions are reluctant to make their fault data publicly available, creating a data barrier and further hindering research of the study on fault diagnosis of PRSM. To address these issues, Federated Learning (FL) is applied for PRSM fault diagnosis. In the FL framework, data remains in local storage, preserving data privacy. To reduce transmission costs, a lightweight model called SResNet18 is proposed. SResNet18 reduces parameters by 95.07 % and 61.93 % compared to ResNet18 and DSResNet18, respectively, which decreases the time needed for parameter uploading, model aggregation, and parameter returning. Additionally, SResNet18 has lower computational complexity, with 92.09 % and 36.66 % fewer FLOPs than ResNet18 and DSResNet18, respectively. Healthy and fault data of PRSM are collected on the PRSM testing rig, and the proposed method is evaluated. Results show that our method achieves the highest accuracy of 99.17 %, improving model performance while maintaining data privacy. The proposed SResNet18 also alleviates overfitting and reduces training time in the FL framework.