Research On Fault Diagnosis Method Of Rolling Bearings Based On Federated Model Transfer Learning

As one of the key components of rotating machinery,rolling bearings are widely used in industrial production,railway locomotives,aerospace and other fields.Timely and effective fault diagnosis can avoid heavy accident.However,the lack of complete bearing data sets often leads to unsatisfactory diagnostic results for single user.At the same time,as data privacy and security are widely concerned by the society,the rolling bearing vibration data of each user’s privacy is isolated and not shared.Therefore,it is of great practical significance to break the data barrier without destroying the data privacy and build an effective diagnostic model together with multiple parties.In view of the above problems,two kinds of fault diagnosis methods for rolling bearings are proposed.(1)A federated learning framework comes up and applied to bearing fault diagnosis.This method improves the federated learning parameter passing process to enhance the security of the passing process.For the vibration data of multiple users,short-time Fourier transform is used to construct the time-frequency graph data sets.Users train local models and upload the parameters to the server.In the meantime,the update of difference value and parameter sparsity method are used to improve the local model parameter transitive strategy in federated learning.The server uses the federated average algorithm to aggregate model parameters and updates the local model.After the iterations,the global fault diagnosis model is established.Experimental results show that the proposed method can improve the diagnostic accuracy without exposing the local data sets.Above method can be used to diagnose faults without data leaving the local area.However,when there are data of different specifications of rolling bearings among federated users,this method has some limitations.Therefore,another fault diagnosis method of rolling bearing is proposed.(2)A federation model transfer learning framework comes up and applied to bearing fault diagnosis.This method is used to solve the problem that the data distribution difference between different users leads to poor model performance.Federated learning and model transfer learning are regarded as the core,using the proposed layer-by-layer thawing strategy,some parameters of the global diagnostic model trained by federated learning are reserved and sent to each user.Users utilize local data to fine-tune the global diagnostic model and obtain a personalized model for each user.Through experimental verification,this method can realize the fault diagnosis of rolling bearings of different specifications under the premise of data island and label scarcity.The method has high accuracy.