Research On Mechanical Fault Diagnosis Method Based On Federated Learning

With the support of the National Natural Science Fund(No.52075236),the Project supported by the Natural Science Foundation of Jiangxi Province,China(Grant No.20212ACB202005),the Project supported by the equipment Pre-Research Foundation of China(Grant No.6142003190210)and other grants,due to the unique advantages of federated learning of effectively protecting the privacy and security of data distributed in different clients and fusing the local models into shared model,federated learning is introduced to mechanical fault diagnosis.A method of mechanical fault diagnosis based on federated learning is proposed and investigated thoroughly.Some innovative result obtained in this article are as follows.(1)Aiming at the problems of uneven distribution and difficult sharing of fault data in engineering,and the inability of the existing fault diagnosis methods to identify new added fault types,a mechanical fault diagnosis method based on federated learning is proposed in this paper.Firstly,the local fault diagnosis models are established respectively by the deep convolutional neural network for the existing fault data and newly added fault data.Then,the weight parameters of the local fault diagnosis models are fused through the federated average algorithm and updated to the global model parameters.Finally,the global model parameters are returned to each local fault diagnosis models to update into a global shared fault diagnosis model,so that the federated fault diagnosis model can identify the newly added fault data.The proposed method is validated with bearing fault data.Compared with the traditional fault model which can only diagnose the existing fault data but cannot identify the newly added fault,the federated fault diagnosis model fused with weight parameters of local models can accurately recognize the newly added fault data.(2)The fault diagnosis model based on federated learning ignores the problem that the distributions between newly added fault data and the existing fault data are different.The transfer learning can solve the data distribution variability well.However,the existing transfer learning only focuses on the effect of model on the target domain but ignores the effect of model on source domain.In order to overcome the shortcomings,combining the advantages of transfer learning and federated learning,a fault diagnosis method based on federated transfer learning is proposed.Firstly,the source domain and target domain fault diagnosis models are established base on deep convolutional neural network in the client.The source domain data with label are used to trained the source domain model,and the parameters of the target domain model are shared by source domain model.Then,in the backpropagation,the loss function with Maximum mean discrepancy is used to reduce the difference of source domain and target domain in order to make the feature spaces of both more similar.Finally,the parameters of each local fault diagnosis are fused using the federated learning to generate a global model,so that the model can not only recognize the fault types in target domain,but also recognize the fault types in source domain.To verify the effectiveness of the proposed method,the both bearings at different rotational speeds are transferred to diagnose the fault.Consider the influence of transmission paths of in the same transmission system,the bearing and planetary gear are respectively as the source domain and target domain for transfer fault diagnosis.Compared with traditional transfer fault diagnosis method,the result show that the federated transfer learning fault diagnosis model can reduce the difference of newly added fault type data,and accurately recognize the fault data in source domain and target domain.(3)The fault diagnosis method based on federated learning is not adaptive enough to the dimensionality of input data,and can only be constructed by the fixed dimensionality of input.Combining the advantage that the multi-head attention mechanism can adaptively construct the input to the sample dimensionality,a fault diagnosis model based on federated learning with attention mechanism are proposed.Firstly,the fault data are processed into information samples with different dimensions by wavelet packet decomposition and variational modal decomposition methods.Then,the fault diagnosis model with multi-head attention mechanism is built in the local,and the inputs of different dimensions are adaptively selected and fused by the multi-head attention mechanism.Finally,the parameters of the model are updated to the global model by fusion through the federation mechanism.The proposed method is applied to experiments on bearing and gearbox faults,and the results show that the attention mechanism federation learning model can effectively and adaptively fuse the multichannel information extracted from the original signals to accurately identify the fault types.(4)The fault diagnosis model based on federated learning ignores the deficiency that the newly added data may be a new task with sparse sample data,while the meta learning are proposed for the poor adaptability of traditional neural network model to new types of tasks and the insufficient generalization ability of existing fault diagnosis models.Combining the advantages that federated learning can recognize the newly added data and meta learning can handle small sample data,a fault diagnosis method based on federated meta learning is proposed.Firstly,model-agnostic meta-learning fault diagnosis models are established in the local client,and the learning algorithms of the models are trained and tuned by the supporting set and query set.Then,the learning algorithms of each the local meta learning fault model are fused by the federated mechanism.Finally,the fused learning algorithms are returned to the local model so that each model is updated to the global fault diagnosis model.The proposed method is applied to small sample fault experiments of bearing and planetary,and the result show that the federated meta learning fault diagnosis models can still recognize the different types of faults with few samples of fault data.