基于联邦学习和可变形卷积网络的风机集群多故障诊断方法OA

To solve the problem of"data islands"generated by different turbines in a wind turbine cluster,which affects the accuracy of multi-fault diagnosis,the FedLVA-WTDCN method is proposed,which achieves cross-equipment and cross-region data co-training,and at the same time guarantees the generalisation ability of the model,which improves the comprehensiveness and reliability of fault diagnosis.In this paper,the aggregation algorithm of federated learning is improved,and an aggregation method based on the joint weighting of loss inverse and variance(FedLVA)is proposed to replace the traditional FedAvg algorithm.In addition,combining with the background of wind turbine fault diagnosis,the latest deformable convolutional network(DCN)is used to the federated learning local model,and WTDCN(wind turbine deformable convolutional network)is proposed,which makes full use of its self-adaptive feature extraction capability to further improve the diagnostic accuracy.Comparison between centralised training and federated learning through open-source wind turbine datasets demonstrates the effectiveness of federated learning in solving the wind turbine cluster fault diagnosis problem;comparison through different aggregation algorithms proves the stability and reliability of the proposed FedLVA algorithm;and the use of different local learning models illustrates that the proposed WTDCN method can be better applied to wind turbine cluster data.