Multi-dimensional Resource Allocation And Optimization For Federated Learning In Intelligent Industrial Internet Of Things

With the rapid development of the 6G network and the ever-increasing scale of the Industrial Internet of Things,the requirements of massive data analysis as well as its privacy and security issues in the modern intelligent industrial scenario have simultaneously increased,which brought serious challenges to the traditional central machine learning framework.As a distributed machine learning framework,federated learning is expected to meet the requirements of real-time data analysis and decisionmaking as well as privacy protection in the Industrial Internet of Things scenario due to its characteristics of high communication efficiency and enhanced privacy protection level.However,due to the complex industrial scenarios and the limited wireless resources,as well as the heterogeneous industrial devices on computation capacity,channel condition,and energy resource,the multi-dimensional resource management and allocation schemes may cause a significant effect on the system performance in terms of federated learning model accuracy.In this context,this thesis focuses on the management and optimization scheme of multi-dimensional resources for federated learning in intelligent Industrial Internet of Things scenarios.The main research works of this thesis can be summarized as follows:(1)A joint optimization framework of multi-dimensional resources for achieving the tradeoff between federated learning performance and learning cost is designed.Based on the complexity of the large-scale Industrial Internet of Things scenario,this thesis proposes a hierarchical collaborative federated learning architecture,which allows the Industrial Internet of Things devices,edge servers,and cloud server to cooperatively update the global federated learning model.In order to mitigate the adverse effect caused by the heterogeneity of Industrial Internet of Things devices and the uncertainty of wireless channels on federated learning model accuracy and learning cost,a weighted learning utility maximization optimization problem is formulated.The problem aims to balance the tradeoff between the federated learning performance and learning cost in the training process by jointly optimizing the edge association,wireless resource block allocation,computation capacity,and transmit power of IIo T device.Due to the NP-hardness of the formulated mixed-integer non-convex optimization problem,this thesis decomposes the original problem into three subproblems,which are solved alternately and iteratively by the methods,such as hypergraph matching and successive convex approximation,to obtain the suboptimal solution of the original problem.Simulation results demonstrate that the proposed joint optimization framework can synthetically improve the weighted learning utility by about 10%-17% compared with benchmark schemes,and effectively reduce the learning cost while guaranteeing the federated learning performance.(2)A Q-learning-based optimization strategy for data offloading in edge-assisted federated learning framework is proposed.Focusing on the “straggler effect” caused by the Industrial Internet of Things devices with limited local computation capacities,this thesis proposes an edge-assisted federated learning framework to improve the training efficiency of federated learning in the Industrial Internet of Things scenario by combining it with multi-access edge computing.In order to obtain the optimal offloading data size of the system and thus improve the learning performance,a federated learning loss function minimization optimization problem is formulated with the latency constraint.Based on the derived optimal offloading strategy in the case of perfect channel state information,this thesis further proposes a Q-learning-based edgeassisted offloading strategy to obtain an efficient solution to the optimization problem in the case of imperfect channel state information.Simulation results show that the proposed edge-assisted federated learning framework and the corresponding edgeassisted offloading strategy can effectively improve the model accuracy of federated learning by about 4.7%,when compared with the benchmark schemes.