Research On Federated Learning-Based Resource Allocation Methods In Industrial Internet Of Things

With the popularization of the Industrial Internet of Things(IIo T),massive industrial data has been generated,which urgently promotes the application of artificial intelligence(AI)to fully release the potential of industrial data to enhance the capabilities of industrial service.Model training in traditional machine learning(ML)methods requires the device to upload its original data,which will cause the privacy leakage.Therefore,in IIOT,in order to take full advantage of data from multiple industrial devices to train models while protecting privacy,federated learning(FL)has become an efficient solution and received extensive attention from both industry and academia.However,the limited communication and computation resources of IIOT networks,the limited computing power and the heterogeneity in data resources of industrial devices,as well as the strict real-time requirements of massive industrial data processing,have brought great challenges to the selection of training devices and the allocation of resources in IIOT.Therefore,how to design efficient device selection and resource allocation methods has become a key and scientific problem,which requires to be solved urgently in the research of FL-based IIOT.In this context,this thesis focuses on the resource allocation methods of FL-based IIOT,and proposes a cost-efficient device selection and bandwidth allocation method for federated edge learning(FEEL)-based IIOT and a device association and resource allocation method in hierarchical federated learning(HFL)-based mobile edge computing(MEC)-enabled IIOT,respectively.The main research contents of this thesis are as follows:(1)Due to the heterogeneity of IIOT devices,the limited channel bandwidth resource,and limited energy consumption of devices in the long-term FL process,a cost-efficient device selection and bandwidth allocation method for FL-based IIOT is proposed to balance the quality of model training and the system resource cost.First,the impact of device selection and bandwidth allocation on the actual cost in FL process is comprehensively analyzed,and the total system cost function is modeled as the difference between the total computation and communication cost and the benefit brought by the number of the total training data.Secondly,under the delay and longterm energy constraints of devices,the long-term time-averaged uniform cost minimization problem is constructed.Finally,considering that it is difficult to solve a long-term problem directly,a Lyapunov optimization-based device selection and bandwidth allocation method is proposed,which uses Lyapunov optimization theory to transform the long-term cost minimization problem into a series of short-term drift-plus-cost minimization problems,and then an iterative algorithm is designed to select device and allocate bandwidth properly in each round.The simulation results show that,the proposed method can achieve better performance in average cost minimization than other methods.In addition,under different data sets and data distributions,the proposed method is superior to other methods in FL performance in terms of accuracy and convergence speed.(2)Because the number of devices that can be accessed to each server in the FEEL framework is limited,the resource allocation method in HFL-based IIOT is further studied to improve learning efficiency in this thesis.For the insufficient computing power of devices and limited computing and communication resources of edge servers in IIOT,a device association and resource allocation method in HFL-based MEC-enabled IIOT is proposed.First,with the help of MEC technology,the impacts of device association,computation offloading and transmission power on learning efficiency are comprehensively analyzed,and the device computation offloading process and HFL process are systematically modeled.The system delay is modeled as the maximum value of the sum of offloading,computation and communication delays among all devices.Secondly,under the constraints of offloading ratio,transmission power and the device association set,the system delay minimization problem is formulated.Finally,to solve the target problem,the divide-and-conquer principle is adopted,which decomposes it into the resource allocation problem under the given device association of each server and the device association problem of all edge servers.By using the convex optimization method,a device association and resource allocation method is designed,so as to obtain the optimal device association and resource allocation strategy.The simulation results show that the proposed method can get low complexity,and achieve better performance in reducing system delay than other methods.