With the rapid development of information technology,mobile communication has made breakthrough in the past few decades.Currently,the widespread deployment of fifth-generation(5G)mobile communication technology has laid the groundwork for various emerging applications by focusing on enhanced mobile broadband,ultra-reliable low-latency communication,and massive machine-type communication.However,to facilitate the crucial transition towards digitalization,networking,and intelligented in industries,and further improve the efficiency of communication networks,Beyond 5G(B5G)has attracted increasing research interests.In B5 G networks,wireless networks,artificial intelligence,and edge computing can be integrated to build an edge intelligence system,which aims to efficiently collect,cache,distribute,process,and analyze data at the edge of network and achieve a low-latency intelligent response.However,it still faces challenges such as the limited communication and computational resources,heterogeneity in edge network.Therefore,this thesis focuses on the key technologies of B5 G edge intelligence including edge caching,edge model training,and edge inference,and explores the system design,optimization,and performance evaluation of edge intelligence for B5 G networks.Specifically,the main content and contributions of this thesis are listed as follows.Firstly,this thesis focuses on the challenge of large-scale data transmission for B5 G edge intelligence,investigateing the edge caching problem in a multi-tier network architecture.In this complex scenario,files and data exist in various caching states and transmission modes,posing significant challenges for system performance evaluation and optimization.To solve this problem,the thesis combines the randomness of edge user locations and the caching states of requested files to determine relay selection protocols and corresponding file transmission modes,deriving a closed-form solution for the system outage probability to accurately evaluate the system performance in this scenario.Furthermore,the thesis derives an asymptotic expression for the system outage probability and,based on the asymptotic results,provides insights for the system design of this network.From the perspective of system outage probability,the thesis optimizes the cooperative caching probability of relay nodes and base stations.By decomposing this non-convex problem and using an improved block coordinate descent method and KarushKuhn-Tucker(KKT)conditions,an efficient caching strategy is obtained to achieve efficient data storage and fast retrieval at edge nodes,reducing wireless transmission outage probability and improving user experience.Secondly,to address the challenge of edge model training in resource-constrained and heterogeneous B5 G edge networks,this thesis investigates a personalized federated edge learning framework based on multiple exits.This framework allows resource-constrained devices to choose earlier exits,thereby training and uploading partial model structures,which can reduce training and communication latency.To improve the training performance under this mechanism,this thesis employs knowledge distillation techniques,using the aggregation of outputs from all exits of the global model to guide local model training.Moreover,the thesis jointly optimizes bandwidth allocation and exits selection for edge users,and propose a scheduling strategy based on a greedy algorithm to enable resource-constrained devices to participate more efficiently in model training,further enhancing the performance of edge model training.Lastly,to address the contradiction between the low-latency requirements of intelligent applications and limited communication and computing resources,this thesis designs a semantic communication-based edge inference framework.This framework encodes and transmits original signals from a semantic perspective,proposing a contrastive learning-based semantic encoding method and an efficient two-stage training framework,achieving adaptive balancing of transmitted pixels and semantic information according to bandwidth compression ratios.Under bandwidth constraints,the framework prioritizes sending semantic information,ensuring the accuracy of edge inference.With abundant bandwidth resources,the framework can also simultaneously send semantic information and global pixel information,achieving a certain signal transmission reconstruction quality and efficiently balancing edge inference accuracy and inference latency.This approach effectively satisfies the low-latency requirements of intelligent applications.In summary,the main contributions and innovations of this thesis lie in proposing a series of new methods and frameworks for edge caching in a multi-tier network architecture,resourceconstrained and heterogeneous edge model training,and addressing the contradiction between the low-latency requirements of intelligent applications and limited communication and computing resources.These contributions provide practical guidance for the development of future mobile communication networks. |