Research On Federated Learning Methods For Spectral Feature Recognitio

Automatic spectrum characteristic recognition(ASCR)aims to automatically extract features from spectrum big data and identify/classify them.This supports various tasks related to spectrum sensing,monitoring,and management,and enables efficient utilization of electromagnetic spectrum resources.Deep neural network-based ASCR methods have shown good recognition results,but this method requires users to send data to a central server,which raises privacy concerns and transmission overheads.In this scenario,Federated Learning(FL),as a new distributed computing architecture,has become a feasible way to reduce transmission overhead,protect users’ privacy and solve the dilemma of data sharing.However,the complexity of mobile edge computing network transmission environment and the limited resources of mobile devices pose two challenges to the implementation of ASCR based on FL architecture.On the one hand,in mobile edge computing networks,mobile devices and edge computing nodes are distributed widely.Mobile devices have limited resources and need to reduce training delay and energy consumption while ensuring the accuracy of multiple FL tasks.Determining the relationship between FL tasks of mobile devices and edge computing nodes is a crucial challenge for improving FL efficiency.On the other hand,in wireless networks,link transmission is often unstable due to the randomness of the transmission environment and the time-varying electromagnetic environment.This instability causes jitter and interruptions in data transmission speed,which significantly affects the training efficiency of FL.Improving the training efficiency of FL in such unstable transmission scenarios remains a significant challenge.Under this background,this dissertation aims to design mobile device scheduling and resource allocation algorithms for ASCR-oriented FL.It focuses on improving training efficiency,reducing communication overhead,and building an efficient and reliable multi-task FL architecture for mobile edge computing networks.The main work of this dissertation includes four aspects:(1)This study summarizes and compares the automatic modulation recognition based on deep learning,constructs a verification environment,implements and verifies multiple mainstream recognition methods.Taking automatic modulation recognition as an example,the current research status of feature recognition is analyzed and the deep learning-based modulation recognition algorithm is divided into 4 categories.The typical solutions for each category are summarized and introduced in detail.Finally,10 typical recognition algorithms are implemented and systematically compared and analyzed on a public dataset.(2)A differential evolution-based mobile device scheduling algorithm is proposed for the multi-task federated learning scenario to reduce the overall system latency and energy consumption.An optimization model is established to minimize the training latency and energy consumption per round of federated learning tasks in the complex mobile edge computing network environment.The proposed algorithm considers factors such as model accuracy,transmission link quality,task latency,and energy consumption.It treats each feasible mobile device scheduling scheme as an individual,with the optimization objective as the fitness value,and obtains the suboptimal solution of the problem through mutation,crossover,and selection operations.The simulation results indicate that when the number of mobile devices increased to 800,compared with the random scheduling method,the proposed algorithm reduced the overall cost by 28%,effectively reducing the time delay and energy consumption during the training process.(3)Proposed an Actor-Critic(AC)based multi-task federated resource allocation algorithm to improve the overall training and communication efficiency in a multi-task federated environment.In a mobile edge computing network,the resources and task data stored on mobile devices differ,and there are also differences in the channel gain between each edge computing node,which leads to different delays and energy consumption when different devices participate in federated learning tasks.In this dissertation,a multi-task federated resource allocation algorithm based on AC is proposed,with the optimization objective being the weighted sum of delay and energy consumption.The algorithm transforms FL model training into a Markov process,improves the accuracy of allocation strategy through network training,and ultimately obtains the best resource allocation strategy.Simulation results show that in large-scale scenarios,after about 2000 iterations,the algorithm’s cost value converges to 1.76,which reduces the overall cost by 4% compared to the optimal benchmark algorithm.(4)A multi-task federated prototype system was designed and validated using typical datasets and models for federated learning architecture.The system consists of five modules:interaction,central management,device management,data integration,and model integration.The interaction module provides a visual interface for monitoring task status and results.The central and device management modules are responsible for parameter instruction transmission and execution of the federated learning architecture.The data and model integration modules include four types of neural network models and four types of federated task datasets represented by spectral feature recognition.Functional validation experiments have demonstrated the effectiveness of the system.