Resource Management And Optimization Based On Deep Learning In Ultra-dense Network

Ultra-Dense Network(UDN)has become a key technology in the new generation of mobile communication systems.By densely and flexibly deploying low-power small base stations to reduce the distance between access nodes and user equipment nodes,the spectrum energy efficiency of the network can be effectively improved.But at the same time,the densification of wireless networks also brings new challenges to issues such as user associ-ation and power allocation in the network.The reliability and practicability of traditional mathematical modeling algorithms to solve these problems depend on accurate modeling of the environment.However,in actual situations,the heterogeneity and complexity of node deployment in UDN make modeling extremely difficult.Besides,most traditional algorithms tend to focus only on the base station side and ignore a lot of information on the user side.Therefore,from the perspective of modeling accuracy and data utilization,traditional algorithms have certain de-fects.By using the function approximation and data processing capabilities of deep learning,this thesis can fully combine the information on the base station side and the user side to solve the user association and power allocation problems in UDN.The algorithms proposed in the article do not require an accurate channel model,instead,they extract the channel gain information obtained from the relevant channel model.The main work of this thesis is as follows:1.Given a typical two-layer UDN system scenario,using a two-stage path loss model,and modeling the small-scale fading between the base station and the user equipment as independent and identically distributed Rayleigh fading,and modeling the shadow fading as A zero-mean Gaussian random variable with standard deviation,thereby obtaining the downlink channel gain between the base station and the user equipment.The mathematical model of the UDN system is established,and the simulation experiment system is realized by writing code.After that,the training and verification of the algorithm proposed in the article are carried out in the simulation experiment system.2.A downlink transmission scenario in UDN is considered,in which base stations and user equipment are randomly distributed in urban areas,and each user is assumed to be served by only one base station.A specific UDN system model is defined,that is,a network scenario and related parameters are designed based on the basic UDN scenario structure,and an optimization objective function that maximizes the total rate of coverage of all base stations under certain constraints is given.For this UDN system model,three cluster-based UDN user association algorithms are proposed.They are user association algorithm based on K-Means clustering,user association algorithm based on a modified version of spectral clustering,and user association algorithm based on deep clustering.The three algorithms use different methods to process information gain and realize the association with corresponding base stations by generating different user clusters.The three algorithms are validated and analyzed in the simulation experiment environment built in this thesis and compared with traditional user association algorithms.The simulation experiment results show that the performance of the three cluster-based UDN user association algorithms under different numbers of small base stations and different maximum transmit powers of small base stations is better than traditional user association algorithms.3.A downlink transmission scenario in UDN is considered,it is assumed that both the base station and the user equipment are configured with a single antenna,and each user selects the best access base station for its service based on the received signal strength,thus forming several base station-user pairs.A specific UDN system model is defined,and an optimization objective function that maximizes the sum rate of users in the network under the premise of total power limitation is obtained.For this UDN system model,four UDN power allocation algorithms based on deep learning are proposed.They are an unsupervised learning power allocation algorithm that uses objective function values for training,and an unsupervised learning power allocation algorithm that uses multiple loss values for training.The semi-supervised learning power allocation algorithm for supervised pre-training and the semi-supervised learning power allocation algorithm for unsupervised pre-training.In the design of the algorithm,the experience of traditional algorithms and feedback from the UDN system are considered at the same time,combined with the methods of supervised learning and unsupervised learning,and the DNN is trained to give a reasonable power allocation plan.The four algorithms are verified and analyzed in the simulation experiment environment built in this thesis and compared with traditional power distribution algorithms.The simulation experiment results show that the four UDN power distribution algorithms based on deep learning perform well and are stable,and the total UDN system rate obtained after the algorithm results are derived is significantly higher than traditional power distribution algorithms.4.A two-layer UDN scenario in an urban environment is considered,the coverage of macro base stations and small base stations in the network is different,and each base station will affect each other.By generating a different transmit power vector for each base station,the coverage area of all base stations is maximized The total rate.To make better use of the channel information in the network,a graph model is constructed for the UDN scenario,in which the adjacency relationship between the base stations is determined by whether there is a piece of common user equipment in their coverage area.An unsupervised graph convolutional neural network is trained on the graph model to output the representation vector of each vertex in the graph,that is,the transmission power vector of each base station,to obtain the power allocation scheme.In the learning process of the graph convolutional neural network,not only the information of the base station node itself can be used,but also the information from its neighboring base stations can be aggregated.At the same time,the optimization problem is considered from the perspective of the base station and the user,reducing inter-cell interference and improving the overall system throughput.Experimental verification and result analysis of the proposed algorithm in the simulation experiment environment built in this thesis prove the convergence of the algorithm.