Design Of Collaborative Transmission Scheme In Millimeter Wave Networks Based On AI

With the rise of technologies such as the Internet of Things and big data,5.5G-oriented future communication needs to meet the requirements of ultra-high throughput,high data transmission speed,ultra-high bandwidth,ultra-low latency,and high reliability.Due to the large bandwidth advantage of the millimeter-wave frequency band,and massive multiple input multiple output(MIMO)beamforming can effectively compensate for millimeter-wave propagation fading by using the array gain,the millimeter-wave massive MIMO technology is considered as one of the key technologies to achieve higher performance metrics in 5.5G mobile communication systems.Considering that millimeter-wave massive MIMO needs to use narrow beams for data transmission,and is very sensitive to the shadowing effect in the channel,collaborative multipoint transmission becomes the main mode for millimeter-wave massive MIMO systems to achieve their performance gain.Currently,Artificial Intelligence(AI)techniques have been used in millimeter-wave massive MIMO systems,but there are still some issues that need to be further researched,such as the lightweight of model,small sample learning,and model adaptation to fast time-varying scenarios.In this thesis,we study the multipoint collaborative millimeter-wave massive MIMO system,the main contents are as follows:1、A user incremental collaborative transmission scheme based on broad learning in millimeterwave massive MIMO system is proposed.Firstly,to address the problem that the current lowoverhead beam training method based on wide beam response cannot be directly applied to scenarios without uplink and downlink channel reciprocity,the local lightweight broad learning network is trained using the downlink wide beam response and narrow beam response collected by each user in the offline phase,and the beam prediction in the online phase is based on the wide beam response of multiple base stations.Secondly,the number of effective training samples that can be collected by a single user is limited in fast time-varying scenarios,and the broad learning model trained based on small samples suffers from insufficient generalization capability.Based on the distributed learning theory,the training problem of neighboring users’ local networks is modeled as a distributed optimization problem with global consistency constraints,and realizes the effective sharing of training data among neighboring users.Finally,due to the fast time-varying scenario also requires high real-time model update,an incremental update method of user local network in collaborative mode is designed by using the characteristics of broad learning network,which can effectively reduce the computational complexity when updating the model.Simulation results show that the proposed scheme can achieve a higher effective system transmission rate than the conventional millimeterwave beam training scheme,and there is no significant performance loss compared with the centralized beam selection method.In a fast time-varying environment,incremental updates can achieve consistent results compared to full updates.2、A base station incremental collaborative transmission scheme in millimeter-wave massive MIMO system based on broad learning is proposed.Firstly,for a multipoint collaborative millimeterwave transmission scenario connected to the central processing unit via fronthaul links,based on the broad learning method,the user uplink omnidirectional wide beam response collected by multiple base stations in the offline phase are used to predict the optimal downlink beam selection results.Secondly,to alleviate the high requirements of the current centralized beam selection method at base station side based on wide beam response on the performance of a single AI engine of the central processing unit as well as the bandwidth of the fronthaul link,a millimeter-wave MIMO base station incremental collaborative beam selection method based on distributed broad learning architecture is proposed,which draws on the idea of vertical federated learning.The original centralized optimization problem is transformed into a distributed optimization problem by vertically cutting the data feature space to achieve an implicit and efficient collaborative training among base stations.At the same time,the communication overhead of the fronthaul link is further reduced by exploiting the sparsity of the intermediate parameters in the training process.Simulation results show that the proposed algorithm realizes the lightweight and significant reduction in update cost of the AI model at the base station side.It can relieve the computational pressure of the central processing unit while achieving the effect of approximating the centralized learning beam selection,and achieves a reasonable compromise between collaboration overhead and beam selection prediction performance.