移动场景下的智能信道预测方法

针对无线通信信道估计老化问题，本文提出了一种基于卷积神经网络的信道预测方法，该方法通过联合轨迹预测和信道重构实现。首先，采用卷积神经网络学习从规划路线和移动终端所在位置到移动方向映射，进而预测出轨迹上多个目标位置；其次，采用卷积神经网络学习从目标位置附近K个位置项的信道，到目标位置信道间映射，用于实现预测轨迹的信道估计。本文利用Wireless InSite为移动方向预测和信道重构模型的训练及测试生成充足的样本，包括规划路线和通过射线跟踪方法获取的信道等。仿真结果表明，本文所提出的方法能有效地估计目标位置的信道特性，与K值较小的K-近邻插值方法和基于全连接神经网络的信道预测方法相比，其信道估计总相对误差更低且鲁棒性较好。 

An Intelligent Channel Prediction Method in Mobile Scenarios

? ?Aiming at the channel aging problem in wireless communication channel estimation， a convolutional neural network （CNN）-based channel prediction method was proposed. It was implemented by joint trajectory prediction and channel reconstruction. The trajectory prediction was implemented by a CNN-based mobile direction prediction model， and the channel reconstruction was implemented by a CNN-based channel reconstruction model. Specifically， first， the prediction problem of the two elements of the direction vector， i.e.， the unit vector in the moving direction， was regard as a multi-task learning problem. Then， we adopted two CNNs to learn the mapping from the planned route and the current location of the mobile terminal， to the two elements of the moving direction， respectively. Next， based on the predicted direction vectors， the predicted position was calculated using kinematic principle. Then， the movement trajectory was predicted. Second， a CNN was adopted to learn the mapping from the channel characteristic matrices of the K closest position terms in the channel path map， to the channel characteristic matrix of target location. The channel estimation of the predicted trajectory was achieved， by predicting the channel characteristic matrix of each target location on the predicted trajectory. We utilized Wireless InSite software to generate sufficient samples for the training and testing of the mobile direction prediction and channel reconstruction models， and to construct the channel path maps. The collected data set included the planned routes， current locations and moving trajectories of the mobile terminal， as well as the multi-input multi-output （MIMO） channel characteristics （e.g.， the powers， arrival angles in vertical and horizontal directions of effective paths， etc.）， which was generated by the ray-tracing algorithm of Wireless InSite. Numerical simulation results showed that， the proposed channel reconstruction model could effectively estimate the MIMO channel characteristics， outperformed the fully connected neural network （FCN）-based channel reconstruction model， and was very close to the error performance of the K-nearest neighbor interpolation method， in terms of error. In addition， the proposed CNN-based channel prediction method had lower sum of relative error and better robustness in channel estimation than the K-nearest-neighbor interpolation method with smaller K values， and the FCN-based channel prediction method.