无线电工程

基于深度学习的大型智能表面毫米波波束赋形问题，引入具有少量有源元件的大型智能表面(Large Intelligence Surface, LIS)后，通过一系列实验对比研究了卷积神经网络(Convolutional Neural Network, CNN)算法下波束赋形的性能。通过使用DeepMIMO数据集中具体场景下的射线追踪信道数据构造通信场景，并采用神经网络模型CNN、反向传播(Back Propagation, BP)和多层感知器(Multi-Layer Perceptron, MLP)算法学习毫米波通信环境。实验考虑了多个参数条件，包括路径数、有源单元数量、发射功率和数据集大小。结果表明，CNN算法在所有参数条件下均优于另外2种算法，在增加路径数和有源单元数量的情况下，CNN算法的性能优势更为显著。此外，增加数据集大小也可以提高CNN算法的性能表现。实验结果为相关领域的研究和实际应用提供了有价值的参考，对于改进毫米波通信系统性能具有重要意义。

To deal with the problem of deep learning-based Large Intelligent Surface(LIS) millimeter-wave beamforming, a LIS with a limited number of active components is incorporated, then the performance of beamforming under the Convolutional Neural Network(CNN) algorithm is scrutinized through a series of experimental comparisons. The ray-tracing channel data from the DeepMIMO dataset are used to construct communication scenarios in specific environments, and the neural network models, CNN, Back Propagation(BP), and Multi-Layer Perceptron(MLP) algorithms are employed to learn the millimeter-wave communication environment. The experiments take various parameter conditions into consideration, including the number of paths, the quantity of active elements, transmit power and dataset size. The results indicate that the CNN algorithm outperforms the other two algorithms under all parameter conditions. Furthermore, as the number of paths and active elements increases, the performance advantage of the CNN algorithm becomes even more pronounced. Additionally, increasing the dataset size also contributes to enhancing the performance of the CNN algorithm. The experimental findings provide valuable references for the research and practical applications in related fields, and hold significant importance in enhancing the performance of millimeter-wave communication systems.