Wireless channel feature extraction via GMM and CNN in the tomographic channel model

Abstract

Wireless channel modeling has always been one of the most fundamental highlights of the wireless communication research. The performance of new advanced models and technologies heavily depends on the accuracy of the wireless CSI (Channel State Information). This study examined the randomness of the wireless channel parameters based on the characteristics of the radio propagation environment. The diversity of the statistical properties of wireless channel parameters inspired us to introduce the concept of the tomographic channel model. With this model, the static part of the CSI can be extracted from the huge amount of existing CSI data of previous measurements, which can be defined as the wireless channel feature. In the proposed scheme for obtaining CSI with the tomographic channel model, the GMM (Gaussian Mixture Model) is applied to acquire the distribution of the wireless channel parameters, and the CNN (Convolutional Neural Network) is applied to automatically distinguish different wireless channels. The wireless channel feature information can be stored offline to guide the design of pilot symbols and save pilot resources. The numerical results based on actual measurements demonstrated the clear diversity of the statistical properties of wireless channel parameters and that the proposed scheme can extract the wireless channel feature automatically with fewer pilot resources. Thus, computing and storage resources can be exchanged for the finite and precious spectrum resource.