Digital-twin-enabled federated learning and CNN-based channel estimation for urban vehicular channels

Abstract

Urban vehicular channel estimation (UVCE) has long been a difficult task due to the inter-carrier interference (ICI) and path loss caused by high-speed vehicle motion and urban geographical features (e.g., buildings, vehicles, and trees). Conventional estimators based on pilot symbols and channel statistics generally assume static signal propagation models, such as Free-space and Rayleigh fading. These models are inadequate for addressing path losses caused by geographical features, leading to limited performance. In contrast, centralized learning (CL)-based estimators can provide higher estimation performance by collecting channel data from a specific geographical area for training. However, when the UVCE is scaled to city size, CL estimators cannot precisely recognize the channel characteristics of each location in the city, resulting in decreased estimation accuracy. To further improve the scalability and accuracy of UVCE, this paper proposes a federated learning (FL) and convolutional neural network (CNN)-based channel estimator, referred to as FL-CNN. FL is used to aggregate multiple local channel models (LCM), which are clustered by the K-Dimension (KD)-tree technique. For each LCM, we employ the ray-tracing model to calculate the path loss caused by geographical features and use CNN to estimate the channel. Our results show that at the low signal-to-noise-ratio (SNR) regime (e.g., 10 dB), the estimation and data recovery performance of the FL-CNN estimator are respectively 61% and 65% higher than those of benchmark estimators on average.