Transfer Learning for Angle of Arrivals Estimation in Massive MIMO System

Angle-of-arrivals (AoAs) estimation is one of the key tasks of localization in massive multiple-input-multiple-output (MIMO) system. Among all AoAs estimation methods, the class of deep learning (DL) based methods has been recognized effective with high estimation accuracy and low complexity. However, the design for methods in this class requires obtaining adequate labeled training data and training a specific network for each of different channels in MIMO system, which is challenging and costly. To overcome this bottleneck, a transfer learning method is proposed in this paper based on the assumption that some com-mon features exist among different channel models. The proposed method is implemented by a deep residual network (ResNet) composed of transfer layers (to extract the common features) followed by fine-tuning layers (to extract other characteristic features) in cascade, thus without need of a specific network for each different channel. As a realization of the proposed framework, the ResNet is trained on the data from the source domain (i.e. the “basic” channel model) to obtain the weights of transfer layers (fixed thereafter), which are used to initialize the fine-tuning layers. Then, small amounts of data from the target domain (i.e., a more sophisticated channel model like spatial channel model), are used to adjust the weights of fine-tuning layers. Finally, extensive simulation results demonstrate that the AoAs estimation accuracy of the proposed method is comparable to that of supervised DL-based methods, for which a specific trained network is required for each channel.