Reconfigurable intelligent surface (RIS) has been considered as an effective approach to achieve extremely large-scale MIMO (XL-MIMO). To mitigate the high-complexity of channel information acquisition brought by the large number of RIS elements, beam training has been applied to select an optimal beam from a predefined codebook for beamforming in large-scale RIS-aided systems. However, practical limitations of RIS, such as its macrocell structure and finite phase shifts, result in the overlap of codeword coverage, i.e., non-orthogonal beams, which degrades the accuracy of beam training. In this paper, we propose a self-correcting hierarchical beam training scheme, where we model the beam training as a sequential process based on a tailored long short-term memory network. Received powers of sequentially selected codewords layer by layer are used to construct quasi-orthogonality for optimizing the beam selection. Unlike traditional deep learning methods relying on sufficient data, which takes substantial overhead to collect, we design a diffusion-empowered generation module given the general data-constrained conditions. The generation module is capable of synthesizing codebook power profiles to enhance the beam training performance. Simulation results demonstrate that our proposed method outperforms existing beam training approaches in terms of accuracy and sum rate, even in the presence of dataset limitations.

Reconfigurable intelligent surfaces (RISs) are considered as an efficient solution for the implementation of extremely large-scale MIMO (XL-MIMO). To mitigate the complexity of channel information acquisition, beam training is identified as an effective solution by selecting the optimal beam from a predefined codebook. However, practical hardware limitations of RIS result in the overlap of codeword coverage, i.e., non-orthogonal beams, which degrades the accuracy of optimal beam selection. In this paper, we propose a self-correcting hierarchical beam training scheme, where beam training is designed as a sequential process based on a tailored long short-term memory network. Received powers of sequentially selected multi-layer codewords are integrated to construct quasi-orthogonality for optimizing the beam selection. Beam training is thus reformulated as a sequential codeword classification problem. A codeword priority adjustment procedure is then designed to prevent the beam selection from looping among layers according to the visit frequency to each codeword of the current selection path. Simulation results show that our proposed method outperforms existing beam training approaches in terms of beam selection accuracy and sum rate in the presence of hardware limitations.

Intelligent omni-directional surfaces (IOS), which can simultaneously reflect and refract incident signals, are considered as a promising solution for enhancing communication quality. To conduct joint beamforming of the BS and IOS, beam training is introduced such that perfect channel state information is not required anymore. However, the propagation environment is usually dynamically varying in practice, leading to frequent beam training procedures and huge training overhead. In this paper, we propose a transfer learning based beam training scheme for the IOS-assisted multi-user system to adapt to the dynamically changing propagation environment. We first build on an offline phase to train a beam prediction model that outputs the optimal beam with the highest data rate given only the received power of a small number of beams as the input. Then a transfer learning based method is developed such that the above beam prediction model can be updated to adapt to the dynamic environment rapidly. Simulation results demonstrate that the proposed scheme outperforms the existing beam training schemes in dynamic environments in terms of the convergence speed and the sum rate.