Personalized Federated Learning Based Adaptive Optical Compensation for Atmospheric Turbulence
Song Song, He Qiao, Yejun Liu, Po Wu, Qiming Sun, Lun Zhao, Ting-Wei Wu, Lei Guo
Abstract
Free-space optical (FSO) technologies face significant challenges from atmospheric turbulence, which can degrade transmission performance. Adaptive optics (AO) techniques offer a promising solution by compensating for turbulence-induced optical aberrations in real-time. In this paper, we propose a deep learning-based AO compensation scheme utilizing a mixed convolutional channel attention prediction network (MCCA-PNet). The network establishes a mapping relationship between distorted beam intensity information and atmospheric turbulence, enabling real-time compensation on a spatial light modulator. In order to address the challenge of limited datasets across diverse atmospheric conditions, we introduce a personalized federated learning approach named federated layer similarity aggregation algorithm (FedLaySim). The proposed method enables the exchange of feature information while preserving data privacy through the use of a layer similarity algorithm. Compared to conventional local training methods, personalized federated training enhances phase screen prediction accuracy by 1.98 dB. The average mode purity of the compensated vortex beams (VB) reaches 0.96. Furthermore, the applicability of our AO compensation scheme in actual, real-world environments is verified through the use of atmospheric simulation in an environmental chamber.