Channel Estimation of SISO FSO System with Deep Inception Residual Convolution Neural Network

Free-space optical (FSO) communication offers advantages such as high data rates, low latency, and cost-effective deployment over long distances without the need for physical cables. However, it is highly susceptible to atmospheric turbulence, misalignment, and noise, which degrade the accuracy of channel state information and affect overall system performance.This paper presents a novel deep learning (DL)-based method for channel estimation in single-input single-output FSO communication systems using deep inception residual convolution neural network (IRCNN). The inception modules allow for multi-scale feature extraction, and the incorporation of residual connections alleviates the vanishing gradient issue, thereby facilitating the efficient training of deeper network architectures. Additionally, the model can mitigate the error floor seen in traditional fixed threshold detectors, leading to enhanced bit error rate (BER) performance. Extensive simulation results demonstrate that the IRCNN model substantially outperforms traditional estimation methods including least squares and other DL models, in terms of mean square error MSE (10−5and BER (10−6at 19dB SNR. It also performs well in harsher channel conditions where the other estimator's performance deteriorates.