Recently, neural networks have gained prominence as potent tools in the realm of channel modeling in optical wireless communication (OWC) systems. However, existing channel networks tend to ignore random intensity noise and phase noise induced by optical sources under spontaneous and excited radiation, and likewise the shot noise of the devices. They could only characterize deterministic noise such as ໿inter-symbol interference (ISI), ໿signal-to-signal beat interference (SSBI), and nonlinearities. In this paper, a novel channel simulator framework to optical wireless communication systems is proposed, which can jointly learn the deterministic noise and random noise of the real channel simultaneously. The proposed approach is a comprehensive noise joint channel estimator (CNJCE) architecture. A comprehensive comparison is carried out in both simulations, free space optical (FSO) experiments and underwater visible light communication (UVLC) experiments, including three classical deep learning algorithms, two tributaries heterogeneous neural network (TTHnet) with posteriori additive white Gaussian noise (AWGN), white Gaussian noise layer-based channel estimator (WGNCE) and conditional generative adversarial network (CGAN) based channel model. In order to represent both deterministic noise and random noise modeling capabilities, an overall estimation score is also proposed, based on frequency mismatch and random noise distribution error. The final experimental results show that CNJCE achieved the highest average estimation scores of 79.86 and 97.52 in real FSO and UVLC channels respectively. In contrast, TTHnet without AWGN or WGNCE only achieved 39.83 and 43.82. And CGAN based channel model completely lost signal characteristics, far worse than CNJCE. Through comparison of various performance, CNJCE demonstrates commendable capabilities in generalization and robustness, concurrently achieving optimal performance in channel modeling. To the best of our knowledge, this is the first time that deterministic and random noise have been fully modeled in optical wireless communication modeling. Its applicability is showcased through successful deployment in real FSO and UVLC systems, thus establishing a foundational framework for comprehensive end-to-end learning paradigms.

Modulation format recognition (MFR) is one of the key technologies in adaptive optical systems and widely used in both commercial and civil applications. With the rapid development of deep learning, MFR algorithm based on neural networks (NN) has achieved impressive success. Due to the high complexity of underwater channels, to gain better performance of MFR tasks in underwater visible light communication (UVLC), the NN tend to be designed with a complex structure, which is costly in computation and hinders fast allocation and real-time processing. In this paper, we propose a lightweight and efficient method based on reservoir computing (RC), whose trainable parameters are only 0.3% of common NN-based methods. To improve the performance of RC in MFR tasks, we propose powerful feature extraction algorithms including coordinate transformation and folding algorithm. The proposed RC-based methods are implemented for six modulation formats, including OOK, 4QAM, 8QAM-DIA, 8QAM-CIR, 16APSK, and 16QAM. The experimental results show that our RC-based methods take only a few seconds for training process and under different pin voltages of LED, the accuracy for almost all exceeds 90%, and the highest is close to 100%. Analysis on how to design a well-performed RC to strike a balance between accuracy and time cost is also investigated, providing a useful guide for RC implementations in MFR.

Fronthaul solutions based on optical wireless communication (OWC) have emerged as a promising technology for 6G communications. OWC systems offer advantages such as flexible deployment and low construction costs, but face significant challenges due to severe signal-signal beat interference (SSBI), nonlinear distortions and atmospheric turbulence. Efficiently optimizing these joint impairments is particularly crucial for large-scale deployment. Additionally, various mobile network architectures typically have distinct fronthaul requirements, and different applications may necessitate a certain level of isolation. In this paper, we propose a multi-band adaptive optimization autoencoder framework (MBAE) to achieve end-to-end (E2E) optimization for OWC-based fronthaul systems supporting multi-access architecture. The impairments from the multiple transceiver systems and the propagation link can be jointly optimized to achieve performance closest to the global optimum. Studies are conducted in a simulated channel and a real optical wireless communication system respectively to verify the feasibility of the proposed channel model and MBAE. Simulation results show that MBAE can obtain an average Q factor gain of 3 dB compared with the multi-band carrier-less amplitude and phase (CAP) modulation with neural network based post-equalization (NN) and decision feedback equalization (DFE). The experimental results show that MBAE can achieve a 1.9 dB received optical power (ROP) sensitivity gain compared with CAP using NN and DFE at a transmission rate of 36 Gbps under the 7% hard decision-forward error correction (HD-FEC) threshold of 3.8×10-3 utilizing a low-cost PD with 10G-class bandwidth and support a link budget of 44.7 dB. In the performance test under turbulent conditions, MBAE also exhibits superior average performance, achieving a mean Q factor gain of 1.41 dB compared with CAP using NN and DFE. The proposed method can be a promising solution to meet the high-speed and low-cost requirements of future optical wireless fronthaul systems.

Modulation format recognition (MFR) is a key technology for adaptive optical systems, but it faces significant challenges in underwater visible light communication (UVLC) due to the complex channel environment. Recent advances in deep learning have enabled remarkable achievements in image recognition, owing to the powerful feature extraction of neural networks (NN). However, the high computational complexity of NN limits their practicality in UVLC systems. This paper proposes a communication-informed knowledge distillation (CIKD) method that achieves high-precision and low-latency MFR with an ultra-lightweight student model. The student model consists of only one linear dense layer under a communication-informed auxiliary system and is trained under the guidance of a high-complexity and high-precision teacher model. The MFR task involves eight modulation formats: PAM4, QPSK, 8QAM-CIR, 8QAM-DIA, 16QAM, 16APSK, 32QAM, and 32APSK. Experimental results show that the student model based on CIKD can achieve comparable accuracy to the teacher model. After knowledge transfer, the prediction accuracy of the student model can be increased by up to 87%. Besides, it is worth noting that CIKD’s inference accuracy can reach up to 100%. Moreover, the parameters constituting the student model in CIKD correspond to merely 18% of the parameters found in the teacher model, which facilitates the hardware deployment and online data processing of MFR algorithms in UVLC systems.

The sixth-generation mobile communication is committed to creating a large-capacity, low-latency, three-dimensional communication network encompassing space, air, land, and sea. Visible light communication (VLC) uses carriers with wavelengths from 380 nm to 760 nm for information transmission, offering a spectrum of 400 THz. Besides, visible light falls within the underwater transmission window, which offers significant potential for sea-based communication network construction. However, underwater visible light communication (UVLC) faces the effects of turbulence, device nonlinearity, attenuation, and noise, which limit the increase of the capacity of UVLC. To address the effects of turbulence, considering that turbulence has different effects on different polarization directions, the received signal can be better reconstructed by increasing the observation dimension, like observing the signal in multiple polarization directions. We proposed the use of circularly polarized light for transmission with a dual-aperture receiver, where the dual apertures detect the horizontally and vertically polarized components of the light beam, respectively. For the fusion of the two received signals, and in order to resist undesirable factors such as device nonlinearity and high-frequency attenuation, we designed a sparse dual-polarization triple-branch network (S-DPTBN) for signal reconstruction and recovery. The sparsification reduces the complexity and improves the robustness of the network, and the triple-branch design can fully extract the information of the horizontal and vertical polarizations as well as the weighted sum of the two. To enhance the communication capacity, we integrated an 8-wavelength laser emission module for signal transmission. Finally, experimental results confirmed that the proposed system achieved data rates of 144.90 Gbps, 143.50 Gbps, and 140.70 Gbps in conditions of no turbulence, medium turbulence, and strong turbulence, respectively, which had an improvement of 14.00 Gbps, 14.00 Gbps, and 13.30 Gbps over traditional point-to-point systems with deep neural network-based equalizers.