Asymmetric CNN-based equalization for clipping distortion compensation in LED-based VLC systems
Visible Light Communication (VLC) systems utilizing Direct Current-biased Optical Orthogonal Frequency Division Multiplexing (DCO-OFDM) are inherently limited by a high Peak-to-Average Power Ratio (PAPR), which leads to non-linear distortions due to the restricted dynamic range of LEDs. Conventional PAPR reduction techniques often impose significant computational overhead on the transmitter or severely degrade the Error Vector Magnitude (EVM). This paper proposes an asymmetric deep learning-based framework to mitigate clipping noise. We introduce a low-complexity clipping approach at the transmitter ((𝑁)) and shift the compensation burden to a Convolutional Neural Network (CNN) at the receiver. Two clipping ratios were evaluated: CR = 0.9 (moderate, 2.9 dB PAPR reduction) and CR = 0.7 (aggressive, 5.8 dB PAPR reduction). Experimental results from 20 spatial positions within a 2.0 × 1.0 m2 test area demonstrate that the proposed CNN effectively recovers the signal under both clipping conditions. For 64-QAM modulation with CR = 0.7,
the CNN achieves a mean EVM of 7.07%, below the 8% reference threshold defined by 3GPP standards, and a mean BER of 8.98 × 10−4, while the conventional MMSE equalizer reaches 15.05% EVM and 5.37 × 10−2 BER. The architecture shows consistent generalization and spatial robustness compared to standard Deep Neural Networks (DNNs), suggesting its potential for LED-based lighting applications.
RVFL-Based Nonlinearity Compensation for MISO VLC Systems: Experimental Validation and Performance Analysis
Visible Light Communication (VLC) systems employing multiple Light-Emitting Diode (LED) transmitters face significant challenges from aggregate nonlinear distortions that degrade signal quality and limit practical deployment. This work presents experimental validation of Random Vector Functional Link (RVFL) networks for post-distortion compensation in Multiple-Input Single-Output (MISO) VLC systems. RVFL’s closed-form training and memoryless architecture offer distinct computational advantages over iterative methods. Comprehensive experimental validation across variable drive amplitude (300-900 mV) and bias current (50-350 mA) in a 2×1 dual-luminaire testbed demonstrates that RVFL achieves EVM of 7.81% for 16-QAM and 3.97% for 64-QAM at 200 cm distance, enabling reliable transmission where uncompensated signals exceed thresholds. Comparative analysis reveals superior computational efficiency: RVFL inference requires 45.5 ms versus 502.5 ms for Artificial Neural Networks (ANN) and 338.5 ms for Volterra, achieving 11× faster inference than ANN. These results position RVFL as an attractive candidate for practical VLC receivers where computational efficiency is critical.