Evaluating DNN and LSTM nonlinear compensators for enhanced performance in DCO-OFDM system

Abstract

Abstract This study proposes a deep neural network (DNN) and long-short-term memory (LSTM) nonlinear compensators method for direct current (DC)-biased optical orthogonal frequency division multiplexing (DCO-OFDM) in indoor visible light communication (VLC) conventional to handle the nonlinearity and retrieve the high-fidelity signals, and compared in terms of performance and complexity. Unlike the data training after fast Fourier transform in existing deep neural network schemes, this study proposes a scheme that uses the time domain waveform data output by photodiodes for direct equalization. The OFDM signal at the receiving end is equalized, which can mitigate hybrid linear and nonlinear impairments and save spectrum resources without requiring the pilots’ assistance. Compared with conventional receivers based on different guide frequencies and existing DL-based reception methods, the proposed adaptive receiver approach yields better bit error rate performance at different signal-to-noise ratios. This research reveals the extreme sensitivity of the LSTM’s performance to system SNR. LSTM outperforms DNN in high signal-to-noise ratio (SNR) situations, but at low SNR, even with high complexity, LSTM falls short of DNN’s performance.