Low-SNR and BER reduction in UWOC systems using DESN and CNN-TCN deep learning models.

Both commercial and scientific underwater wireless optical communication (UWOC) systems are essential and significant for several applications with the ability to provide high data transmission rates over distances up to tens of meters. There are several research gaps for UWOC like the limitation of the UWOC performance which is impacted by the intrinsic characteristics of ocean water which includes losses that reduce the signal-to-noise ratio (SNR) and impede communication quality. The main objective of this study is to enhance the performance of the UWOC system by utilizing recent artificial intelligence (AI) technologies to reduce the bit error rate (BER) with different neural network (NN) models. We apply both pulse amplitude modulation (PAM) and quadrature phase shift keying-orthogonal frequency division multiplexing (QPSK-OFDM) with several NN models across a range of underwater transmission ranges with PAM of 100, 125, 167 m and QPSK-OFDM of 100, 110, 120, 130 m. Deep echo state network (DESN) and recursive least square (RLS) are the two channel estimation techniques used. A variety of deep learning (DL) models and convolutional neural network (CNN) are assessed to improve system robustness and efficiency. CNN-RNN-AM shows performance gains of 18.9, 23.5, 33.3, and 29.4% at 100, 110, 120, and 130 m, respectively. CNN-LSTM-AM shows gains ranging from 10 to 17.6%; and TCN-LSTM-AM shows the largest gains, 29.4, 41.17, 53.3, and 71.4%, at the same distances. As implications of these results that, by comparison, DESN improves by up to 25% at 167 m, while PAM with TCN-LSTM-AM only improves by 17.3% at 100 m utilizing RLS. The most notable improvements are seen in CNN-LSTM-AM and TCN-LSTM-AM, which achieve accuracy gains of + 10.41 and + 11.80%, respectively, along with notable decreases in error metrics like mean squared error (MSE) of 58.46 and 66.15%, root mean square error (RMSE) of 30.48 and 41.90%, and mean absolute error (MAE) of 34.94 and 40.96%. Remarkably, TCN-RNN-AM exhibits a significant improvement in accuracy of 7.25% and RMSE of 32.38%, but it also experiences a minor rise in MAE of −4.82%, suggesting less consistent absolute error handling. When AM is added to CNN-LSTM, the accuracy increases from 94.33 to 96.77%, and the MSE, RMSE, and MAE decrease to 0.0054, 0.073, and 0.054, respectively. The TCN-LSTM model with AM also attains the maximum accuracy of 97.99%, indicating a 3.65% improvement. These findings demonstrate how well DL works in conjunction with sophisticated modulation and channel estimation methods to significantly enhance UWOC system performance in demanding underwater conditions.