Underwater optical wireless communication system performance improvement using convolutional neural networks

Many applications that could benefit from the underwater optical wireless communication technique face challenges in using this technology due to the substantial, varying attenuation that affects optical signal transmission through waterbodies. This research demonstrated that convolutional neural networks (CNNs) could readily address these problems. A modified CNN model was proposed to recover the original data of a non-return to zero on–off keying modulated signal transmitted optically through a tank full of Gulf seawater. A comparison between the proposed CNN model and a conventional fixed-threshold decoder (FTD) demonstrates the excellent performance of the proposed CNN model, which improved the bit error ratio (BER), signal-to-noise ratio (SNR), and effective channel length. The BER of the optical signals that are transmitted at powers of 24, 26, and 27 dBm and a bit rate of 10 Mbit/s at a distance of 3 m from the transmitter when FTD is used is 7.826 × 10−7, 5.049 × 10−8, and 8.38 × 10−10, respectively. When the CNN decoder is used at the same distance and powers, the BER is 6.23 × 10−14, 1.44 × 10−16, and 2.69 × 10−18, respectively. In conclusion, the BER decreased by about seven orders of magnitude, the effective channel length increased by four times, and the SNR decreased by about 20 dB. The simplicity of the proposed CNN decoder is independent of the prior knowledge of the channel conditions. Furthermore, the magnificent obtained results make the proposed CNN decoder an ideal substitute for ordinary underwater optical wireless communication decoders.