Wireless ultraviolet scattering communication is a wireless communication technology based on atmospheric particle scattering. Due to its strong scattering characteristics, wireless ultraviolet can be applied to special scenarios such as non-direct vision. However, this strong scattering effect can lead to an obvious multipath effect of wireless ultraviolet and cause serious pulse broadening. In the case of a high data rate, this phenomenon will cause inter-symbol interference and even cause information misjudgment, leading to the increase of bit error rate and poorer communication performance. To improve wireless ultraviolet communication, it is necessary to study the signal processing technology for ultraviolet scattering channel. As a key technology in wireless optical communication, channel equalization can effectively suppress or eliminate inter-symbol interference. As an artificial intelligence method, deep learning has developed rapidly in recent years. With wide application, it can also be applied to the signal processing of wireless optical communication, which inspires channel equalization. In this paper, we combine deep learning technology with ultraviolet optical communication to achieve more efficient and intelligent wireless ultraviolet optical communication.

We study the channel problem of wireless ultraviolet (UV) scattering communication, and establish a single scattering channel model for non-line-of-sight UV. We analyze the scattering channel characteristics in terms of impulse response and path loss, to provide a suitable channel model for subsequent equalization. Then, we combine long short term memory recurrent neural network (LSTM) and deep neural network (DNN) to develop a blind equalization method for UV scattering channel based on a hybrid neural network, which can preprocess the training data into a time sequence, and process the temporal dependence of the input signals through LSTM to extract useful temporal features. The nonlinear features of the signal data are further explored using DNN to enhance the prediction performance of the model, which features flexibility, adaptivity, and nonlinear modeling capability, and is capable of learning and adapting to complex UV scattering channels without prior information. With sufficient training sample data and the learning capability of the hybrid neural network, the signal can be equalized accurately and efficiently.

Based on the bit error rate (BER) and mean square error (MSE) as indicators, the proposed scheme (LSTM-DNN), the classical adaptive equalization algorithms [least mean square (LMS) and recursive least square (RLS)], and the DNN-based channel equalization scheme are comprehensively compared and analyzed. When the signal-to-noise ratio (SNR) is less than 5 dB, the BER curves of the algorithms are close to coincident; when the SNR is greater than 5 dB, it is observed that the BERs of DNN and LSTM-DNN begin to be gradually lower than those of LMS and RLS, with the BER of LSTM-DNN significantly lower than that of the DNN; when the SNR exceeds 9 dB, the BER of LSTM-DNN can be reduced by 0.5 to 2 orders of magnitude compared with that of the traditional algorithm [Fig. 10(a)]. Similarly, when the SNR is less than 5 dB, the MSE curves of LSTM-DNN and DNN are close to coincident, and the MSE is slightly lower than those of LMS and RLS; when the SNR is greater than 5 dB, the MSE of the LSTM-DNN is the lowest of all the algorithms [Fig.10(b)]. These results show that with a high SNR, the neural network model can better capture the difference between signal and noise, so DNN and LSTM-DNN show better equalization performance when the SNR is greater than 5 dB, while the LSTM in LSTM-DNN can automatically capture the temporal correlation in the signal, so it is more suitable for feature extraction of signal sequences.

Aiming at the serious pulse broadening and signal attenuation of non-line-of-sight wireless UV optical communication due to various factors such as atmospheric scattering, we propose a blind equalization method based on a hybrid neural network for the UV scattering channel. In this method, LSTM and DNN are combined, and the received signal is treated as a time sequence, without the need to study prior channel information. Also, LSTM’s powerful learning ability regarding temporal memory sequence is used to extract the characteristics of the received signal to recover the original signal. Simulation results show that when the SNR is greater than 11 dB, the BER of the proposed algorithm can be reduced by one to two orders of magnitude and the MSE is reduced by more than 0.5 orders of magnitude compared with the LMS algorithm and the RLS algorithm; compared with DNN, the equalization performance of the proposed algorithm is better, and the BER and the MSE of the proposed algorithm are reduced by 81.0% and 27.8% respectively when the SNR is equal to 11 dB, proving that the hybrid neural network has stronger noise suppression ability.