深度学习辅助的超奈奎斯特速率光空间脉冲位置调制

Abstract

Objective As an innovative multipleinputmultipleoutput (MIMO) technology, optical spatial modulation (OSM) resolves antenna interference and synchronization challenges in MIMO systems by selecting a single antenna to carry information and collectively transmits the antenna index as additional information. However, existing OSM research predominantly adheres to the orthogonal transmission criterion, and imposes limitations on enhancing the transmission rate of the system although the research is effective in avoiding intersymbol interference. To this end, the introduction of nonorthogonal transmission via Faster -Than -Nyquist (FTN) technology compresses symbol intervals during pulse shaping, enabling an increase in transmission rate within the same bandwidth per unit time. As a result, we propose a novel Faster -Than -Nyquist rate optical spatial pulse position modulation scheme that combines OSM with FTN to further enhance the transmission rate and spectrum efficiency of the system. Additionally, in response to the highly complex receiver issue, a multiclassification neural network (MNN) decoder is proposed to significantly reduce computational complexity and achieve approximate optimal detection. Methods At the transmitting end, the input binary bit stream is divided into two groups of data blocks after serial/parallel transformations. The first group of data blocks is mapped to the index of the selected lasers for each symbol period, whilethe second group is mapped to pulse position modulation (PPM) symbols. An FTN shaping filter is employed to compress the PPM symbols. Then, the compressed PPM -FTN signals are loaded onto the chosen lasers for transmission. The signal traverses the Gamma -Gamma channel, and it is received by photodetectors (PDs) and converted into an electrical signal for further signal processing at the receiving end. Initially, downsampling is conducted to obtain a signal with the same dimensionality as the input signal. The downsampled signal is then classified based on its effective features, with each class being assigned the corresponding label. Subsequently, different samples with varying signaltonoise ratios (SNRs), along with their associated label values, are utilized as input and output for offline training of a neural network model. The objective is to achieve optimal decoding accuracy by defining average loss and learning rate parameters to construct an MNN, which helps determine the number of hidden layers and neurons. Finally, the wellconstructed MNN is employed for online signal detection. Then, inverse mapping is conducted on output label values from the decoder to recover the corresponding modulation symbols and laser index.Results and Discussions Monte Carlo simulations are conducted to evaluate the proposed scheme in a Gamma -Gamma channel. We first derive an upper bound of the average bit error rate (ABER) of the system and provide a comparison of the simulated BER with the ABER in Fig. 3. The results show that the two curves asymptotically coincide at high SNRs, which de...