Recognizing prohibited items in X-ray security checking images automatically based on Convolutional Neural Networks (CNNs) has attracted attention increasingly. However, there are no suitable X-ray security checking image databases to train a reliable CNN model. Therefore, we propose a data augmentation method for X-ray security checking images. First, a lot of new X-ray prohibited item images are generated using the improved Self-Attention Generative Adversarial Network (SAGAN). Next, a Cycle GAN based method is proposed to transform the item natural images into the X-ray images. It can enrich the diversity of the new images, including item shape and pose. Then, we combine the prohibited item images with background images to synthesize the new X-ray security checking images. Finally, two single shot multi-box detector (SSD) models are applied to verify whether the enlarged database has achieved data augmentation. Experimental results show that the performance of SSD model trained by the enlarged database is better than the SSD model trained by the original database. It implies that our method can achieve data augmentation for X-ray security checking images effectively.

High-fidelity modeling is of great importance for digital twin (DT) technology. A Transformer-based modeling method for radio over fiber (RoF) link is developed and experimentally demonstrated, in which the Transformer is a kind of deep learning network. The modeling performance of the constructed model by using the proposed method is evaluated, which is better than the ones based on fully connected neural network (FCNN), convolution neural network (CNN), and generative adversarial network (GAN). In addition, the modeling consistency with different laser launch power, fiber length, carrier frequency, and signal-to-noise ratio (SNR) is also investigated. Experimental results show that the fitting waveforms predicted by the proposed method agree well with the actual waveforms at the receiver end. The normalized mean square errors of the fitting waveforms are lower than the acceptable bound of 0.02.

A data-driven digital demodulator based on a Fourier layer Transformer network (FTnet) for radio over fiber (RoF) transmission system with quadrature amplitude modulation (QAM) is developed and experimentally demonstrated. The FTnet combines the Transformer encoder with the Fourier layer to learn the waveforms and directly recover the bitstreams from the impaired received signals. The FTnet-based demodulator does not rely on a series of digital demodulation algorithms such as frequency offset compensation, down-conversion, equalization, and decoding, making the process more efficient and accurate. The 10 GHz 2 Gsym/s 25 km RoF transmission systems are established to evaluate the proposed FTnet-based digital demodulator experimentally. The results show that the bit error rates (BERs) performance of the proposed demodulator for the 16-QAM RoF is better than the ones based on a fully connected neural network, Transformer, and traditional digital demodulator with the least mean square error equalizer (TDD-LMS). The optical receiving sensitivity for the 64-QAM RoF system based on the proposed demodulator is improved by 3 dB compared to TDD-LMS under a BER limit of 3.8×10 ^-3 . Furthermore, our proposed demodulator outperforms other demodulators for the RoF system with wireless transmission at different received optical powers and wireless distances.

Multipath self-interference (MSI) between the transmitter and receiver is an inevitable problem in realistic in-band full-duplex (IBFD) transmissions. Due to the dynamic and intricate MSI channels, it is crucial for the MSI cancellation system to quickly obtain the optimal configuration of multi-dimensional parameters. To overcome this problem, a photonics-assisted adaptive MSI cancellation scheme using deep reinforcement learning (DRL) is proposed and experimentally demonstrated. Multiple local reference signals are obtained using optical wavelength division multiplexing to achieve MSI cancellation. An adaptive algorithm based on DRL is applied for adaptive optimization of multi-dimensional optical parameters to achieve real-time MSI cancellation. To manipulate the amplitude and time delay of the multiple-path local reference signals, a multipath optical tunable delay line (MOTDL) module is introduced. A proof-of-concept experiment was carried out to verify the feasibility of the proposed scheme. By superimposing three-path local reference signals, a MSI cancellation depth of 24 dB with a bandwidth of 1 GHz was achieved. After undergoing random exploration, the adaptive algorithm learns the state relationship between the MSI and multiple-path local reference signal. This facilitates achieving optimal cancellation within just 5 steps from most random starting point. In the experiment, a 16-QAM OFDM signal with a bandwidth of 600 MHz was successfully recovered from the MSI signal with a 1 GHz bandwidth in an IBFD transmission. Furthermore, the adaptive capability of the proposed algorithm is also validated in response to varying MSI conditions.

End-to-End (E2E) learning has increasingly become a dominant method for enhancing the performance of communication systems. In this paper, an E2E Radio Over Fiber (RoF) transmission system utilizing self-supervised learning (SSL) is proposed. This E2E approach enables automatic optimization of the transmission system's transmitter modulator and receiver demodulator based on specific channel characteristics. The SSL architecture comprises four deep neural networks: TransNN for symbol mapping, SamplNN for upsampling, ChannelNN for channel modeling, and ReceivNN for demodulation, collectively replacing traditional components in the RoF link. Indeed, the E2E system adjusts the geometric shape of the constellation and upsampling rules according to RoF channel properties, facilitating transmission with enhanced received sensitivity. Perturbation noise is incorporated during the training phase to improve the ability to generalize the SSL. Notably, traditional demodulation methods cannot demodulate the RF signals transmitted in the E2E system, thereby introducing an additional layer of confidentiality to the transmission process. Numerical simulations have been conducted in 10 GHz 2 Gsym/s RoF transmission systems. The results indicate that compared to traditional approaches, the received sensitivity of the E2E system improved by 3.5 dB, under a BER limit of 2.4e-2. Compared to optimized only at the receiver side, achieved a sensitivity improvement of at least 3 dB. The numerical experiments also validate the importance of the SamplNN and perturbation training for the E2E system. These elements can improve the system's transmission performance and generalization ability, as well as enhance the system's security.

Fast and accurate modeling techniques are crucial for improving the performance of long-distance polarization division multiplexing (PDM) transmission. Traditional methods, such as the Split-Step Fourier Method (SSFM), suffer from high computational complexity, especially for long-distance and large-scale data. To overcome this challenge, a BERT-based modeling method for long-distance PDM transmission channel is proposed, which can model the various orders of dispersion, nonlinear effects, and polarization mode dispersion (PMD) in PDM systems. The BERT-based method enables fast and accurate modeling of transmission channels without requiring iterative processes, in which BERT is a deep learning model leveraging the attention mechanism. Numerical experiment results show that the fitting waveforms of the proposed method closely match the actual waveforms over 1200 km across various launch powers, modulation formats and erbium-doped fiber amplifier (EDFA) noise, achieving normalized mean squared errors below 0.003. Furthermore, the computational complexity of the BERT-based method is approximately 1/3800 of that of the SSFM.