One of the key challenges in wideband spectrum sensing is that it has to be performed with extremely low latency and high accuracy over a large band to detect tiny spectrum holes. Motivated by this, we introduce a parallel convolutional neural network (ParallelCNN)-based wideband spectrum sensing approach, which aims to detect tiny spectrum holes with high sensing accuracy and low latency. The ParallelCNN contains two parallel CNNs, where the I/Q samples are split into two parts and processed concurrently by the two CNNs. Afterwards, the outputs of the two CNNs are combined to improve spectrum sensing performance. We extensively evaluate the accuracy, latency and generality performance of ParallelCNN. Performance evaluations indicate that the ParallelCNN approach outperforms other deep learning (DL)-based approaches such as the DeepSense and TCN, in terms of a higher precision, a higher recall and a lower latency, under a wide range of signal-to-noise (SNR) levels. In addition, the ParallelCNN can be generalized to different protocols such as LTE-M uplink transmissions, 5G NR downlink transmissions and WiFi transmissions.

This paper proposes a new digital signal recovery (DSR) technique for next-generation power efficient low Earth orbit (LEO) satellite-to-ground communication systems, which feature additive white Gaussian noise (AWGN) channel and significant power variation. This technique utilizes the prior knowledge (i.e., nonlinearities of radio frequency power amplifiers (RF-PAs)) of space-borne transmitters to improve the quality of the signal received at ground stations by modeling and mitigating the imperfection using deep neural networks (DNNs). Benefiting from its robustness against noise and power variation, the proposed DNN based DSR technique (DNN-DSR), can correct high signal distortions caused by the nonlinearities and hence allows RF-PAs to work close to their saturation region, leading to a high power efficiency of the LEO satellites. This work has been validated by both simulations and experiments, in comparison with the power back-off technique as well as memory polynomial based DSR solutions. Experimental results show that the DNN-DSR technique can increase the drain efficiency of the space-borne RF-PA from 32.6% to 45% while maintaining the same error vector magnitude as the power back-off technique. It has also been demonstrated that the proposed DNN-DSR technique can handle a signal power variation of 12 dB, which is challenging for conventional solutions.

A novel joint satellite-transmitter and ground-receiver (JSG) digital pre-distortion (DPD) (JSG-DPD) technique is proposed to improve the linearity and power efficiency of the space-borne active phased arrays (APAs) in low Earth orbit (LEO) satellite communications. Different from the conventional DPD technique that requires a complex RF feedback loop, the DPD coefficients based on a generalized memory polynomial (GMP) model are extracted at the ground-receiver and then transmitted to the digital baseband front-end of the LEO satellite-transmitter via a satellite–ground bi-directional transmission link. The issue of the additive white Gaussian noise (AWGN) of the satellite–ground channel affecting the extraction of DPD coefficients is tackled using a superimposing training sequences (STS) method. The proposed technique has been experimentally verified using a 28 GHz phased array. The performance improvements in terms of error vector amplitude (EVM) and adjacent channel power ratio (ACPR) are 7.5% and 3.6 dB, respectively. Requiring limited space-borne resources, this technique offers a promising solution to achieve APA DPD for LEO satellite communications.

In cognitive radio (CR), compressed spectrum sensing (CSS) is an attractive technology for detecting spectrum holes over a large band with a sub-Nyquist sampling rate. Traditional compressed spectrum (CS) reconstruction methods exhibit poor reconstruction performance and high complexity, due to the unknown and dynamically changing sparsity order of the observed signal, the hand-crafted optimization parameters involved in solvers, and dozens or hundreds of iterations. To cope with these challenges, we propose a wideband signal CS reconstruction algorithm by merging the iterative shrinkage thresholding algorithm (ISTA) and deep learning (DL) through deep unfolding and refer to it as the ISTA deep reconstruction network (ISTADR-Net). The parameters in ISTADR-Net (e.g., regularization and step size parameters) are learned end-to-end rather than being adjusted manually. Next, we propose a convolutional neural network (CNN)-based wideband spectrum sensing (WSS) network (CNNWSS-Net), which makes full use of the sparse characteristics of the reconstructed signal in the frequency domain to improve the WSS performance. The simulation results show that the proposed CSS scheme combining ISTADR-Net and CNNWSS-Net works well in low and high compression ratio scenarios, low and high signal-to-noise ratio (SNR) regions. Compared to the traditional CSS methods and other DL-based CSS methods, the proposed CSS scheme yields superior performance in terms of the probability of detection and the probability of false alarm. Furthermore, the computational complexity and computation time of the proposed CSS scheme are reduced compared to the traditional CSS methods.

Deep learning (DL) has exhibited immense potential across several domains, including image classification, speech recognition, and language translation, among others. Furthermore, it has been effectively employed in the domain of wireless communications, addressing intricate scenarios that involve dealing with uncertain channel models, high-speed data transmissions, and precise processing demands. To this end, we present CRNN-ResNet, a DL-based architecture to replace the carrier frequency offset compensation, in-phase and quadrature mismatch correction, channel estimation, equalization, and soft demapping modules of the orthogonal frequency division multiplexing (OFDM) receiver. We propose a novel residual neural network (ResNet) architecture that alternately stacks convolutional and bidirectional long short-term memory (BiLSTM) layers. This approach leverages the state prediction capabilities of BiLSTM and the feature extraction capabilities of convolutional layers to enhance the performance of receivers. Simulation results show that CRNN-ResNet achieves comparable performance to the ideal linear minimum mean square error (LMMSE) method and outperforms the conventional approximate LMMSE (ALMMSE) method and other known DL-based receivers under various non-ideal factors including radio frequency impairments, multipath fading, and Doppler frequency shift. Furthermore, CRNN-ResNet is more robust than other DL-based receivers when the cyclic prefix is omitted. We also demonstrate that CRNN-ResNet has lower computational complexity or time consumption compared to the other DL-based receivers.

In this letter, we introduce a novel blind equalization algorithm based on the convolutional neural network (CNN), to improve the bit error rate (BER) performance of equalizers against the multipath fading effect and nonlinear distortion. In contrast to existing neural network (NN)-based blind equalization algorithms, the proposed algorithm performs equalization and soft demapping jointly, which allows it to obtain soft bits directly from received data. In addition, the input preprocessing module is used to reorganize the received data to fully exploit the feature information of the received signals. Afterwards, the two-dimensional (2D) convolutions can be operated on a three-dimensional (3D) input array obtained by input preprocessing to recover the transmitted bits. Simulation results show that the proposed algorithm outperforms other CNN-based blind equalization algorithms, in terms of BER performance, under a wide range of channel models, signal-to-noise (SNR) levels and modulation schemes. Strikingly, the BER obtained by the proposed algorithm approaches the theoretical BER of multiple quadrature amplitude modulation (M-QAM) in linear channels.