Under the condition of low SNR, enhancing the precision of parameter estimation for linear frequency modulation (LFM) signals and diminishing the complexity of the relevant methods represent crucial challenges that are presently being confronted. To address this problem, a parameter estimation method for LFM signals based on the High-Resolution Representation network (HRRnet) is proposed. The fundamental concept underlying this method lies in the employment of a strategy that combines the expansion of the receptive field with the fusion of multi-scale features. This enables the efficient extraction of both global and local information, which in turn augments the expressive power of the inherent signal characteristics and consequently mitigates the impact of noise interference. Based on this strategy, a high-resolution representation of the time–frequency spectrum of the signals is performed to improve the distinguishability of the time–frequency spectrum, and it further improve the accuracy of parameter estimation for LFM signals. In addition, the network utilizes dilated convolution to expand the receptive field while reducing the dependence on network depth, so as to control the network complexity and further optimize the computational efficiency. Experimental results show that when the SNR is greater than −12 dB and the tolerable error is equal to 0.1, the average accuracy of the HRRnet method for estimating the initial frequency and frequency modulation coefficient of LFM signals can reach above 95.53% and 91.19%, respectively, and its number of parameters and computational complexity are reduced to more than 20.47% and 20.37% of those of the existing methods.

Currently, in the application scenario of generative adversarial networks, determining how to improve the quality of the generated signals and ensure the modulation recognition accuracy of convolutional neural networks are important problems. In this paper, a generative sample quality screening method for the problem of low-quality samples generated by generative adversarial networks under few-shot conditions has been proposed, which innovatively establishes a sample expansion mode without fixing the network parameters, realizes the learning of the real data distribution by constantly updating the network weights, and enhances the quality of the expanded samples by adopting the quality screening method with two quality screenings. A generative adversarial network has been designed for this method, which reduces the time investment required for generating samples by extracting different features of few-shots of signals. The experimental results show the few-shot conditions, under the signal-to-noise ratio of −8∼12 dB and three expansion ratios of 1:1, 1:2 and 1:3. Compared with the general method expansion, the average modulation mode recognition accuracy of the QCO-CFGAN method expanded with the quality screening method is improved by 2.65%, 2.46% and 2.73%, respectively, which proves the effectiveness under this condition.

In response to the problem where neural network models fail to fully learn signal sample features due to an insufficient number of signal samples, leading to a decrease in the model’s ability to recognize signal modulation methods, a few-shot signal modulation mode recognition method based on reinforcement metric meta-learning (RMML) is proposed. This approach, grounded in meta-learning techniques, employs transfer learning to building a feature extraction network that effectively extracts the data features under few-shot conditions. Building on this, by integrating the measurement of features of similar samples and the differences between features of different classes of samples, the metric network’s target loss function is optimized, thereby improving the network’s ability to distinguish between features of different modulation methods. The experimental results demonstrate that this method exhibits a good performance in processing new class signals that have not been previously trained. Under the condition of 5-way 5-shot, when the signal-to-noise ratio (SNR) is 0 dB, this method can achieve an average recognition accuracy of 91.8%, which is 2.8% higher than that of the best-performing baseline method, whereas when the SNR is 18 dB, the model’s average recognition accuracy significantly improves to 98.5%.

As drone technology develops rapidly and many users emerge in airspace networks, various forms of interference have caused the wireless spectrum to exhibit a dense, diverse, and dynamic trend. This increases the probability of spectrum conflicts among users and seriously impacts the quality and transmission rate of communication. How to effectively improve the detection accuracy of each frequency component in the electromagnetic space mixed signals and avoid spectrum conflicts will become one of the crucial issues currently faced by unmanned aerial vehicle (UAV) communication technologies. However, the existing methods overlook the mutual interference among the component signals as well as the noise during the frequency detection process, which affects their detection performance. In this paper, we propose a mixed-signal frequency detection method based on the reconstruction of sparse feature signals. Without information such as frequency range, bandwidth, and the number of components, it can utilize the autoencoder network to learn the sparse features of each component signal in the high-dimensional frequency domain space and construct a nonlinear reconstruction function to reconstruct each component signal in the mixed signal, thereby realizing the separation of signals. On this basis, complex dilated convolution and deconvolution are used successively to perform feature extraction on the separated signals, which enhances the receptive field and frequency resolution ability of the network for signals, reduces the interference between noise and different component signals, and realizes the accurate estimation of the number of components and carrier frequencies. The simulation results show that when $SNR\ge$ 6 dB, the detection accuracy of the number of component signals is greater than 96.3%. The detection error and detection accuracy of component frequencies are less than 3.19% and greater than 90.7%, respectively.

Polar codes are closer to the Shannon limit with lower complexity in coding and decoding. As traditional decoding techniques suffer from high latency and low throughput, with the development of deep learning technology, some deep learning-based decoding methods have been proposed to solve these problems. Usually, the deep neural network is treated as a black box and learns to map the polar codes with noise to the original information code directly. In fact, it is difficult for the network to distinguish between valid and interfering information, which leads to limited BER performance. In this paper, a deep residual network based on information refinement (DIR-NET) is proposed for decoding polar-coded short packets. The proposed method works to fully distinguish the effective and interference information in the codewords, thus obtaining a lower bit error rate. To achieve this goal, we design a two-stage decoding network, including a denoising subnetwork and decoding subnetwork. This structure can further improve the accuracy of the decoding method. Furthermore, we construct the whole network solely on the basis of the attention mechanism. It has a stronger information extraction ability than the traditional neural network structure. Benefiting from cascaded attention modules, information can be filtered and refined step-by-step, thus obtaining a low bit error rate. The simulation results show that DIR-Net outperforms existing decoding methods in terms of BER performance under both AWGN channels and flat fading channels.