The application of deep learning improves the processing speed and the accuracy of automatic modulation recognition (AMR). As a result, it realizes intelligent spectrum management and electronic reconnaissance. However, deep learning-aided AMR usually requires a large number of labeled samples to obtain a reliable neural network model. In practical applications, due to economic costs and privacy constraints, there is a small number of labeled samples but a large number of unlabeled samples. This paper proposes a Transformer-based contrastive semi-supervised learning framework for AMR. First, self-supervised contrastive pre-training of the Transformer-based encoder is completed using unlabeled samples, and data augmentation is realized through time warping. Then, the pre-trained encoder and a randomly initialized classifier are fine-tuned using labeled samples, and hierarchical learning rates are employed to ensure classification accuracy. Considering the problems of applying Transformer to AMR, a convolutional transformer deep neural network is proposed, which involves convolutional embedding, attention bias, and attention pooling. In experiments, the feasibility of the framework is analyzed through linear evaluation of the framework components on the RML2016.10a dataset. Also, the proposed framework is compared with existing semi-supervised methods on RML2016.10a and RML2016.10b datasets to verify its superiority and stability.

Automatic modulation recognition (AMR) technique is utilized to extract signal features and identify modulation types under non-cooperative conditions. The development and application of artificial intelligence have greatly enhanced the accuracy and efficiency of AMR. However, existing deep learning-based models typically design intricate backbone networks to extract deep features of signals. Moreover, the training of networks relies on supervised information, and the number of samples is crucial to model performance. This paper proposes a masked transformer model for AMR (MTAMR), comprising two stages: mask training and fine-tuning. The signal is preprocessed into in-phase orthogonal, amplitude phase, and Fourier transform vectors, and then concatenated into a multimodal sequence to obtain a more efficient representation. In the mask training stage, the multimodal sequence is masked by random zeroing, and global features are extracted through the Transformer encoder. Subsequently, these features are processed by two methods to obtain the prediction sequence and classification probability. Considering the homogeneous uncertainty of the prediction and classification tasks, a joint loss with learnable weights is constructed. In the fine-tuning stage, the unmasked multimodal sequence is taken as input, and the cross-entropy loss with a low learning rate is used to update the model. Simulation results show that MTAMR with a simple backbone structure, achieves significant recognition accuracy on multiple datasets by leveraging mask modeling and multimodal fusion representation, while the computational complexity is feasible.

The application of deep learning (DL) has improved the reliability and intelligence of automatic modulation recognition (AMR). However, real-world scenarios often involve a limited number of signal samples. Furthermore, most existing DL-based AMR models improve performance at the cost of computational complexity. Therefore, we propose an efficient contrastive multistage sparse attention network (CMSSAN) model for few-shot AMR without auxiliary datasets. Specifically, supervised contrastive learning is utilized to enhance the feature representation of the signal, and a joint loss with dynamic weights is constructed to balance the representation and classification tasks. In addition, a lightweight MSSAN encoder is proposed to enhance the recognition performance with lower computations and parameters. Simulation experiments are conducted on the ablation experiment and hyperparameter analysis of the proposed model, and the superiority of the model is verified on several datasets.

By exploiting the sparsity of the channel in the delay and angle domains, compressed sensing (CS) algorithms can be used for channel estimation of massive multiple-input multiple-output (MIMO) systems to reduce pilot overhead. Due to the Doppler frequency shift, however, the intercarrier interference (ICI) and the rapid change of the channel state result in the poor estimation effect of doubly selective (DS) channel. In this paper, we propose the block sparsity adaptive matching pursuit (B-SAMP) algorithm to solve this problem. Firstly, the complex exponential basis expansion model (CE-BEM) is used to convert numerous channel tap coefficients into BEM parameter vectors and then the sparsity adaptive channel estimation scheme based on compressed sensing is proposed. Specifically, the ICI-free model is obtained by using the proposed equally placed pilot group scheme, and the B-SAMP algorithm is proposed by using the spatio-temporal common sparsity of the channel to complete the estimation of DS channel. Finally, a linear smoothing method is used to reduce the error caused by CE-BEM, thereby further improving the accuracy of the estimation. The simulation results show that the proposed method not only improves the estimation accuracy compared with the existing scheme but also requires fewer pilots.

In massive multi-input multi-output orthogonal frequency division multiplexing (MIMO-OFDM) systems, accurate channel state information (CSI) is essential to realize system performance gains such as high spectrum and energy efficiency. However, high-dimensional CSI acquisition requires prohibitively high pilot overhead, which leads to a significant reduction in spectrum efficiency and energy efficiency. In this paper, we propose a more efficient time-frequency joint channel estimation scheme for massive MIMO-OFDM systems to resolve those problems. First, partial channel common support (PCCS) is obtained by using time-domain training. Second, utilizing the spatiotemporal common sparse property of the MIMO channels and the obtained PCCS information, we propose the priori-information aided distributed structured sparsity adaptive matching pursuit (PA-DS-SAMP) algorithm to achieve accurate channel estimation in frequency domain. Third, through performance analysis of the proposed algorithm, two signal power reference thresholds are given, which can ensure that the signal can be recovered accurately under power-limited noise and accurately recovered according to probability under Gaussian noise. Finally, pilot design, computational complexity, spectrum efficiency, and energy efficiency are discussed as well. Simulation results show that the proposed method achieves higher channel estimation accuracy while requiring lower pilot sequence overhead compared with other methods.

Deep learning-based automatic modulation recognition (AMR) methods assume that all classes have the same number of samples. However, in practical applications, all classes are usually imbalanced and exhibit a long-tailed distribution, i.e., only some classes have many samples while the rest have fewer. There is a scarcity of research addressing this issue in AMR. In this paper, we propose a long-tailed AMR network based on supervised contrastive learning (SCL). Specifically, a reversed sampler augmentation module is introduced to improve the diversity of tail classes with fewer samples. The uniformly sampled and the reversely sampled augmentation samples in the batch are passed through the weight-sharing encoder and the projection head to obtain a balance feature space by the SCL loss. To achieve a trade-off between feature representation and classification, the SCL loss and the cross-entropy loss form a joint loss of dynamic weights that decay exponentially to update the model end-to-end. For the encoder, a Transformer with global dependency modeling is employed to enhance the representation of signals. Simulation results show that the proposed network effectively improves the recognition accuracy of tail classes under multiple sample size settings for different signal datasets.