Automatic modulation classification (AMC), as a key technology of cognitive radio, has become a focal point of research. However, most deep learning-based AMC methods require an extensive number of labeled signals to acquire a comprehensive understanding of modulation types, placing substantial pressure on signal acquisition and labeling. To solve this issue, we propose a few-shot AMC (FSAMC) method to facilitate rapid generalization and recognition with limited data, namely multi-scale feature fusion and distribution similarity network (MS2F-DS). Firstly, we design a multi-scale feature fusion (MS2F) model, which aims to extract features with varying fields of view and boost feature fusion, enabling the derivation of contextual information from the signal. Furthermore, we introduce a distribution similarity (DS) classifier to address the insufficient measurement of current similarity measurement functions by considering both micro and macro perspectives of vectors, further increasing intra-class compactness and inter-class separability. Finally, extensive experiments were conducted on 3-way 1, 3, and 5-shot FSAMC tasks using public datasets RML2016.10a and RML2016.10b, and the results demonstrated the effectiveness of our method.

Recently, significant progress has been made in deep learning, which has been widely applied in automatic modulation classification (AMC) with remarkable outcomes. However, current deep learning based AMC (DL-AMC) algorithms show limitations in their ability to accommodate dynamically changing communication scenarios. With the increasing number of modulation types, most DL-AMC algorithms often need to be re-trained, making it hard to transfer previous knowledge to the new models. Also, modulation classification faces the difficulty of acquiring and annotating a large number of signals. To address these challenges, we have modeled a few-shot class-incremental AMC (FSCI-AMC) task and proposed a pseudo classes and stochastic classifiers based network (PASS-Net) to accomplish it. Firstly, the pseudo classes are generated to reserve space for new types, enhancing the model’s continuous learning capability. Additionally, stochastic classifiers ensure the reliability of generated pseudo classes. Finally, in the incremental session, both real and pseudo classes are used for modulation classification. To evaluate the proposed approach, experiments were conducted on 7 modulation types as base classes and another 7 modulation types as incremental classes. The results show superior performance in 7 sessions of 1-way 1-shot and 1-way 5-shot class-incremental experiments compared to other competitive methods.

The rapid advancement of B5G/6G and wireless technologies, combined with rising end-user numbers, has intensified radio spectrum congestion. Automatic modulation recognition, crucial for spectrum sensing in cognitive radio, traditionally relies on supervised methods requiring extensive labeled data. However, acquiring reliable labels is challenging. Here, we propose an unsupervised framework, Multi-Representation Domain Attentive Contrastive Learning, which extracts high-quality signal features from unlabeled data via cross-domain contrastive learning. Inter-domain and intra-domain contrastive mechanisms enhance mutual modulation feature extraction across domains while preserving source domain self-information. The domain attention module dynamically selects representation domains at the feature level, improving adaptability. The experiments through public datasets show that the proposed method outperforms existing modulation recognition methods and can be extended to accommodate various representation domains. This study bridges the gap between unsupervised and supervised learning for radio signals, advancing Internet of Things and cognitive radio development.

In recent years, deep learning has achieved remarkable success in synthetic aperture radar (SAR) automatic target recognition (ATR). However, the performance of most deep learning-based models heavily depends on a large quantity of high-quality labeled data, which presents severe challenges in the collection and annotation of SAR images. To improve the stability and accuracy of few-shot SAR ATR, we propose a multi-level contrastive learning driven and dependency matrix measured method, termed MCL-DMM. This method comprises two key components: the multi-level contrastive learning (MCL) module and the dependency matrix-based measurement (DMM) module. The MCL module enhances model performance by integrating both individual and cluster discriminative tasks, thereby leading to a plausible and powerful representation learning at both the individual and cluster level. This module aims to construct a robust and unbiased feature extractor with good intra-class tightness and inter-class separateness. Building on this foundation, the DMM module models the dependencies within the feature matrix, providing a comprehensive measurement for few-shot SAR ATR. The cooperation between the robust feature extractor and the precise classifier enables the model to achieve superior performance. Experimental results on the standard MSATR dataset demonstrate that MCL-DMM outperforms current models, positioning it as a highly efficient solution for few-shot SAR ATR.

Automatic modulation recognition(AMR) is a key technology in the field of electronic reconnaissance. Benefiting from the powerful feature extraction ability of deep neural networks, the AMR algorithm based on supervised deep learning usually achieves pleasant performance. However, in non-cooperative scenarios high-quality and reliable modulation labels are difficult to obtain. Therefore, In this article, a novel unsupervised automatic modulation recognition method based on multi-domain representation contrastive learning (MRC-based UAMR) is proposed. By carefully utilizing unlabeled signal data with a self-supervised contrastive training step across multiple diverse representation domains, which is a more effective way than directly using IQ data as inputs, MRC aims to obtain high-quality signal feature information from unlabelled signals. The conducted experiments clearly showed that MRC remarkably outperforms existing unsupervised algorithms in modulation recognition and exhibits superior generalization ability. Our work significantly reduces the disparity between supervised and unsupervised representation learning in the field of automatic modulation recognition.

The study of open-set recognition for modulation types in communication signals is of high research significance, as it addresses critical challenges in wireless communication systems, such as spectrum monitoring, interference identification, and secure transmission. Traditional closed-set recognition methods are limited to pre-defined modulation types, which restrict their adaptability in real-world environments where new or unknown modulation types may emerge. In contrast, Automatic Modulation Open-Set Recognition (AMOSR) not only classifies known modulation types but also explicitly recognizes unknown ones, thereby addressing the inherent limitations of closed-set approaches. However, existing AMOSR methods face challenges in achieving a balanced optimization between the empirical risk for known modulation types and the open space risk for unknown modulation types. This issue results in two adverse effects: (1) low rejection accuracy for unknown modulation signals, and (2) degraded classification performance on known types due to overfitting to the closed-set training data. To address these issues, we propose a multi-level feature distribution learning (MLFDL) method for AMOSR, which integrates a pseudo modulation placeholder method with feature distribution constraints. To simulate unknown modulation types during training, we introduce a pseudo modulation placeholder via manifold mixup, which constructs synthetic samples in the feature space to approximate the behavior of unseen modulations. We design a multi-level feature distribution constraint method, combining sample-centroid contrastive learning with a max-min feature constraint, ensuring adequate feature space for unknown modulation signals and enhancing the separation between known and unknown signals. Comprehensive experiments across different datasets show that MLFDL achieves state-of-the-art performance on both AUROC and OSCR metrics, with minimum sample-to-centroid distance serving as a robust decision criterion for unknown signal recognition.