7 March 2025 in IEEE Transactions on Aerospace and Electronic Systems
In recent years, with the development of deep learning, deep learning (DL)-based specific emitter identification (SEI) has outperformed the traditional methods in accuracy. However, DL-based methods often depend on huge labeled samples for training, which are not available in certain situations. Semi-supervised methods can be trained with a large number of unlabeled samples and few labeled samples, thereby reducing the need for manual labeling signals. In this article, a semi-supervised SEI method based on contrastive learning and data augmentation, SSMPCL-SEI, is proposed. SSMPCL-SEI modifies the self-supervised contrastive learning to cope with multiple positive samples and learns category information from both labeled and unlabeled samples. Furthermore, data augmentation for the classifier is used after pre-training to improve the classification ability. The experimental results with a UAV dataset and an ADS-B dataset show that SSMPCL-SEI yields better identification performance than other state-of-the-art semi-supervised SEI methods.
1 January 2026 in IET Radar, Sonar & Navigation
In modern electromagnetic reconnaissance, the ability to recognise intra‐pulse modulation patterns in complex environments has become increasingly critical. Traditional recognition algorithms are typically designed under ideal conditions and handcrafted features, and the presence of substantial noise interference in complex electromagnetic conditions can significantly degrade the performance of these algorithms. Although transformers have achieved great success in computer vision tasks, existing vision transformers lack the ability to capture long‐range dependencies. To remedy these flaws, this paper proposes an accurate automatic modulation recognition algorithm based on the vision transformer. First, we utilise traditional image processing techniques to provide the position of signals in the time frequency image and introduce a radial basis function to generate the positional embeddings. This procedure enhances the spatial awareness of the model. Second, dynamic scales of 2D patches are embedded as inputs to the encoders. Furthermore, inspired by dilated convolutions, vanilla attention is reformed in the cascade encoders to promote the training and inference procedures. Experiments and ablation studies exhibit the proposed algorithm that performs well across various signal‐to‐noise ratio conditions, demonstrating its robustness and accuracy in modulation recognition.