Small Sample Wireless Technology Identification Enhanced by Cyclic Shift and Attention Mechanism

Small sample wireless technology identification has significant application value in complex electromagnetic environments. However, the generalization ability and recognition accuracy of models are often limited due to insufficient data. This paper proposes a small sample wireless technology identification method enhanced by cyclic shift and attention mechanism. The method is applied to a six-class classification task involving CV2X, Five_G, ITSG5, LTE, Noise, and WiFi signal categories. By incorporating cyclic shift data augmentation during the training phase, the model's robustness to time-series data is improved. Meanwhile, the SE attention mechanism is employed to optimize the feature space and strengthen the representation of key features. Experiments were conducted with different sample sizes (100, 200, 300, 400, 500), comparing the testing accuracy, confusion matrices, and feature distributions of the models before and after optimization. Results demonstrate that the proposed method significantly enhances the classification performance of the model under small sample conditions. Compared to the initial model, the optimized model achieves higher testing accuracy across all sample sizes, with more concentrated feature distributions and markedly improved category separability. Furthermore, confusion matrices reveal that the optimized model shows a particularly notable improvement in classifying small sample categories, with a significant reduction in misclassification rates. This method provides an efficient solution to the challenges of small sample wireless technology identification.