Deep Convolutional Transformer Network With Optional CFO Compensation: A Framework for Automatic Modulation Recognition

Automatic modulation recognition (AMR) methods in wireless communication systems can identify the modulation type of received signals without relying on any prior information. However, the feasibility of these methods is limited by parameters, computational complexity, and prior knowledge, while the impact of carrier frequency offset (CFO) on recognition accuracy is often neglected. A Deep Convolutional Transformer Network with Optional CFO Compensation (DCT-OCC) is proposed for the classification of unknown signals. It is a hybrid architecture combining deep one-dimensional convolutional neural networks (CNN) with multi-layer Transformer encoders, which can effectively extract temporal and structural features from raw in-phase (I) and quadrature (Q) signals for modulation recognition tasks. In addition, an optional CFO estimation and compensation module is introduced to address the lack of robustness in practical environments. The experiments are conducted on the RadioML2016.10A and RadioML2018.01A dataset and compared with other methods. The results demonstrate that DCT-OCC achieves superior classification performance compared with traditional deep classification models.