A Novel and Effective Model for Automatic Modulation Classification Prediction Based on Multi-BIGRU, Multi-Encoder, and Hyper-Cross

Automatic Modulation Classification (AMC) is a pivotal technology in various communication systems. In recent years, deep learning (DL) has been widely applied in AMC methods due to its powerful feature extraction capabilities. However, currently proposed AMC methods still have room for improvement in classification performance. To further enhance the prediction accuracy of AMC, we propose a new model called MMH-AMC: Automatic Modulation Classification with Multi-BIGRU, Multi-Encoder, and Hyber-Cross. Multi-BIGRU employs a multi-layer bidirectional GRU architecture to extract deep distribution patterns of AP data from both directions simultaneously. Multi-Encoder adopts a multi-layer encoder architecture with a core of Multi-Head Attention, aiming to extract deep-level IQ data distributions. To unify the feature distributions outputted by Multi-Encoder and Hyber-Cross modules, we designed and utilized the Hyber-Cross module. To validate the model’s performance, we compared six different deep learning models and achieved the best performance in various scenarios.