Automatic Modulation Classification Based on Multimodal Coordinated Integration Architecture and Feature Fusion

With the rapid advancement of the 5G wireless communication technology, automatic modulation classification (AMC) is not only faced with more complex communication environment, but also needs to deal with more modulation styles, which increases the difficulty of modulation recognition invisibly. However, most deep learning (DL)-based AMC approaches currently merely use time domain or frequency domain monomodal information and ignore the complementarities between multimodal information. To address the issue, we exploit a signal statistical graph domain-I/Q waveform domain multimodal fusion (SIMF) method to achieve AMC based on AlexNet, complex-valued networks and multimodal technology. The extracted multimodal features from signal statistical graph domain and I/Q waveform domain are fused to obtain a richer joint feature representation and T-SNE algorithm is used to visualize the extracted feature. Moreover, coordinated integration architecture was adopted to achieve mutual collaboration and constraints between multiple modalities to maintaining the unique characteristics and exclusivity of each modal. Simulation results demonstrate the superior performance of our proposed SIMF method compared with unimodal model and feature fusion model.