Classification intelligente des modulations radio par Deep Learning

Abstract

Automatic Modulation Classification (AMC) is a key technology for many modern wireless applications, such as cognitive radio, spectrum monitoring, and intelligent networks. This thesis presents the study and development of a hybrid Deep Learning architecture designed to automatically identify the modulation type of received radio signals without prior information. The proposed model combines Convolutional Neural Network (CNN) layers for spatial feature extraction with a Transformer, a Bidirectional Gated Recurrent Unit (BiGRU), and an attention mechanism to efficiently capture long- term sequential dependencies in I/Q signals. The approach is validated using the benchmark RadioML 2016.10a dataset, which includes 11 classes of analog and digital modulations. This architecture aims to improve the robustness and accuracy of modulation recognition in complex transmission environments affected by noise, frequency offsets, and channel impairments.