An adaptive neural network approach for automatic modulation recognition

This paper presents a novel adaptive approach to automatic modulation recognition (AMR) using artificial neural networks (ANN). Two types of features have been combined to ensure the robustness of detection for all types of modulations. An adaptive system of feature extraction and neural network is designed to distinguish the target modulation between 5 different modulation types namely AM, ASK, FSK, PSK and QAM, distinctly. In a novel approach by introducing modulation pairs, the decision is made based on the majority of each individual output of the 10 minimally complex - binary ANNs which have been trained using sufficient and wide range of modulation parameters for each pair of modulation types. We have made safe error criterions to indicate the performance of our design, and simulation results express the performance and accuracy of proposed system model according to detection error versus signal level, and detection error versus number of iterations through convergence.