Performance analysis of deep learning-based automatic modulation recognition over wireless communication

Abstract

Automatic Modulation Recognition (AMR) based on Deep Learning (DL) is an efficient technique to improve spectrum utilization by replacing the old way of detecting modulation type through the allocation of modulation information in the signal frame. However, DL models have the problem of low recognition accuracy when dealing with a dataset containing in-phase and quadrature channel data. Hence, in this work, the enhancement of DL models that automatically recognize different types of modulation techniques with an increase in recognition accuracy was carried out. The two utilized dataset were RadioML2016.10a and RadioML.2016.10b. Convolutional Neural Network with RadioML2016.10a (ECNN-1) and RadioML2016.10b (ECNN-2) and Long Short-Term Memory with RadioML2016.10a (ELSTM-1) and RadioML2016.10b (ELSTM-2) were implemented in Python 3 using Google Colab. Adam optimizer was applied to optimize the hyperparameters of DL models. ECNN-1 and ECNN-2 have recognition accuracy values of 81% and 88%. The accuracy values obtained for ELSTM-1 and ELSTM-2 were 79% and 85%. The ROC AUC score for the ECNN-1, ECNN-2, ELSTM-1, and ELSTM-2 were 89.63%, 92.90%, 90.92%, and 92.81%, respectively. The experimental results showed an improvement in modulation recognition accuracy for both enhanced CNN and LSTM models.