Joint use of Bivariate Empirical Mode Decomposition and Convolutional Neural Networks for Automatic Modulation Recognition

The AMR (Automatic Modulation Recognition) domain has recently shown an increase of interest, particularly as an application for monitoring the physical layer of wireless transmissions. In this work, it is proposed to study the combination of BIEMD (BivariatelEmpirical Mode Decomposition) and CNN (Convolutional Neural Network) in order to improve the modulated signals recognition rates. On the one hand, B/EMD is able to decompose signals into a finite number of IMFs (Intrinsic Mode Functions) whose features can be analyzed. On the other hand, CNNs are already used in the scientific literature to classify modulations. Adopting a publicly available data base containing numerous modulated signals under varying channel conditions, it is possible to decompose the complex signals using B/EMD and inserting them into a CNN. It is shown that the shape of the data has an important impact on the classification. Furthermore, the performance is compared and it is shown that employing the IMFs provided by a BEMD (Bivariate Empirical Mode Decomposition) instead of the original IQ signal improves the overall accuracy of the recognition by 2% in the frame of the parameters used in this paper. More than that, it is also able to increase the recognition rates up to a maximum of 4.4% for hizher sianal to noise ratio values.