An Ensemble based Modulation Recognition using Feature Extraction

Due to the increase in wireless devices, there has been an increasing demand for spectrum utilization and intelligent radios. This is done by sensing the spectrum to collect important information regarding the received signal and is generally used in military applications to identify or decode enemy signals. Software-defined radio (SDR) and Cognitive radio (CR) are important civilian applications for this field. These types of task fall into the broader study of signal recognition, which is quite difficult due to many different parts of the signal to be identified just by using the received signal. In this paper, we explore the viability of different hand crafted features from various domains such as Time, Frequency, and Statistics for the task of Automatic Modulation Classification (AMC). For the classification we are using an ensemble learning approach with Random Forests (RF) model, and comparing it with Gini-based Decision tree (DT) algorithm. We found that the accuracy of RF model is 10% higher than the DT model for medium to high SNR values and also one feature is dominating the prediction in case of DT approach. Using the scikit-learn machine learning library, we are able to get $\approx$75% accuracy without using any complex Deep Learning architectures or expensive hardware.