Classification of Wireless Digital Modulations under Bad Urban Channel Using Convolutional Neural Networks

Modulation recognition using convolutional neural network (CNN) has received great attraction from academia during recent years. Most of the published papers use an artifact dataset in which signals have been passed through random time off set, scaling, rotation, phase, channel response, and noise effects. Although this kind of choosing parameters represents harsh environment and it can be interpreted as a worst case scenario, it is important to know how much this condition is pessimistic and what the performance of system is in presence of a real channel effects. In this paper, we apply fading channel parameters of bad urban COST 207 [1] on commonly used digital modulation signals and compare our result with that of [2], which we call it OCC model. Then, we use different CNN based architectures for classification as follows, the hybrid CNN-SVM, OCC-FC model and OCC with maxpooling layers. Our result show that using maxpooling can significantly affects classification rate in average and especially low SNRs.

Keywords:

modulation recognition , deep learning , fading , software radio , convolutional neural network