Novel Training Methodology to Enhance Deep Learning Based Modulation Classification
Venkatesh Sathyanarayanan, A.J. Jolly, Peter Gerstoft · 2021 55th Asilomar Conference on Signals, Systems, and Computers · 2021
Automatic Modulation Classification (AMC) is central to dynamic spectrum sensing. This work aims to improve the performance of deep learning (DL) models applied to AMC. Novel training methodology is introduced to improve performance. Over-the-air (OTA) data is collected using software-defined radio (SDR) over a range of modulation types and SNR levels. Collected dataset is partitioned into subsets across SNR levels. A group of identical models is trained on these multiple subsets and performance is compared against model trained on whole dataset. Efforts are taken to identify and isolate corrupted OTA data caused by interferences. Convolutional neural network (CNN) based architecture is used. We show an average improvement of 6% in the classifier’s performance on OTA data using this SNR partitioning approach.