Emerging mobile network technologies such as 5G promise high data rates with increased reliability and low latency. However, the implementation of these technologies requires efficient spectrum management and utilization. Due to an exponential increase in the number of connected devices over the past decade, the spectrum has become a scarce and expensive resource. To increase spectral efficiency, cognitive radios have generated keen interest. To enhance the performance of these devices for signal detection, real-time modulation recognition tasks have to be performed.
In this thesis, we explore deep learning algorithms for Automatic Modulation Classification (AMC) tasks using raw signal data in the form of I/Q samples. A comprehensive analysis of the performance of algorithms under the influence of noise, hardware heterogeneity, Carrier Frequency Offset (CFO), and modulated co-channel interference is performed. Further, a method to explain model predictions is also discussed.
- XGBoost
- CNN
- ResNet-101