Abstract
In this work, we contribute to the emerging field of deep learning (DL) methods of automatic modulation classification (AMC) for cognitive radios. Traditional AMC methods rely on expert-based knowledge of the wireless channel and incoming signals. This method suffers from a lack of generalizability to real-world channels that may be severely impaired or unknown. DL does not require a priori or expert-based knowledge and has seen success in other fields such as image processing and natural language processing. In recent years, DL has been explored as an alternative to traditional methods; however, currently proposed DL AMC methods suffer from high training times due to the many layers used to improve classification accuracy. We propose the use of induced class hierarchies to decompose the AMC task into subcomponents, while still maintaining deep architectures for improved classification accuracy. A publicly available, synthetic radio data set is used, which models severe channel impairments under a range of various signal-to-noise (SNR) levels. Three hierarchical convolutional neural network (h-CNN) architectures are developed: a single-level, baseline model; a two-level hierarchical model, termed model A; and a three-level hierarchical model, termed model B. Model A achieves a 4% improvement in classification accuracy over the baseline model while model B maintains comparable accuracy. Moreover, the training times of both models are reduced, with 50% improvement with model A and 28.6% improvement with model B, from the baseline model.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have