Complex-valued Neural Networks for Classifying RF Wireless Signals

Complex numbers are common in science and engineering fields, such as state vectors in Quantum Physics and radio waves in Electrical Engineering. In this work, we investigate the value and effectiveness of applying Complex-Valued Neural Networks (CVNNs) to the task of modulation recognition and classification of radio frequency (RF) wireless signals. The results of an extensive experimental evaluation show that CVNNs consistently outperform real-valued Neural Networks under different settings (distorted or undistorted, overlapped or non-overlapped) and different cross-validation strategies (temporal or non-temporal). We also investigate how CVNNs are affected by distortions caused by changing wireless environment. Our experiments demonstrate that Complex-Valued Neural Networks have a more robust and stable accuracy curve when these distortions occur.

Co Advisor: Hamdaoui, Bechir
Co Advisor: Weng-Keen Wong
Committee: Fu, Xiao
Committee: Lee, Stefan
GCR: Bolte, John