Spectrum Sensing for Modulated Radio Signals Using Deep Temporal Convolutional Networks

Amir Ghasemi, Chaitanya Parekh, Paul Guinand

Abstract: Detecting the presence of radio signals in noise has been a decades-long problem in the signal processing domain. More recently, there has been a resurgence of interest in detecting signals with relatively low signal-to-noise ratio in the context of cognitive radios and spectrum sharing. A majority of existing algorithms rely on detection criteria which are carefully-crafted by domain experts to exploit specific features of the target signal. Motivated by the ability of deep neural networks to learn useful representations directly from raw data, in this work we use a deep temporal convolutional network for detection of signals, under multi-path fading and noise, directly from raw complex baseband samples with no other pre-processing or feature engineering.Our results indicate that the proposed deep learning approach outperforms a popular eigenvalue-based method without requiring any expert feature engineering. We further show that the performance is robust to new signal types not used during the training of the neural network.