A Deep Learning-based Signal Classification Approach for Spectrum Sensing using Long Short-Term Memory (LSTM) Networks

Signal detection and identification plays an important role in cognitive radio (CR) and spectrum sensing applications. This paper proposes a deep learning (DL) technique for signal classification to identify the wireless signals in a certain radio frequency (RF) environment. The proposed approach is based on a long short-term memory (LSTM) neural network which can classify different radio signals based on their modulation type and pulse shape. We evaluate the classification accuracy of the LSTM network by using raw data and cyclostationary features as an input function. We show that the classification accuracy of the raw data-based LSTM classifier outperforms the one of the cyclostationary-based classifier even under low signal-to-noise ratio (SNR) conditions. The proposed raw data-based approach can classify 8 types of signals with an accuracy of more than 86% at an SNR ≥ 0dB.