Comparison of Deep Architectures for Indoor RF Signal Classification

Abstract

In this paper, we study the performance of three different Deep Learning (DL) network architectures in Radio Frequency (RF) signal classification tasks considering an indoor environment. We compare the classification accuracy of 7 modulation types (BPSK, QPSK, GMSK, 16-QAM, 64-QAM, GFSK, and CPFSK) with Convolutional Neural Network (CNN), Long Short Term Memory (LSTM) network, and Residual Network (ResNet) architectures by varying receiver positions in a building layout along with two different transmitter positions. It is seen that, in the considered scenario, for a given transmitter position, CNN and LSTM architectures provide better classification accuracy depending on the receiver positions. It is also seen that in certain receiver positions, some of the modulation types perform better than others.