Classificação automática de modulações utilizando deep learning

Abstract

In this work, the generation of a deep learning model capable of predicting about six types of modulation signals was proposed. 8PSK, B-FM, BPSK, DSB-AM, AFSK and DPSK modulations were chosen. To obtain these signals, a system composed of a RTL-SDR to capture real signals and a code in Matlab to generate synthetic signals was used. The signals were generated in complex components, which were processed in order to generate signals in the time domain and thus extract relevant characteristics to the model. The mel-cepstral, pitch and spectral centroid coefficients are used to measure the shape of signals. The architecture used is composed of convolution neural networks, these divided into max-pooling and dropout layers. The result obtained for the model was an accuracy of 98% for the validation set. The precision was 98% to 8PSK, 62% to AFSK, 97% to B-FM, 100% to BPSK, 100% to DSB-AM