Projeto de classificadores para sistema de reconhecimento automático de modulações

Abstract

Underutilization of the frequency spectrum is a recurring problem today, and with the increasing demand of users using remote communication systems, it was necessary to search for a more efficient way to allocate users in the spectrum, appearing thus, the techniques that apply cognitive radio. Cognitive radio detects spectral holes and dynamically allocates users in these unused spaces. With this in view, automatic modulation classification techniques have come to provide a priori information that aid in spectrum sensing. In this dissertation it is proposed to classify modulated signals using a range of supervised multiclass classifiers based on machine learning and deep learning, with its pre-established parameters. Among the classifiers encompassed in machine learning, we approach algorithms based on decision tree and probabilistic classification algorithm, Naive Bayes. Within deep learning, artificial neural networks were applied through a multilayer perceptron network fully connected with backpropagation using Levenberg-Marquardt algorithm to update the network weights. Accuracy rates of 95.2866% and 93.1253% were obtained in the decision tree-based classifiers, 87.4% in the neural network, and 74.7845% in the Naive Bayes In the literature, we found a study with a similar database qualitatively to that used in this dissertation and its accuracy was 89.72%, while the best accuracy presented in this dissertation was 95.2866%.