Machine Learning-Based Frequency Bands Classification for Efficient Frequency Hopping Spread Spectrum Applications.

This paper is focused on the performance evaluation of nine supervised machine learning algorithms in terms of classification accuracy applied to perform two radio scene analysis tasks: 1. blind binary frequency band occupancy classification: vacant or occupied; 2. interference type classification: sine wave interference, or modulated signal or additive white Gaussian noise (AWGN) for the frequency hopping spread spectrum cognitive radio application. Twenty-nine features derived from the time-, frequency-domain and RSSI, have been used as classification inputs to the evaluated machine learning classifiers. Classifiers training and validation have been performed offline in Matlab Classification Learner and Neural Networks applications using four data sets, generated in the controlled experiment, covering both classification tasks in AWGN and mixed channel propagation conditions (AWGN and Rician fading). Data samples have been generated using a hardware signal generator and recorded on the target application receivers' front end as the time-domain complex signals. The highest classification accuracy of 98.71 % has been demonstrated by Feed Forward Neural Network (FFNN) for the binary occupancy classification in Kfold validation for the mixed data set containing both AWGN and Rician fading channel samples. For the interference type classification, FFNN has demonstrated classification accuracy of 99.82 % for K-fold validation and 99.71 % for hold-out validation. FFNN has been concluded as an acceptable algorithm for further adaptation and embedded deployment on our target radio application for both binary classification between occupied or vacant frequency bands and interference type classification.