A hybrid convolutional neural network with fusion of handcrafted and deep features for FHSS signals classification

Abstract

Shared spectrum utilization is unavoidable because of the continuous rise of wireless usage and bandwidth needs. Effective spectrum sharing can be done by spectrum monitoring that involves detection, parameter estimation, and classification of signals of interest. Signal classification becomes challenging by frequency-hopping spread spectrum (FHSS), which outspread the signal across a vast bandwidth while the carrier frequencies are switched swiftly under a pseudorandom number. Interference from background signals with additive white Gaussian noise complicates classification even further. A hybrid convolutional neural network (HCNN) system with the fusion of handcrafted and deep features is developed in this paper for the FHSS signals classification in the occurrence of the former and latter. The CNN is used as a deep feature extractor by transforming the intermediate frequency signal to the time–frequency representation and used as a two-dimensional input image, whereas the three-layer fully connected network is used as a classifier. The issue of an imbalanced dataset occurred due to unequal observations among classes, which is resolved by performing the random erasing (RE) and synthetic minority oversampling technique (SMOTE). Monte Carlo simulation is performed to verify the performance of the CNN and HCNN. The signal-to-noise ratio (SNR) ranges at 90% probability of correct classification (PCC) for the former and latter with balanced datasets: - 0.18 to 1.4 dB and - 1.58 to - 0.66 dB, respectively. Consequently, the HCNN-RE-SMOTE outperformed the CNN-RE by 1.4 to 2.06 dB of SNR.