Spectrum Monitoring Based on End-to-End Learning by Deep Learning

Numerous autonomous wireless deployments have become invaluable for understanding and investigating the radio frequency environment. However, machine learning techniques have their drawbacks and there are situations where such strategies are unreliable. The purpose of the present paper is to present an end-to-end learning framework based on deep learning (DL) and to evaluate different methods of wireless signal classifiers implementation and signal representation for spectrum monitoring. Furthermore, we tend to investigate the significance of wireless data representation selection for varied spectrum monitoring tasks. For each case study, modulation recognition (MR) and wireless interference identification (IId), three deep learning networks are evaluated for the subsequent wireless signal representations, temporal I/Q data, the amplitude/phase, frequency domain and Hilbert and wavelet transform representations. From our analysis, the accuracy of wireless signal identification is proved to be affected by the network classifier and wireless data representation. For different signal-to-noise ratio values, the classification accuracy of the three DL networks are evaluated. The results of the experiments indicate that the representation of data influences network accuracy. In MR case, in high SNR (18), the first, second and third networks have the best results in the db3 mother wavelet, amplitude/phase and Hilbert samples, respectively. In the medium and low SNR (0, − 8) in all three networks, almost the best results is obtained from Hilbert data representation with the accuracy variation up to 4%. In IId case, for three SNR (− 8, 0, 18) in the three presented networks almost the best results is obtained from the FFT and wavelet data representations with 0.5% accuracy variations.