Wireless Standard Classification Using Convolutional Neural Networks
Samuel R. Shebert, Anthony F. Martone, Richard Michael Buehrer · 2021 IEEE Global Communications Conference (GLOBECOM) · 2021
The growing prominence of spectrum sharing technologies has spurred interest in spectrum monitoring technologies with the ability to identify unknown wireless signals. This paper presents a convolutional neural network (CNN) deep learning model to classify 4G LTE downlink, 4G LTE uplink, 5G NR downlink, 5G NR uplink, IEEE 802.11ax (WiFi 6), and Bluetooth Low Energy (BLE) 5.0 signals. The classifier operates on In-phase and Quadrature (I/Q) samples and does not require synchronization with the unknown signals. To improve the generalizability of the classifier, comprehensive signal datasets are generated to include a wide range of signal configurations found in the standards. These signals are impaired with additive white Gaussian noise (AWGN), Rayleigh or Ricean multipath fading channels, frequency offsets, and I/Q imbalances to make the signals more realistic. The exploration of time domain, frequency domain, and time-frequency domain features reveals high frequency resolution time-frequency domain features perform best. The proposed CNN model achieves a high classification accuracy in the presence of all of the aforementioned impairments, achieving over 94% accuracy for signal to noise ratios (SNR) greater than 0 dB.