Deep Learning Applied to Automatic Modulation Classification at 28 GHz

Automatic Modulation Classification (AMC) is a fast-expanding technology, which is used in software defined radio platforms, particularly relevant to fifth generation and sixth generation technology. Modulation classification as a specific topic in AMC applies Deep Learning (DL) in this work, which contributes a creative way to analyse the signal transmitted in low Signal to Noise Ratio (SNR). We describe a dynamic system for the Millimeter wave (mmW) bands in our work. The signals collected from the receiving system is without phase lock or frequency lock. DL is applied to our system to classify the modulation types within a wide range of SNR. In this system, we provided a method named Graphic Representation of Features (GRF) in order to present the statistical features in a spider graph for DL. The RF modulation is generated by a lab signal generator, sent through antennas and then captured by an RF signal analyser. In the results from the system with the GRF techniques we find an overall classification accuracy of 56% for 0 dB SNR and 67% at 10 dB SNR. Meanwhile the accuracy of a random guess with no classifiers applied is only 25%. The results of the system at 28 GHz are also compared to our previous work at 2 GHz.