Multi-carrier Signal Detection using Convolutional Neural Networks

For efficient spectrum sharing between non-cooperating networks a fast spectrum scan must be implemented. Frequency, power, bandwidth and modulation have to be quickly estimated to adapt to the environment and cause minimal interference for other users even when protocol is not known. Here we propose to apply convolutional neural network for multi-carrier signal detection and classification as it can measure all these parameters from one short data sample. For the classification and detection tasks, six multi-carrier signal modulations were generated. We have measured detection probability and classification accuracy over wide range of signal-to-noise ratios and have estimated the hardware resources needed for the task. In addition, we have studied impact of signal augmentation during training phase on classification accuracy when only portion of the signal is available. We show that signal four times shorter than 5G radio subframe can be sufficient for the task.