A Deep Learning Framework for Blind Time-Frequency Localization in Wideband Systems

Abstract

In this paper, we propose a blind timefrequency localization method for wireless signals present in a wideband radio frequency (RF) spectrum. The signal detection problem is transformed into an object detection problem by converting the RF time-series captures into spectrogram images. A deep learning system based on the Faster RCNN [2] is then configured to suit the signal detection task. Guidelines are provided to make design choices in terms of both data pre-processing and the FRCNN modeling, for example, on the Short Time Fourier Transform (STFT) parameters, the spectrogram sizes, and the anchor sizes. Experiments with artificially generated WiFi high throughput data [3] reveal that (i) the proposed framework can achieve up to a mean average precision (mAP) of 0.9 for captures with positive signal-to-noise ratio (SNR), (ii) the proposed framework is fairly robust to the number and size of the anchors, and (iii) the proposed framework is sensitive to the disparity in the signal sizes, giving us few insights into possible future extensions.