Signal Analyzing Network Tool using Deep Learning

Deep learning can identify different signals and extract a range of useful features or track a signal source. Semi and selfsupervised learning techniques can be used to teach networks the underlying dynamics of a problem and broaden generalizability. We demonstrate preliminary results on machine learning software capable of identifying the source of a target and extracting key pieces of information to help resolve or identify the source including angle of arrival. A Ushaped convolutional network may be trained to classify signals based on IQ samples according to modulations or other select features while reconstructing the clean signal. Use of semi-supervised learning training schedule including Barlow Twins on the generated latent space was demonstrated on combinations of real and synthetic radiofrequency (RF) signals. These signals were augmented under various common signal obfuscations such as Raleigh fading, reflections, varying noise and background signals. Group structure of the signals may be displayed through latent space visualizations. Classification accuracy on unseen test sets was used as the primary measurement of performance under varying levels of obfuscation. From this base, we attempted to combine this network with directional sensitivity in order to enable beam steering or identifying the source. A similar augmentation route enhanced by similar semi and selfsupervised techniques was deployed to improve tracking accuracy under realistic conditions. Statistical techniques may be used to identify frequency regions of interest during the prototyping of this signal identification network. This Deep network framework may be applied across a variety of domains and regimes for sensing and tracking.