Machine learning for dynamic spectrum access in passive radio sensing

About This Grant

This project develops methods and tools that use artificial intelligence and machine learning (AI/ML) to help overcome the problem of radio-frequency interference (RFI) in measurements made by radio telescopes. The techniques could benefit other sensitive receivers threatened by RFI, for example weather radars. RFI is a growing challenge for instruments like telescopes and radars due to increasing usage of the radio spectrum by mobile wireless communications and other applications. The undesired signals from nearby transmitters overpower the faint signals from far away that the instruments intend to measure. To continue performing their missions, these instruments require improved techniques to filter out RFI, and where that is not possible, to identify and discard data that are irretrievably corrupted by RFI. Recent progress in AI/ML offers new approaches that could be applied to RFI mitigation, with the promise of being fast enough to keep up with the high volumes of data that stream from state-of-the-art instruments. This project will deliver documented data sets, open-source AI/ML models, and software tools. The models and tools are intended for use by non-experts in AI/ML to address RFI challenges in their instruments or scientific studies. The project starts by assembling the largest-ever dataset of RFI observations by radio telescopes, through doing an RFI survey at selected US astronomy facilities. The results inform the development of digital twins: data simulators that satisfactorily reproduce RFI-affected data. In parallel, the project designs a large neural network specialized for RFI mitigation in passive sensing data, whether astronomical or otherwise. This effort may include innovation in efficient computer vision algorithms for unsupervised segmentation learning. The network is then trained to serve as a baseline model, akin to a “foundation model” for language processing. The baseline model is fine-tuned through training with real-world data to operate effectively in the interference environment specific to different sites, frequency ranges, and observing modes. The new approach is evaluated through quantitative comparisons against the existing iterative sum/threshold approach, starting with a single data set for simplicity (total-intensity data from the GReX instrument) and progressively growing to more complex cases, e.g. multiple instruments and frequency ranges. Fine-tuned versions of the baseline model are evaluated at multiple telescopes. Computational costs and power budgets are evaluated to determine feasibility of implementation in real-time processing pipelines. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.