Reliable and Explainable Data-driven Methods for ELINT Signal Detection/Classification - RAS Proposal

The REDDMEAT (REDDME Assessment and Test) SBIR develops data-driven signal processing methodologies providing explainability and interpretability of Machine Learning (ML) model results. Intercepted/incoming data statistics at various ML network layers are computed and compared to ML results. This enables robust algorithm development where assumptions and thresholds can be quantified and verified.In Phase I, the RAS and MTRI team develop the explainable methods and tests with existing representative ML models (e.g. Intentional Modulation Recognition and Low Probability of Intercept (LPI) characterization) as might be used in EW/ELINT systems. Effectiveness is assessed using synthetic data-sets (with known a priori statistics for evaluation and comparing input data-driven probability densities with ML results.Phase II refines the approach with the Phase I model, and extends it to other ELINT tasks such as characterizing and classifying PRI, Scan, and RF Agility Types and in signal identification. The prototype will be demonstrated, and its performance quantitatively assessed relative to other approaches.  An embedded computing approach for tactical timeline requirements is developed.Multiple Phase III applications are envisioned across the entire DoD enterprise to improve characterization, identification, and tracking in airborne radar and ELINT and Electronic Attack.  Explainable models also have numerous commercial ML applications.