Uncertainty-Guided Label Rebalancing for CPS Safety Monitoring
Researchers have developed a new method called U-Balance to improve safety monitoring in Cyber-Physical Systems (CPS) by addressing extreme class imbalance in telemetry data. This approach utilizes behavioral uncertainty, which is correlated with safety outcomes, to rebalance datasets. U-Balance trains an uncertainty predictor and then employs an uncertainty-guided label rebalancing mechanism to relabel uncertain 'safe' windows as 'unsafe', thereby enriching the minority class without generating synthetic data. Evaluated on a UAV benchmark, U-Balance achieved a 0.806 F1 score, significantly outperforming existing methods. AI
IMPACT Enhances safety monitoring in critical systems by improving the detection of rare unsafe events.