New Method Improves Changepoint and Root Cause Analysis with Corrupted Data

Monitoring complex engineered systems requires robust methods to detect when their statistical behavior shifts and to pinpoint the exact component responsible for the change. Traditional methods often struggle when data is corrupted by noise or outliers, leading to overly broad or uninformative results, even if statistical guarantees are technically preserved. This new research proposes Weighted CONCH (W-CONCH) and Weighted CROC (W-CROC) to address this challenge. These techniques build upon existing conformal methods by introducing a weighting mechanism that downplays the influence of potentially corrupted observations. By integrating uncertainty signals, such as those from evidential deep learning, the approach significantly narrows the confidence sets for changepoint localization and root cause identification, providing more precise insights without sacrificing statistical reliability.