Data-Driven Versus Period-Driven Change Detection for Process Monitoring

Avenhaus and Jaech’s landmark 1981 paper dampened hopes that frequent nuclear materials accounting, referred to as near-real-time accounting (NRTA), would allow diversion detection probabilities goals to be met. The 1981 paper showed that NRTA leads to lower detection probabilities for some types of protracted diversion, and the inevitable conclusion was that for some diversions “less frequent balance closure is better.” Now in 2012, process monitoring (PM) in some cases leads to very frequent model residual checks, using residuals that arise for example in comparing in-tank volumes to predicted volumes every few minutes. Such model residuals are analogous to material balances in NRTA. This paper re-examines the 1981 paper which invoked period-driven change detection, and argues for instead using data-driven sequential testing applied to residuals from both NRTA and PM. We also show that while frequent sequential testing using NRTA and PM data can meet detection probability goals for some specified scenarios, and is recommended, such frequent testing is not a panacea, but is a reasonable approach. Simple numerical examples are given to contrast period-driven from data-driven change detection probabilities in the case of using NRTA residuals alone, without PM residuals.