Resolution of an Inventory Difference Loss Trend Caused by Holdup Non-Destructive Assay

In 2016, an effort was undertaken to address a cumulative inventory difference (ID, also known as material unaccounted for or MUF) loss trend. It was caused by a measurement bias involving high-efficiency particulate air (HEPA) filter houses of various sizes, shapes, and configurations measured in-situ by non-destructive assay (NDA) for annual inventories. The overestimation by hundreds of kilograms of uranium per year led to a loss trend spanning over a decade, which caused adverse regulatory and financial impacts. First, a capability analysis was employed to investigate the uranium accountancy uncertainty contributors plant-wide. Secondly, a change to the statistical approach, rather than a change to the NDA method, was used to improve the holdup estimation method and confirm the finding. This involved focusing maintenance and operations resources on de-inventorying these filter houses sequentially after in-situ measurements were taken to establish an input-output relationship. The resulting continuous/categorical ordinary least squares (OLS) linear model accurately predicted subsequent filter house cleanout values with high precision and accuracy prior to the cleanout of the last filter house in each data set. The statistical NDA calibration reduced the estimated relative uncertainty of the in-situ method from ~70% to ~1%. Lastly, the uranium IDs have since been analyzed and demonstrate a healthy pattern. This paper discusses how the loss trend was identified, how the holdup measurement method’s statistical calibration was improved, and how the trend was verified to be resolved.Supplemental InformationThis occurred at Westinghouse’s low enriched uranium Columbia Fuel Fabrication Facility (CFFF). This holdup method was greatly improved without the need for NDA expertise or sophisticated computer modeling of the equipment being measured. It simply took a coordinated effort during normal operations to develop input-output data sets sufficient to investigate major categorical differences in equipment populations, develop the OLS model on a spreadsheet, and test each new output’s contribution to model predictability and uncertainty. This potentially enables personnel less conversant in NDA techniques to challenge assumptions and improve methods in lieu of more expensive alternatives.