Trace element analysis is now receiving more and more attention in strengthened safeguards to verify the completeness of the State declaration with respect to the origin and production of Uranium materials. For the trace element analysis, samples are usually taken from the feed, intermediate, and product materials at Uranium concentration and/or conversion plants, the trace element compositions are analyzed using advanced techniques such as ICP-MS, and the analytical results are evaluated to assess the consistency with the declared origin of the given Uranium materials.
Statistical data evaluation is one of the most important steps in trace element analysis. Two sequential modules, detection and classification, are applied for the evaluation. The detection module is to determine whether a new sample is significantly different from the historical dataset for the samples collected from the given Uranium material at a facility. If an abnormal sample is detected, the classification module will then be triggered to determine which material in the available reference database has the highest probability to belong to.
Linear techniques have been pursued for the statistical data evaluation. Principal Component Analysis (PCA) is used as a dimensionality reduction technique in the detection module. PCA is optimal in dimensionality reduction in terms of capturing the variance of the data while the correlations among trace elements are taken into account. The decision making for detection is based on T2 statistics to determine whether a new sample is out-of-control in terms of systematic variation. In the classification module, Partial Least Square (PLS) is used as a modelling tool to predict the class memberships of a new sample. By using data decomposition method for maximizing the covariance between the input and the output, the PLS model developed based on historical dataset is able to capture the relationship between the trace element compositions and the class memberships for different origins.
The performance of the procedure described above is demonstrated using real data. In addition, special difficulties encountered in the implementation are also discussed, which include statistical treatment of missing values, results reported as less than detection limits, and removal of susceptible results due to poor quality in the sampling or measurement process. It is recommended that new statistical methods be studied such that uncertainties of trace element results reported by analytical laboratories can be effectively used in the statistical evaluation.