Classifiers for Dissolution Events in Processing Facility Using Effluents Measurements

Inferring occurrences of dissolutions at a radiochemical processing facility using measurements from an independent monitoring system could be an important part of compliance assessments. We consider on/off classification of such dissolution events using effluents measurements. Recently, there has been an explosive growth in classifier methods applicable to this case, but their vast variety makes it extremely challenging to select suitable ones. Accuracy estimates based on training data may help to identify fewer promising classifiers; however, any individual classifier is subject to over-fitting, and consequently may not provide a strong generalization property, which depends on its underlying function class, for example, its Vapnik-Chervonenkis dimension. For Radiochemical Engineering Development Center (REDC) facility, we overcome this limitation by fusing four promising but disparate classifiers that are trained on measurements collected over four months. We study classifiers using measurements of effluents, Kr-87, Xe-135, Xe-138, and Cs-138, of which first and second provide lowest and highest training accuracy, respectively. Specifically, the Ensemble of Trees (ET) and Support Vector Machine classifiers, respectively based on non-smooth and smooth functions, achieve higher training accuracy. Additionally, we consider the Naive Bayes and k Nearest Neighbor classifiers, which are based on different statistical and structural principles. We study two fusers that combine the outputs of these four classifiers, thereby incorporating their diversity while preserving training accuracy. We implement Chow's Fuser (CF) by utilizing the confusion matrices of classifiers to derive weights and thresholds; it is computationally simple and is optimal under certain statistical independence conditions. We also apply the Ensemble of Trees Fuser (ETF) which satisfies the isolation property, and therefore ensures as at least good performance as its component classifiers. We present training accuracy estimates of CF and ETF applied to classifiers that use individual isotope measurements and jointly Xe-135, Xe-138, and Cs-138 measurements. In our analysis, for example based on 5-fold cross-validation, ETF achieves (i) accuracy 83.51%, which is higher than of best Xe-135 classifier, and (ii) accuracy 82.08% that matches that of best three isotope classifier.