Update 2020-CoDA/Abtract.pdf, 2020-CoDA/README.md files

I presented a poster titled "Classification of Dissolution Events at Radiochemical Processing Facility using Effluents Measurements".

Additionally a related journal paper is in the works and this section will be updated when the paper is published.

## Abstract:

We address the problem of inferring dissolution events at a nuclear processing facility during a single isotope production campaign using radiation measurements of effluents at an off-gas stack. We utilize datasets collected at the Oak Ridge National Laboratory's Radiochemical Engineering Development Center (REDC), which is a multipurpose radiochemical processing facility involved in the production of a variety of radioisotopes. The abatement system of REDC is instrumented with a high purity germanium (HPGe) detector that records gamma spectra of the effluents. Using the spectra, activity levels of 15 radionuclides, including isotopes krypton, xeon, and iodine, are computed. The time series of these isotopic measurements along with Boolean labeling of plutonium-dissolution events at 1 hour periods are used to design, train, and test a suite of classifiers and fusers. 

We utilize a diverse set of classifiers, namely, ensemble of trees, support vector machine, naive Bayes, linear discriminant, and k-nearest neighbor that are based on different design principles which results in varying prediction performance and biases. To mitigate this effect, we utilize classifier fusers or meta learners - such as Chow's rule, meta ensemble, and meta linear discriminant - to combine the outputs of individual classifiers, which generally yielded performance as good as the best performing individual classifier.

Furthermore, to improve the classification performance, we pre-process the isotopic measurements using two transformations prior to training: a convolution operator and a moving average filter. A convolution filter using the Hilbert transform potentially captures the effects of non-stationary structural components that are mechanically or pneumatically coupled to the measurement system, which in turn are reflected in the frequency domain.The moving average filter exploits the nature of a Poisson distributed random processes; appropriately long time windows reduce the uncertainty of measurements, and thus improve the classifier performance, since our classifiers do not explicitly consider the measurement uncertainty. Resulting from an analysis of the respective decay chains and isotopic half-lives, a measurement window of 24 hours is identified. 

In general, classifier performance improved with filtering and classifier fusion. Specifically, the Hilbert transform improved the area under the receiver operating characteristic curve (AUROC) for the meta linear discriminant classifier by 9.4% relative to the original AUROC. The best individual classifier, k-nearest neighbor, with the 24 hour average filter resulted in 24.6% improvement relative to the unfiltered case. This result is further improved by the meta linear discriminant classifier which improved relative to the unfiltered case by 26.2% with a final AUROC of 99.18%.