| Machine learning and AI are concepts that have had a large impact in daily life since 2000. It is unlikely that most people at this point in time do not have some sort of interaction with an AI system on a daily basis. This research effort looked to contribute to the field of nuclear safety and explore ways to expand the use of machine learning through the ap-plication of AutoML. This project consisted of four major phases. In the first phase, data was collected from a GPWR simulator for five different reactor events, creating a dataset with over 30,000 points. Six different machine learning models were trained using theAutoML package TPOT. The results from this test were positive with all models produc-ing accuracies in the high 90% range. The models were also able to perfectly distinguisha reactor operating normally from one experiencing a transient. In the next phase, thedataset was expanded using the GPWR, the number of classes was increased to 12 andthe new dataset consisted of over 110,000 points. Models were retrained and while manyof models suffered in validation, three of the models were still able to score results in thelow 90% range. The models were then examined looking at model redundancy by drop-ping key features, examine variation due to changes in random state, exploring ways to improve the model and identify the reasons behind misclassifications. The third phase of the project explored the use of autoencoders to identify GPWR data that had been al-tered. The model was able to identify all points at high levels of noise, but performance dropped off as the noise was decreased. Still, the technique has validity to help with se-curity concerns and identify sensor malfunctions. The final phase of the project was to explore different AutoML approaches and compare and contrast their performance, easeof use and functionality. These were TPOT, H2O and Google Cloud AutoML. Each of these approaches were found to have different advantages and issues, but all performed with models produced using GPWR data, with results in the mid to high 90% range.xvii
Keywords: Machine Learning, Nuclear Safety, AutoML, Anomaly Detection, NuclearSimulation, Data Science |