Left to right: Abhinav Gupta, Maria Avramova, Igor Bolotnov, Joomyung Lee, Nam Dinh, and Kostadin Ivanov

Joomyung Lee successfully defends dissertation

On March 4, Joomyung Lee successfully defended his PhD dissertation, Development of the Machine Learning Based Safety Significant Factor Inference Model for Advanced Diagnosis System. Joomyung ’s committee consisted of his advisor, Nam Dinh, and members, Abhinav Gupta, Igor Bolotnov, Kostadin Ivanov, and Maria Avramova.

Abstract

LEE, JOOMYUNG. Development of the Machine Learning Based Safety Significant Factor Inference Model for Advanced Diagnosis System (Under the direction of Dr. Nam Dinh)

Diagnosis is the process of identifying the nature of abnormal occurrences to provide the information for predicting the consequences in the operation of the management and control system. The Nearly Autonomous Management and Control (NAMAC) system is an integrated system that delivers recommendations to the operators to make a decision through diagnosis, prognosis, and strategy assessment based on the Artificial Intelligence (AI) technical support for accident management. The AI-guided diagnosis aims to identify the plant damage states during abnormal states or accident conditions by using the Machine Learning (ML) algorithm, which is trained with the data from the best-estimate simulation code, as the primary stage in the NAMAC system.

In this dissertation, the ML-based Safety Significant Factor Inference Model (SSFIM) is developed to infer the physical plant damage states by training the Recurrent Neural Network (RNN) during transient in Experimental Breeder Reactor II (EBR-II). To generate the training dataset for the ML algorithm, GOTHIC simulates the transient resulting from loss-of-flow (LOF) accident owing to the pump operation status in a simplified EBR-II reactor. The Safety Significant Factor (SSF), a surrogate index used to represent the physical damage states during a transient, is assumed to a fuel centerline temperature that is a directly risk-related variable. The objective of the ML-based SSFIM is accurately inferring the SSF in real time from the measured physical variables and its model performance should be as good as a thermocouple’s, even though the measurement data includes the random noise.

The RNN is confirmed as an appropriate algorithm to deal with time dependent physical phenomena in the reactor system. Based on the RNN with Reduced Learning Rates on Plataea (RLRP) scheme, the ML-based SSFIM is well trained to infer the fuel centerline temperature within 1% accuracy, even if there is 5% random noise in the mass flow rate data. The RNN proves its capability by allowing the ML-based inference model to extend the information range in the aspect of both data amount and time. The adapted ML-based SSFIM is able to map the temperature distribution in the reactor system from only one mass flow rate variable and become the faster-than-real-time inference model by predicting the one second later SSF from the current measurement. Also, the ML-based SSFIM shows good inference ability through the extrapolation test in which it covers ±20% of pump operation status even though the transient scenarios are out of range of the training dataset, however, the model performance becomes poor when the extrapolated range in testing is far from the training dataset. To compensate for the weakness in diagnosis, a large amount of training data with a wide range of scenarios is required for training the model. In conclusion, since the ML-based model is developed by not only the AI techniques but also the best-estimate code simulation result, the ML-based SSFIM is completed as a knowledge-based data-drive model.

The proposed developmental workflow and developed robust SSFIM contribute to the design of the AI-guided diagnosis model in the NAMAC system. To improve the SSFIM’s model performance, there are several suggestions: 1) reduction of the sensor noise, 2) subdivision of the physical variables in the normalization process to train the model, 3) head event study to generate a large amount of data with a wide range and various types, and 4) identification of the head event. Then, the ML-based SSFIM is expected to be utilized to develop the advanced AI-guided diagnosis model in various ways.