XAI - Core Anomaly Detection

Download Intro_XAI_Core_Anomaly_Detection

The Recommended Publication for Citing

Hanjoo Kim, Sang-Rae Moon, Deokjung Lee*, “A Comparative Analysis of eXplainable AI Techniques for Nuclear Reactor Core Anomalies,” RPHA 2023, Gyeongju, South Korea, October 24-26 (2023). download paper
Hanjoo Kim, Sang-Rae Moon, Deokjung Lee*, “Feasibility Study of an Explainable AI-based Anomaly Detection for Nuclear Reactor Core Operation in PWRs,” KNS Spring Meeting, Jeju, South Korea, May 18-19 (2023). download paper
Hanjoo Kim, Yugwon Jo, Deokjung Lee*, “Feasibility Study of AI-based Prediction for CRUD Induced Power Shift in PWR,” KNS Autumn, Korea (hybrid on-offline), October 20-22 (2021). download paper

Introduction

This study aims to enhance the reliability of AI-based early diagnosis technology for nuclear core anomalies by incorporating explainable AI (XAI) techniques, which were developed in prior research. By doing so, the goal is to increase the practical applicability of AI-based early diagnosis technology for nuclear core anomalies.

Traditional AI, driven by machine learning (ML), achieves high-level decision-making from input data but has limitations due to the complex neural network structures that make interpreting results difficult. This lack of transparency, reliability, and fairness in AI decisions limits its use in critical decision-making areas such as healthcare, law, and finance.
XAI is a technology that enables humans to understand the rationale behind AI decisions, allowing for human involvement in parts of the decision-making process to improve trust in the outcomes.
In preceding research, an AI technology was developed to diagnose nuclear core anomalies early by analyzing operational data in real-time. The efforts included establishing a system for generating AI training data, producing training data, training the AI, optimizing AI models, and developing a diagnostic GUI for nuclear core anomalies, all of which received high marks in evaluations.
The AI model developed in prior research demonstrated high diagnostic accuracy for nuclear core anomalies, but it has a limitation in that it does not provide the basis for its judgments.
To develop a reliable and practically applicable AI-based early diagnosis technology for nuclear core anomalies, this study will address and improve limitations from prior research, such as data imbalance. It will include data collection for XAI development, production of XAI training data for early anomaly detection, implementation and validation of XAI technology, alignment with technical guidelines on nuclear core anomalies, and development of a GUI to enhance usability and visualization.

Key Anomalies

Axial Offset Anomaly (AOA) caused by CRUD (CIPS)
Unintended misalignment of control rod positions
Incorrect indication from Control Element Assembly Calculator (CEAC)
Signal errors due to in-core instrument malfunction
Signal errors resulting from cross-installed in-core instruments at the beginning of a cycle
Unintentional Power Distribution Indication Limiter (PDIL) violations during power control
Asymmetric coolant temperature inflow due to coolant pump anomalies

Methods

Generation of RAST-K data simulating anomalous situations, incorporating actual operational history
Implementation and training of AI models considering the specific characteristics of core anomalies
Development of GUI-connected interfaces for the trained model in collaboration with partner institutions
Visualization of diagnostic results from the model
Visualization of the model's diagnostic rationale and process using XAI

Brief Results

AOA early prediction