A Bias-Reducing Framework for Enhancing Trustworthiness of AI-Based Operator Support Models in Nuclear Power Plants

Recent advances in artificial intelligence (AI) have led to increasing research efforts to support human operators in nuclear power plants (NPPs). However, before deploying AI-based systems in actual operational environments, their trustworthiness must be thoroughly verified. Trustworthy AI requires robustness and bias mitigation to ensure reliable decision support. This study focuses on bias mitigation in AI models developed for NPP operator support and proposes a bias-reducing framework to enhance both fairness and reliability. Most existing AI-based operator support systems depend on simulation-generated data to compensate for the limited availability of real operational data. However, even when simulation diversity is considered, the individual characteristics of plant systems and components make it difficult to predict or eliminate potential bias across various operational scenarios. Furthermore, a domain gap between the simulated training domain and the actual operational domain can amplify bias and degrade model generalization performance. To address these challenges, this study proposes a bias-reducing framework that considers a fairness point into the training process of AI-based operator support models for NPPs. The proposed training stage in this framework extends beyond the main task structure to include the sub-task structure designed for domain adaptation and bias mitigation strategies in the neural network. Through this approach, the study establishes a methodological foundation for developing trustworthy AI-based operator support models that can reliably support operator decision-making in safety-critical environments such as NPP operations.

Keywords: Trustworthy AI, bias mitigation, adversarial training, nuclear power plant, operator support system.