Development of a Confidence Estimation based Domain-Adaptive Industrial AI Framework

Abstract

This study presents CONDA-AI, a confidence-estimation–driven domain-adaptive industrial AI framework designed to maintain reliable decision-making under complex distribution shifts. Unlike conventional industrial predictors that treat confidence as a post-hoc statistic, CONDA-AI formulates confidence as a control variable that determines whether the system should act autonomously, defer decisions, or trigger conservative mitigation under risk. The framework separates industrial shifts into long time-constant domain drift and short time-constant environmental volatility and addresses them through two complementary modules: Module 1 constructs a domain-stable feature space with structured confidence to prevent confident-but-wrong behavior under drift, while Module 2 produces robust, physically meaningful indicators and prioritized control targets via physics-guided mapping and constrained residual learning. CONDA-AI is validated across three industrially grounded domains, including high-throughput continuous casting defect forecasting for selective surface treatment, large scale NCM precursor co-precipitation monitoring under long-horizon drift, and rapid-change acoustic environments requiring real-time robustness. Across these case studies, the proposed framework improves stability of confidence–correctness alignment, supports risk–coverage operating policies, and enhances operational readiness through integrated evaluation and governance protocols. The results suggest that reliability in industrial AI is best achieved by co-designing data handling, shift-aware learning, and decision policies, enabling scalable deployment of trustworthy AI systems across heterogeneous plants and evolving operational conditions.