SHIFT: Self-Healing Intelligence in Feature Store and Data Store Transitions

Abstract

This study presents an AI model management framework for maintaining the prediction performance of virtual metrology (VM) systems in steel manufacturing processes. The proposed method addresses concept drift (CD), a major threat to long-term model reliability, by employing a dual-update strategy that autonomously updates both the feature store and the data store. Unlike conventional model maintenance approaches based on static inputs and accumulated data, our framework continuously reassesses feature relevance and adapts retraining datasets to reflect evolving process conditions. In a real-world steel sintering process, the method achieved significant recovery in predictive accuracy—from 65% post-deployment to over 93% after updates. This study highlights the importance of proactive maintenance for AI models and demonstrates how an intelligent update strategy can extend the usable life and reliability of VM systems in critical production workflows. The findings contribute to the broader field of AI-enabled asset management by offering a scalable solution for condition monitoring, retraining scheduling, and long-term performance assurance.