A Local-Model Integration Approach to The Transition State Prediction of Continuous Annealing Processes

In a continuous annealing process (CAP), an accurate and timely temperature prediction is crucial, particularly during transitions between different types of strips. Predicting transition states is challenging due to process nonlinearities and time-varying characteristics induced by these transitions. In addition, extensive historical datasets hold valuable insights that can be harnessed to detect consistent and similar changes within transition groups. In this study, we introduce a data-driven framework known as the “progressive search and parallel model” to forecast transition states rapidly. The “progressive search” is executed by categorizing subclasses based on parameter properties and spatial locations, each with varying priority levels. Specifically, parameters are subdivided into segments along a piecewise linear scale, and a just-in-time learning framework is employed to establish local models at these levels concurrently. We apply this framework in both stable and transitional CAP conditions, demonstrating its robust performance in effectively tracking shifting trends during state transitions.