"PRESAGE: Physics-Anchored Digital Twin of a Hot Strip Mill -Two-Year Dataset from Controlled Deployment (April 2024-March 2026)"

Abstract

"This dataset comprises 6.99 million evaluated passes across 583,468 coils over 24 months (April 2024\u2013March 2026) from a 12-stand hot strip mill operating under controlled conditions. It contains comprehensive pass-level records with 56 columns spanning full PLC signal integration (444 tags), per-coil summaries, roll-change logs, incident records, and high-rate 100 ms time series for representative operational windows. The dataset validates a physics-anchored digital twin incorporating safety-routed online residual learning, demonstrating 1.84 % force prediction error via a River-based ensemble combining AMRules, SGD, and EWMA learners across all 12 stands.Competition participants and researchers may use this dataset to: (1) reproduce, validate, or improve upon the reported routed prediction accuracy and safety routing mechanisms; (2) develop novel early-warning monitors for real-time anomaly detection and condition-based maintenance; (3) optimize gauge control strategies or advance flatness prediction algorithms; (4) benchmark federated learning architectures across distributed mill stands; (5) explore physics-informed machine learning approaches for industrial process control.Companion analysis tools include the physics validation suite (vphysics.py), the learner engine verification harness (verify_engine.py), and the derived-quantity recomputation script (recompute_derived.py). Submissions should report force absolute percentage error, gauge error, router fallback rate, and drift detections against both the 48-hour reference window from fresh rolls and the full two-year campaign."