Causal‐Invariance Transfer Learning via Process‐Topology‐Aware Invertible Flows for Weak Cold‐Start Quality Prediction in Multicondition Process Manufacturing

Weak cold‐start quality prediction in process industries faces two coupled challenges: new grades provide only a handful of labels, and cross‐condition shifts make observed correlations environment‐dependent. The task is formulated as preserving transferable quality‐formation mechanisms while suppressing environment‐specific disturbances, and an end‐to‐end framework is proposed that uses mechanism stability as the transfer criterion. In an invertible representation space, structured routing disentangles stable predictive information from spurious variation under a process‐chain topology prior. Multienvironment conditional likelihoods model admissible environment‐dependent variation in the base distribution, and adversarial regularization suppresses residual environment‐discriminative information. Under weak supervision, causally consistent augmentation expands target‐domain support, and likelihood‐based screening limits off‐manifold drift and extrapolation. Conditional quantile modeling captures central tendency and tail uncertainty, yielding calibrated prediction intervals. Evaluation uses a hot‐rolling dataset of 3391 coils and 129 variables spanning steelmaking to coiling, covering 17 grades, seven of them target grades with fewer than 50 labels each. Against thirteen baselines, the framework achieves the lowest mean absolute percentage errors, 1.2%–4.3% for tensile strength, yield strength, and elongation, a 37%–45% average reduction. The nominal 80% prediction interval attains 82% empirical coverage. The framework therefore supports quality assessment and risk screening for new grades before inspection data accumulate.

Authors

Institutions

Publication Details

Journal
Advanced Intelligent Systems
Published
2026-11-17
DOI
https://doi.org/10.1002/aisy.70631
Primary Topic
Fault Detection and Control Systems
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Causal‐Invariance Transfer Learning via Process‐Topology‐Aware Invertible Flows for Weak Cold‐Start Quality Prediction in Multicondition Process Manufacturing

Romain Bérard, Zhimin Lv, Johannes De Smedt, Yiquan An et al.
Advanced Intelligent Systems
Fault Detection and Control Systems
article

Causal‐Invariance Transfer Learning via Process‐Topology‐Aware Invertible Flows for Weak Cold‐Start Quality Prediction in Multicondition Process Manufacturing

Romain Bérard, Zhimin Lv, Johannes De Smedt, Yiquan An, Yihang Li, Xi Sun
article en

Abstract

Weak cold‐start quality prediction in process industries faces two coupled challenges: new grades provide only a handful of labels, and cross‐condition shifts make observed correlations environment‐dependent. The task is formulated as preserving transferable quality‐formation mechanisms while suppressing environment‐specific disturbances, and an end‐to‐end framework is proposed that uses mechanism stability as the transfer criterion. In an invertible representation space, structured routing disentangles stable predictive information from spurious variation under a process‐chain topology prior. Multienvironment conditional likelihoods model admissible environment‐dependent variation in the base distribution, and adversarial regularization suppresses residual environment‐discriminative information. Under weak supervision, causally consistent augmentation expands target‐domain support, and likelihood‐based screening limits off‐manifold drift and extrapolation. Conditional quantile modeling captures central tendency and tail uncertainty, yielding calibrated prediction intervals. Evaluation uses a hot‐rolling dataset of 3391 coils and 129 variables spanning steelmaking to coiling, covering 17 grades, seven of them target grades with fewer than 50 labels each. Against thirteen baselines, the framework achieves the lowest mean absolute percentage errors, 1.2%–4.3% for tensile strength, yield strength, and elongation, a 37%–45% average reduction. The nominal 80% prediction interval attains 82% empirical coverage. The framework therefore supports quality assessment and risk screening for new grades before inspection data accumulate.

Advanced Intelligent Systems
Chinese Academy of Sciences (CN), National Science Library (CN), University of Science and Technology Beijing (CN)
Fault Detection and Control Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon