A hybrid KSFA-ICA-MVNCA framework for shape quality monitoring and anomaly tracing in cold rolling of high-strength steel

Abstract

Cold rolling, a critical process in steel manufacturing, plays a decisive role in determining flatness quality. However, conventional monitoring techniques exhibit limitations in defect detection and root cause tracing. This study leverages data mining theories and methods to analyze and trace flatness quality anomalies in high-strength steel. A monitoring and tracing framework, KSFA-ICA-MVNCA, was developed by integrating kernel slow feature analysis (KSFA), independent component analysis (ICA), and multivariate network causal analysis (MVNCA). The framework has two main features: (1) the KSFA-ICA model effectively handles the nonlinearities, dynamics, and mixed Gaussian/non-Gaussian distributions inherent in cold rolling processes to achieve robust anomaly monitoring; (2) MVNCA addresses the challenges posed by numerous and strongly coupled variables, thereby enabling precise root cause analysis of anomalies. Based on experimental results from 13 coils of high-strength steel production data (approximately 1500 sampling points per coil, totaling nearly 20000 samples), the KSFA-ICA joint monitoring model achieved an average monitoring rate of 84.21% with a false alarm rate of only 2.84%, demonstrating overall superior performance compared with traditional monitoring models such as PCA and PLS. MVNCA enabled the identification of key factors influencing flatness quality through bidirectional information flow. For validation, feature maps were constructed based on the temporal variations of features variables and strip flatness (IU), thereby confirming the effectiveness of the proposed method in anomaly tracing during the cold rolling process.