MIQ-TRA: A Novel Modeling and Anomaly Detection Framework in Ironmaking Cyber-Physical System Integrating Stationarity and Nonstationarity

Siwei Lou, Chunjie Yang, Chao Liu, Hanwen Zhang

Abstract: In the contemporary era of big-data-driven steel industry, artificial intelligence technologies have emerged as powerful enablers, providing substantial support to field engineers and alleviating their workloads. Nevertheless, the pronounced uncertainty and nonstationarity inherent in the highly intricate modern ironmaking cyber-physical system (ICPS) pose formidable challenges to accurate process modeling and safety assessment. To surmount these obstacles, this research presents a novel framework for molten iron quality-related total regression analysis, namely MIQ-TRA. Initially, a time-constrained interval type-2 (TC-IT2) fuzzy representation is devised to handle data uncertainty and nonlinearity across diverse time windows. This approach not only extracts interpretable nonlinear features but also effectively captures the complex interrelationships within the data. Following this, concerning the stationary features associated with MIQ, a stationary regularized regression objective function is formulated by ingeniously integrating stationarity and regressivity. Subsequently, a decoupled model solver is derived in a closed-form solution, which process effectively partitions the data into the MIQ-related stationary subspace and nonstationary residual. For the nonstationary residuals, a nonstationary clustering partial least squares (PLS) method is introduced. It enhances the nonstationary features, performs pattern clustering, and models the nonstationary information related to MIQ with PLS. In the integration stage, three components-TC-IT2 fuzzy representation, stationary feature separation, and nonstationary residual modeling-are comprehensively integrated into MIQ-TRA framework. When applied to numerical simulations and real-world ICPS, our MIQ-TRA demonstrates superiority over existing methods. For anomaly detection, it exhibits lower false alarms and higher accuracy levels, while MIQ prediction attains lower prediction errors. These outcomes empower field engineers to make more reliable decisions, thereby underscoring the practical significance.

External IDs:doi:10.1109/ticps.2025.3569214