Causal Structure Discovery for Industrial Process Monitoring: A Novel Approach for Nonlinear Systems With Complex Noise

Modern industrial processes generate massive volumes of sensor data requiring automated causal discovery for process optimization and predictive maintenance. However, industrial big data exhibits mixed linear-nonlinear relationships with non-Gaussian noise, challenging existing methods. This paper presents a scalable causal discovery framework leveraging polynomial chaos expansion to handle complex industrial data characteristics. Our approach utilizes a multi-criteria strategy to minimize false discoveries critical for industrial safety. The method processes high-dimensional sensor data in polynomial time while providing uncertainty quantification through bootstrap confidence intervals. Validation on real industrial datasets from chemical processes demonstrates 88.9% F1-score, outperforming baseline methods by 27.7%. Furthermore, experiment on synthetic benchmark confirms robust performance across varying problem dimensions, with improvement margins increasing from 26.7% to 63.9% as dimensionality grows. The framework successfully identifies both linear and nonlinear mechanisms from data, enabling data-driven decision support for operators. This work advances data-driven causal discovery for smart manufacturing applications where understanding complex variable interactions is essential.