Dynamic reconstruction principal component analysis for process monitoring and fault detection in the cold rolling industry

Abstract

An improved method for process monitoring and fault detection called dynamic reconstruction principal component analysis (DRPCA) is proposed. By extracting direct dynamic connections between samples, DRPCA utilizes the overall dynamic information of the training data set and improves the monitoring performance of dynamic industrial processes. In DRPCA, the optimal orthogonal transformation is used to reconstruct the past sub-matrix with the current sub-matrix. The reconstructed increment matrix preserves the raw data’s basic, incremental, and dynamic information as much as possible. In addition to the traditional T-squared and SPE statistics, a new statistic SPE-R is proposed for DRPCA based on the reconstruction accuracy. We evaluate the performance of the proposed method on a cold rolling mill system, and the results show that DRPCA outperforms DPCA and its improved versions in terms of faster computation speed, more timely alerts, higher detection rates, and lower false alarm rates. Our study demonstrates that DRPCA is a superior method for monitoring dynamic processes.