Locality preserving randomized canonical correlation analysis for real-time nonlinear process monitoring

Abstract

Hazard identification and analysis is an important step in the process safety assessment/management in the modern process industry. For hazard identification, real-time monitoring of process operations plays a critical role to establish required safety measures. In this paper, a novel locality preserving randomized canonical correlation analysis (LPRCCA) method is proposed for real-time nonlinear process monitoring. The basic idea is to map the original data onto a randomized low-dimensional feature space through random Fourier feature map, and then integrate the local geometric structure information to improve data mining performance. On the basis of explicit low-dimensional random Fourier features, the computational cost of the online feature extraction is dramatically reduced. The proposed LPRCCA method is significantly more favorable than kernel-based methods for nonlinear process monitoring. The applicability and effectiveness of the proposed process monitoring scheme are verified through a numerical example and an industrial benchmark of the Tennessee Eastman process (TEP) by the comparisons with other relevant methods.