Data Based Process Monitoring of Industrial Processes Using Multiscale Nonlinear Multivariate Statistical Methods

Data Based Process Monitoring of Industrial Processes Using Multiscale Nonlinear Multivariate Statistical Method.

Chiranjivi Botre^a, Majdi Mansouri^b, Mohamed N. Nounou^c, Hazem N. Nounou^b and M. Nazmul Karim^a

^a Artie McFerrin Dept. of Chemical Engineering, Texas A&M University, College Station, Texas 77843, USA

^b Electrical and Computer Engineering Program, Texas A&M University at Qatar, Doha, QATAR,

^c Chemical Engineering Department, Texas A&M University at Qatar, Doha, QATAR

Abstract

Process monitoring is an important aspect in an industrial processes to ensure safety of the plant and maintain product quality at desired level. One of the important steps in process monitoring is fault detection and diagnosis. Multivariate statistical methods are data based fault detection methods, which are powerful tools capable to handle noise and correlated data set. Two commonly used multivariate methods are partial lease square (PLS) and principle component analysis (PCA).

In this work wavelet-based approach is applied to developed multivariate statistical model to improve fault detection performance. Wavelet based approach provides three distinct advantages: denoise the original signal, de-correlate auto-correlated data, and wavelet coefficients have Gaussian distribution irrespective of original signal. Most of the industrial processes are nonlinear in nature; to improve fault detection performance of the nonlinear processes, Kernel based PLS algorithm is developed. Our fault detection algorithm consists of wavelet decomposition of the input data; the resultant data set is modeled at each scale using KPLS algorithm, and generalized likelihood ration test (GLRT) is used to detect fault at individual scales and at global scale; GLRT is a composite hypothesis testing method and has shown better fault detection performance compared to conventional T² and Q statistic [3]. Fault identification is carried out using contribution plots, GLRT statistical values are computed for all the process variable to identify the violated variables that contribute towards fault in the process.

The developed fault detection model is optimized based on missed fault detection rate, false alarm rate and early fault detection. These three criteria are also used to demonstrate the effectiveness of multiscale-KPLS based GLRT over convention KPLS model. KPLS model can also be used as a nonlinear regression model to predict the quality variables from continuous online variables. Fault detection performance is illustrated through Tennessee Eastman process problem (TEP), which is a continuous process problem based on Eastman chemical company. TEP simulator is used to simulate wide variety of faults occurring in a chemical plant. Results show that our method has superior performance.

Keywords: PLS, GLR, KPLS, wavelet function, fault detection, Tennessee Eastman process.

References:

1.     Bakshi, Bhavik R. ÒMultiscale PCA with Application to Multivariate Statistical Process Monitoring.Ó American Institute of Chemical Engineers. AIChE Journal 44, no. 7 (July 1998): 1596.

2.     MacGregor, J.F., and T. Kourti. ÒStatistical Process Control of Multivariate Processes.Ó Control Engineering Practice 3, no. 3 (March 1995): 403–14.

3.     Mansouri, M., Nounou, M., Nounou, H., Karim, N., 2016. "Kernel pca-based glrt for nonlinear fault detection of chemical processes." Journal of Loss Prevention in the Process Industries 26 (1), 129–139.