Locally Weighted Canonical Correlation Analysis for Nonlinear
Process Monitoring
Qingchao Jiang and Xuefeng Yan*
Key Laboratory of Advanced Control and Optimization for Chemical Processes of Ministry of Education, East China University of
Science and Technology, Shanghai 200237, P. R. China
*
S
Supporting Information
ABSTRACT: A locally weighted canonical correlation analysis (LWCCA)
method is proposed to achieve efficient nonlinear process monitoring.
The basic idea of the LWCCA is to approximate a nonlinear process
through several local linear canonical correlation analysis (CCA) models, in
which the determination of sample weights is a key step. Slowly decreasing
weights will ignore the local behaviors, whereas rapidly decreasing weights
will lead to significant false alarms. A randomized algorithm-based approach
is proposed to determine the tunable parameter for calculating the weights.
Thus, the LWCCA model explores as much local behavior as possible with
the false alarm performance guaranteed. When a local CCA model that
characterizes the process input and process output correlation is estab-
lished, optimal fault detection residuals are generated, and monitoring sta-
tistics are established. Two experimental studies are conducted through
which the efficiency of the LWCCA method is verified.
■
INTRODUCTION
Process monitoring is crucial in maintaining the long-term safe
operation of a production plant. Rapid advancement of data
collecting and transmitting techniques has created an abundance
of process data that contain meaningful available process infor-
mation.
1−3
Especially, data-driven multivariate analysis (MVA)
methods play a more important role in process monitoring.
4−9
Generally, a classical MVA monitoring method follows offline
modeling with online monitoring procedures. During the offline
modeling, a multivariate data analysis technique is employed to
explore the relationship among variables and to construct the
feature spaces for monitoring. In the online monitoring proce-
dure, a query sample is projected into the monitoring spaces, and
the process status is determined according to monitoring
statistics.
Principal component analysis (PCA), partial least-squares
(PLS), and canonical correlation analysis (CCA) are the basic
MVA methods. PCA focuses on the variable relationships of the
entire process (without discriminating process input and
output).
10−12
It constructs a dominant subspace and a residual
subspace according to the features’ importance for reconstruct-
ing the original data. PLS focuses on the quality-related process
monitoring.
13−15
It constructs the quality-related subspace and
the residual subspace according to their relationship with
difficult-to-measure quality variables. A fault in the quality-
related subspace is important, because the fault will generally
affect the production quality. CCA explores the correlation
between two sets of variables. It is generally used in two different
ways for process monitoring, as follows: one way focuses on the
relation between the process input and process output; and the
other way focuses on the relation between two coupled
units.
16−20
It is proven that the CCA generates optimal fault
detection residual when only one set of variables are affected by a
fault.
18
Given the efficiency, PCA, PLS, and CCA-based moni-
toring methods have been intensively extended to solve various
monitoring problem. Other MVA-based monitoring methods
were also proposed and efficiency has been reported.
21−23
How-
ever, these methods generally assume that the measured vari-
ables are linearly related, which limits their applications in
nonlinear processes.
For nonlinear process monitoring, the neural networks (NN)-
based methods,
24,25
the kernel learning methods,
26−28
and the
just-in-time learning methods
29,30
are the basic ones. The
NN-based methods extract features through nonlinear mapping,
which is relatively sophisticated and generally requires a large
amount of computation. Moreover, the structure and parame-
ters should be determined in designing an NN for monitoring,
which remains a challenge. The kernel-based methods replace
the nonlinear mapping through kernel functions and then
extract features in the high-dimension space. Although efficiency
is shown, the selection of related parameters and kernel func-
tions is subjective, which significantly affects monitoring results.
Recently, just-in-time learning (JITL) approaches have been
developed for nonlinear process modeling and monitoring.
29−31
The basic idea of a JITL method is to establish several local
Received: April 26, 2018
Revised: August 7, 2018
Accepted: September 13, 2018
Published: September 13, 2018
Article
pubs.acs.org/IECR
Cite This: Ind. Eng. Chem. Res. 2018, 57, 13783−13792
© 2018 American Chemical Society 13783 DOI: 10.1021/acs.iecr.8b01796
Ind. Eng. Chem. Res. 2018, 57, 13783−13792
Downloaded via EAST CHINA UNIV SCIENCE & TECHLGY on December 30, 2018 at 07:25:16 (UTC).
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.