J Control
Theo
η
Appl
2013
11
(4)
615-622
DOI
1O
.1007/s11768-013-2102-2
Nonlinear fault detection based on
locally linear embedding
Aimin MIAO, Zhihuan
SONG
飞
Zhiqiang
GE
, Le
ZHOU
, Qiaojun
WEN
State
Key
Laboratory
o
fI
ndustrial
Control
Technology
,
Zhejiang
University
,
Hangzhou
Zhejiang
310027
,
China
Abstract:
In
this paper, a
new
nonlinear fault
d
巳
tection
technique based
on
locally linear embedding
(LLE)
is
de-
V
巳
loped.
LLE
can
efficiently compute
the
low-dimensional embedding of
the
data
with
the
local neighborhood structure
information preserved.
In出
is
method, a
data-d
巳
pendent
kernel matrix
which
can
reflect the nonlinear data
structur
巳
is
defined. Based
on
the kernel matrix,
the
Nyström formula
makes
the
mapping extended
to
the
testing data possible.
With
th
巳
kernel
view
of
the
LLE,
two
monitoring statistics
are
constructed. Together
with
th
巳
out
of
samp
1e
extensions,
LLE
is
used for nonlinear fault detection. Simulation cases
w
巳
re
studied
to
demonstrate
the
p
巳
rformance
of
the
proposed method.
Keywords: Locally linear embedding; Fault detection; Nonlinear dimension
r
巳
duction
1 Introduction
Data-based multivariate statistical
techniqu
巳
s
have be-
come
one
of
the research hotspots in industrial fault de-
tection. As a representative data-driven technique
, princi-
pal component analysis (PCA) has been
eff
,巳
ctively
applied
in this area [1-3]. Compared to traditional model-based
fault detection methods
[牛毛],
PCA
is more sophisticated
for addressing high dimensional
, noisy and strongly corre-
lated process data. By projecting the data onto the lower di-
mensional space where most
of
the variance in the original
data is preserved
,
PCA
helps
achiev
巳
the
process monitoring
with great profìciency. However
, due to the linear assump-
tion
,
PCA
may
not perforrn well in monitoring some non-
linear chemical processes. Such shortcoming has recently
been brought into focus and various extensions
of
the lin-
ear
methods have been
report
巳
d.
Kr
amer
且
rst
developed a
nonlinear
PCA
method based
on
the auto associative neural
network [1]; Hiden combined genetic programming and lin-
ear
PCA
to produce a nonlinear
PCA
algorithm [7]; Maulud
proposed an orthogonal nonlinear
PCA
based on the wavelet
decomposition [8]; and Karnpjarvia introduced neural net-
works and the
PCA
based fault detection and isolation sys-
tem
[3]. Kern
e1
principal component analysis (KPCA) is
another powerful tool and has been widely appreciated for
nonlinear dimension reduction [2
, 9]. Via the kernel trick,
KPCA
extends the linear
PCA
to nonlinear case based on
the idea
of
perforrning
PCA
in
th
巳
high
dimension fea-
ture space. Without tedious nonlinear optimization proce-
dur
时,
KPCA
has been successfully applied and showed su-
perior perforrnance to their linear counterparts for nonlin-
ear
process monitoring [2, 10]. Although KPCA can cap-
ture the higher-order relationships among the data points
,
the structure
of
the manifold data has not been we
l1
con-
sidered which may also has signifìcant impact
on
the moni-
toring perforrnance.
On
the other hand, the perforrnances
of
such methods largely depend
on
the kernel function, e.g., a
Received
9
May
2012;
revised
14
September
2012
poor kernel choice may lead to signifìcantly impaired per-
forrnance [11].
Manifold learning is a recently proposed nonlinear di-
mensionality reduction algorithm
, which was found power-
ful in discovering the intrinsic geometrical structure
of
the
nonlinear data. In manifold learning
, nonlinear dimension-
ality reduction is achieved with certain local neighborhood
information preserved. Locally linear embedding (LLE) is
one
of
such representative
m
巳
thods
which expects each data
and its neighbors on a locally linear patch [12].
Th
巳
local
neighborhood structure relationship
of
these patches is de-
scrib
巳
d
and preserved in a low-dimensional space through
an optimal way.
LLE
seeks to map the
c1
ose-by points in the
original space to be
c1
0se in the low dimensional represen-
tation
, while the global structure
focus
巳
d
algorithm, such as
PCA
may map the faraway points nearby. In
LLE
, the low
dimension embedding is obtained
by
th
巳
data
itself, as a re-
sult
, it can
effectiv
巳
ly
discover the intrinsic nonlinear data
structure hidden in the high-dimensional process data. In
contrast
, the traditional KPCA method which needs a prior
kernel function for the low dimension mapping
, may distort
the geometry structure
of
the data [11].
The original
LLE
method yields mappings defìned only
on the training data points; it is un
c1
ear how to evaluate the
mapping for new data points. For online process monitor-
ing
, it is required that the mapping should be able to con-
duct dimensional reduction
on
newly measured data sam-
ples. To obtain the maps
on
the
n
巳
w
testing data points and
make it suitable for practical applications
, some
improv
而巳
ments have been reported in
t
白
h
巳
literature
[口
13-15
叮].
Th
巳
10-
cal
st
位
ructωure
巳
eι
七
as
巳
d
algorit
出
hmn
巳
eighbor
由
hood
pres
巳创
rving 巳
m-
b
巳
dding
(NPE)
i
妇
s
a
lin
巳
ar
approx
对
imation
algorithm
of
LLE
to evaluate the maps on the new points [14]. However
, it is
essentially a linear dimensionality
techniq
肘,
the ability in
capturing the nonlinear data structure and dealing with the
nonlinear processes is limited. Thus
, it is intrinsically not a
↑
Coπesponding
author.
E-mail:
songzhihuan@zju.edu.cn.
Te
l.:
+86-571-87951442-8007;
fax:
+86-571-87951921
Th
is
work
was
supported
in
part
by
由
e
National
Basic
Research
Program
of
China
(973
Program)
(No.
2012CB720505)
and
th
巳 National
Natural
Sci
巳
nce
Foundation
ofChina
(No.
61273167).
@
South
China
University
ofTechnology
and
Acad
巳
my
ofMath
巳 matics
and
Systems
Science
,
CAS
and
Springer
斗
'erlag
Berlin
Heidelberg
2013