Information Technology Reference
In-Depth Information
An eigenvalue decomposition of the matrix S is
S
=
VΛV
T
(3)
which reveals the correlation structure for the covariance matrix, where
Λ
is a
diagonal and
V
is orthogonal. The Principal Components (PCs) are represented
by the loading and score vectors so that PCA can decompose the observation
matrix
X
as:
X
=
t
1
p
1
+
t
2
p
2
+
...
+
t
k
p
k
+
E
=
TP
T
+
E
(4)
t
i
,i
=1
,...,k
are vectors, also named scores of the data, which contain informa-
tion on how the samples are related to each other.
p
i
,i
=1
,...,k
are the loadings,
also are the eigenvectors of the covariance matrix,
E
is the residual matrix. The
important statistic for PCA monitoring is given by Hotelling's
T
2
, which is the
sum of normalized squared scores defined as [5],
T
i
=
t
i
λ
−
1
t
i
=
x
i
pλ
−
1
p
T
x
i
(5)
where
t
i
is the ith row of
k
score vectors from PCA model, and
λ
−
1
is a diagonal
matrix containing the inverse eigenvalues associated with the
k
eigenvectors. The
equation represents the distance in principal component model subspace, which
shows that
T
2
is a measure of the variation within the training data set.
However, monitoring the output variable via
T
2
based on the first
k
PCs is
not sucient. If a totally new type of special event occurs which was not present
in the reference data used to develop the PCA model, the new observations
X
new
will move off the plane. Such new events can be detected by computing
the squared prediction errors (SPE) of the residuals of new observations [6].
SPE
X
new
=
i
=1
(
X
new,i
−
X
new,i
)
2
(6)
This statistic is referred to as the Q-statistic, or distance to the model. It rep-
resents the squared perpendicular distance of a new observation from the plane.
When the process is normal, this value should be small. Upper control limit for
this statistic can be computed from training data set. Q statistic indicates how
well each sample conforms to the PCA model, it is a measure of the amount of
variation in each sample not captured by the
k
principal components retained
in the model.
2.2 Support Vector Machine
SVM has become an increasingly popular technique for machine learning activ-
ities including classification, regression, and outlier detection [7]. The idea of
using SVMs for separating two classes is to find support vectors, which refers to
a representative of training data points, to define the bounding planes, in which
the margin between the both planes is maximized [8].
SVM maps the input vectors into some high dimensional feature space through
nonlinear methods. In this space a linear decision surface is constructed with
special properties that ensure high generalization ability of the network. The
number of support vectors increases with the complexity of the problem.
Search WWH ::
Custom Search