Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 154

Tabl e 4. 5 ( continued )

Reference Model class Main features Applications

RBF NN Soft sensor based on radial basis

function (RBF) NN. Data is sub-
ject to clustering and a RBF NN
is developed for each cluster. The
SS output is given by the sum of
the several RBF NN weighted by
the respective membership func-
tions to each cluster.

Tests in high purity distillation
column.

Wang,
1996 [37]

Gray (empirical
+ phenomeno-
logical)

Soft sensor for flow in open chan-
nels having an overshot (sub-
merged) gate, using as inputs to
a gray box model upstream and
downstream levels and gate an-
gle.

Tests in open channel point to the
selection of one of three proposed
models.

Wendt,
1999 [38]

State Generic soft object oriented de-

velopment tool for soft sen-
sor development includes on-line
model updating.

Applications to industrial SAG

and ball milling circuits.

Herbst,
1999 [39]

LS-SVM SS model design using least

square regression and SVM. Non-
linear regression transformed to
linear regression but in higher
dimension space. Cost function
includes a term for penaliz-
ing model complexity. A simple
training algorithm is given.

Tests using data of gasoline ab-
sorbing rate in cracking unit of
arefinery. Better results are re-
ported than those obtained if NN
SS model is used, including less
data needed in training.

Zhang,
2006 [40]

PCA Identification of sensor faults us-

ing sensor validity index.

Tests using boiler process data.

Dunia et
al. , 1996
[34]

— Problems arising in soft sensors

for the hydrocarbon processing
industry are analyzed, including
warnings of wrong interpreta-
tions and procedures. The subject
may be extended to other indus-
tries.

Several graphs are given based on
industrial results to support the
different subjects considered.

King,
2004 [41]

—- System for handling SS in in-

dustrial environment. Includes:
several soft sensor models per
primary measurement, selected
according to analysis of sec-
ondary measurement quality and
availability; parameter estimation
while actual sensors are available;
monitoring of soft sensor errors
and decision to update parameter
of models using appropriate exci-
tation signals. Flexible configura-
tion using text strings.

Tested on particle size estimation
in an industrial grinding plant.

Barrera,
1996 [42]

Notation: LIP = Linear-in-parameters; NARX = Nonlinear ARX; NN = Neural Network;
SVM = Support Vector Machine SS = Soft Sensor; PCA = Principal Component Analysis;
RMS = root mean square; LS =least squares; RBF = radial basis function; SAG = Semi-
autogenous.