Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 176

4.2.5.2 Soft Sensor Models and Plant models
From the point of view of soft sensor models (and in many other cases: control
system design, fault isolation and detection systems) it is very convenient to have
a dynamic or static model of the plant. If a model is available the soft sensor's
performance may be evaluated for different operating conditions. But since plant
models in the mineral processing industry are usually quite complex, in most cases
a model which is qualitatively good but only approximate from the quantitative
point of view is sufficient to design and test the soft sensor's basic behavior. The
final adjustment of the soft sensor may then be performed with plant data using
the best models determined. For example, plant models have been used to design
and test soft sensors in [12, 18, 20, 28]. Figure 4.25 in Section 4.4.2 shows the
SAG mill circuit for which a dynamic model able to handle two types of ore was
developed, and used for designing a grindability index soft sensor [22]. In Figure
4.9 the response of the model circuit variables to a step change in fresh ore feed
flow is shown.
4.2.6 Soft Sensors Designed From Measurement Features
A different way to approach soft sensor design is by analyzing characteristics or fea-
tures extracted from measured variables, such as variances and wavelet transforms,
possibly after undergoing transformations using PCA and discriminant analysis.
Such is the case of the on-line estimation of the grindability index based on con-
catenated wavelet transform variances. Tests done by simulation using a dynamic
SAG circuit state model of a large industrial grinding circuit show good results, as
seen in [22] and in Section 4.4.2. See also Figures 4.24 and 4.27.
Another example is the estimation of ore lithological composition by a classi-
fier using genetic algorithms on the results of image processing of the ore carried
on a conveyor belt. Seven lithological classes are considered and 130 features are
extracted from each rock sample out of a database consisting of 760 digital images
[23, 25]. Calibration was done in the laboratory using ore samples obtained at a
mine site .
Hy otyniemi et al. [32] have used image analysis of flotation froth to extract fea-
tures such as speed of froth movement, color, area, aspect ratio, perimeter, roundness
and transparency of bubbles in the froth. These features are used for categorizing the
froth into classes, and for flotation control by means of an expert control system.
4.2.7 Soft Sensors in Control Loops
One of the most important factors determining a suitable control strategy for a plant
is the availability of sensors to measure not only controlled and manipulated vari-