Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 208

4.5 Final Remarks
Soft sensors have become widely used in many industrial plants to replace sensors
either as a back up for installed sensors that cease to be available, or as alternative
sensors when no real sensors are installed, if appropriate conditions are met.
The growing area of soft sensor R&D and applications has also encompassed
the mineral processing industry, where the relatively difficult modeling problems
are inherited by the problem of finding suitable soft sensor models. Soft sensors
have been developed for ore particle size, slurry density, ore grade, ore grindability,
mill contents, operational work index and weight in mineral processing plants. But
sensors or soft sensors are still missing for on-line measurement or estimation of
important variables in mineral processes.
Different classes of models that have been used include LIP models, which may
be linear or nonlinear in plant measurements. Among these models, nonlinear mod-
els of the NARX class have the advantage of allowing the introduction of process
phenomenology, while retaining the relatively simple and fast updating properties
of the linear models. An added advantage of LIP models is the geometrical insight
involved. Also employed are neural network models, fuzzy models, support vector
machine based models, and soft sensors designed using characteristics of measures
variables such as mean, variances of the variables themselves or of transformations
of them.
The issue of soft sensor robustness, specifically low sensitivity to disturbances
caused by failing secondary measurements and to changes in operating conditions
and plant characteristics, has been approached. On the one hand, by using models
that cover relatively large regions of plant characteristics and operating conditions,
such as models based on neural networks, fuzzy sets, clustering, and models of the
NARX class. On the other hand, robustness to secondary measurement unavailabil-
ity has been obtained by introducing a set of candidate soft sensors - instead of
having only one soft sensor - based on models whose components have forcibly
been made independent of different secondary measurements. Soft sensor shell sys-
tems have been designed to manage these and other aspects of soft sensor operation
in industrial plants.
New data processing and sensor technologies in general, and new equipment in
the mineral processing industry may be expected to promote the development of
new sensors and soft sensors. On the one hand, nanotechnologies, mechatronics and
developing wireless technologies will allow networks of very high speed miniature
processors, allowing in turn networks of miniature sensors and soft sensors pro-
fusely located throughout the plant equipment. New sensors for as yet unmeasured
variables, developed using these new technologies, will surely appear. On the other
hand, it is also reasonable to expect that the development of new mineral process-
ing equipment based on new technologies may require new sensors and soft sensors.
All these developments should have an important impact on improving plant control
and maintenance and, as a result, on plant technical and economic performance.