Success Stories

for Control

Performance Monitoring for Mineral Processing

Monitoring and Control with Computer Vision
Most of the developments were based on the application of computer vision
systems in areas where no devices were previously in place to measure key
variables in reaction systems:

• Customized computer vision algorithms are used to estimate

the proportion of fines in coal feed systems accurately
enough to allow online control (Figure 1). Excessive fines
in the feed may adversely affect gas flow through the
reactor burden, leading to substantial losses associated with
suboptimal operation. Previous methods for online analysis
of particulate feeds were either nonexistent or based on
The chemical and metallurgical process inefficient automated sieve samples taken from belts.
Figure 1: An image from a
industries face stiff challenges in the form • Multivariate feature extraction from platinum froth images computer vision system

of increasing energy costs, increasingly could be used as a basis for system identification and on a coal conveyor is
advanced control of froth flotation systems. With the aid automatically processed
stringent environmental regulations, and of advanced process models, these features can be used to to estimate the proportion
global competition. Although advanced reliably estimate prevailing process conditions, which is not of fines in the coal.

control is widely recognized as essential to otherwise feasible (Figure 2).

meeting these challenges, implementation

is hindered by more complex, larger-scale dataset 1

circuit configurations, the tendency toward

plantwide integration, and in some cases an 0.05
L
increased lack of trained personnel. In these 3 0

environments, where process operations -0.05

dataset 2 0.1
0.1
are highly automated, algorithms to detect
0.05
and classify abnormal trends in process -0.1
0
measurements are critically important.
L -0.05 -0.3 L
2 1
Advanced algorithms and measurement
control systems have been designed and dataset 1 dataset 2
implemented for process performance Figure 2: Change in process conditions is detected by an online computer vision system by
monitoring and operational performance projection of froth image features to a control chart.

management, yielding substantial benefits

in operating installations. These techniques have been developed by the Anglo Platinum Centre for Process
Monitoring and their associates at the University of Stellenbosch in South Africa to
enable their industrial partners to realize the competitive advantages afforded by
advanced control and monitoring systems.

Contributor: Chris Aldrich, University of Stellenbosch, South Africa

From: The Impact of Control Technology, T. Samad and A.M. Annaswamy (eds.), 2011. Available at www.ieeecss.org.
 Advanced process monitoring
Reference Data Fault Condition Data

technology can expedite safe

Secondary Mill Power

Secondary Mill Power

1200 1200

operation closer to production 1150 1150

limits, eliminate major process 1100 1100

upsets, and reduce minor, often 1050 1050

260 280 300 320 260 280 300 320

unreported, abnormal situations. Secondary Mill Load Secondary Mill Load

Monitoring of an autogenous mill on an industrial concentrator plant

Capacity increases of 3-8% have
been estimated for a range
of process industries with
monitoring and related technologies
(www.asmconsortium.net); this
amounts to $200—500 million per
year for the South African
platinum industry alone. Prognostic monitoring of wall temperatures for a metallurgical furnace

Process Fault Detection and Identification

The Centre and its associates have developed • The detection and identification of different operational states in autogenous
state-of-the-art algorithms (Figure 3) for and semi-autogenous mills can at times be realized with an accuracy of up to
process fault diagnosis that are currently 80%—providing a basis for better control and potentially large reductions in
used plantwide by one of the premier mineral energy usage by mineral processing plants.
processing companies in South Africa. The
following benefits have resulted from use of
the algorithms: E (nxm)
residual Squared
• The turnaround time associated with space Predition
Error (SPE) Model
attending to plant faults has been reduced breakdown
from several weeks or even months to less
than three days.

• The frequency of large events leading to

process circuit and equipment downtime— Residual space
X (nxm) X’ (nxm)
and associated losses in revenue—has been original reconstructed
process data data
substantially reduced.

• Plant alarms and the cost of alarm Feature 1

forward reverse
management have been reduced mapping mapping
T (nxa)
considerably. feature Process out
space Score space of control
• The development of large-scale process Feature 2
monitoring systems has enabled early
detection of thermal runaway in
metallurgical furnaces.

Figure 3: Monitoring the model prediction error in an appropriately

constructed feature space enables advance warning of process upsets.

For further information: The Anglo Platinum Centre for Process Monitoring, www.sun.ac.za/cpm