5
into a rule of thumb as such. Additionally, such
rules of thumb require faith that the target value
can be effectively and accurately measured. If sen-
sors drift over time or fail entirely, then they may
become inaccurate, useless, or actively harmful.
3. Favor encoding stable controls on stable measure-
ments: In many cases this involves creating opti-
mizing controls instead of on/off controls or PID
controls. Especially with X-ray fluorescence val-
ues, which are known to drift slightly over time
in our operations, it is often preferable to target
controls based on how the grade of a given ele-
ment is changing over time rather than for specific
numbers. The drift on the absolute values of any
given element reading can become significant, but
the drift on the difference between two consecutive
measurements of the same element is in our expe-
rience generally negligible. Thus, a control which
minimizes or maximizes a reading is often more
reliable overall than one which responds to a spe-
cific target value.
4. Mill operators are always on the driver’s seat, can
run the whole, any, or no part of Digital One.
When operator disengages expert control or opti-
mization to make any setpoint change and back
into optimization, expert control follows the new
setpoint but limits within 10% change around
new setpoint for 30 minutes, honoring but “cross
checking” operator’s decision.
5. Therefore, always accept operator input: Every
setpoint calculated by the expert system can be
constrained or guided by the operator as needed,
and when the operator changes any of these set-
points the expert control is designed to support the
change.
6. Do not rely on forecasts: No matter how successful
the system has been thus far flotation is considered
a stochastic process for good reason. Any forecasted
variable which is critical to the control system will
inevitably be forecasted critically incorrectly, and
the control system must handle that gracefully.
The ultimate goal of these guidelines is that the system
must be able to stay on production line. Equivalently, it
must be able to survive the operators, in that they should
not find it to be too intrusive or too cumbersome, nor
should they need to be overriding it all the time. It should
handle problematic local events appropriately, typically by
allowing the human operators to address the situation with-
out the expert system’s interference.
For specifics, the expert control implemented at Buick
Mill is particularly focused towards the detection of prob-
lematic marcasite ores and to automatically account for
them when they are detected. The primary indicators here
are the iron assays in the Pb and Zn concentrate streams.
When marcasite (or any unusually high iron) is detected,
the iron depressant is enabled and controlled based on the
iron grade until the iron is reduced to acceptable levels in
the concentrate again.
Similarly, the other hybrid reagents are triggered based
on X-ray assay information. The Pb collector is used to
reduce the amount of Pb in the Zn concentrate, as that is
where it ends being problematic if it does not make it to
the bulk concentrate. The Zn collector is added when the
Zn in final tails stream is unusually high. The Cu collector
is added when the Cu in the Pb concentrate is too high, as
the Cu/Pb separation floats the Cu while sinking the Pb.
Pb depressant is added when the Pb in the Cu concentrate
becomes too high, for the same reasoning. These cutoff val-
ues are based on shippable concentrate targets.
The control system for the hybrid reagents has two
modes: basic and advanced. In basic mode, these controls
are on/off controls with hysteresis. They are enabled when
they pass the upper threshold value and are disabled when
the target variable is reduced below a lower threshold value.
The dosing of the reagents is constant when the reagents are
enabled. In advanced mode, the thresholds remain the same
but the dosing is exponential feedback controlled to rapidly
increase the dosage when the targets are significantly above
the threshold values.
The next step in this part of the control system is to uti-
lize techniques from machine learning to optimize the on/
off thresholds and dosing targets. This will move the system
past simply ensuring concentrate quality is maintained into
directly using these reagents to optimize the overall recover-
ies and plant performance.
The first bulk reagent to be addressed by the expert
control system is copper sulfate. Copper sulfate is used in
the Zn flotation to activate the Zn minerals (overwhelm-
ingly sphalerite), and the presence of sufficient copper sul-
fate can be strongly detected based on the bubble size. This
is a parameter that is detected by the VisioFroth ™ system.
Thus, copper sulfate is added when Zn is being lost to the
tails, but not if the bubble size is already small and indica-
tive of sufficient copper sulfate being present.
The air and MIBC for each part of the flotation cir-
cuit are also controllable. For these the primary param-
eter is bubble texture, another soft property reported by
the VisioFroth ™ system. It is used as a reliable indicator of
into a rule of thumb as such. Additionally, such
rules of thumb require faith that the target value
can be effectively and accurately measured. If sen-
sors drift over time or fail entirely, then they may
become inaccurate, useless, or actively harmful.
3. Favor encoding stable controls on stable measure-
ments: In many cases this involves creating opti-
mizing controls instead of on/off controls or PID
controls. Especially with X-ray fluorescence val-
ues, which are known to drift slightly over time
in our operations, it is often preferable to target
controls based on how the grade of a given ele-
ment is changing over time rather than for specific
numbers. The drift on the absolute values of any
given element reading can become significant, but
the drift on the difference between two consecutive
measurements of the same element is in our expe-
rience generally negligible. Thus, a control which
minimizes or maximizes a reading is often more
reliable overall than one which responds to a spe-
cific target value.
4. Mill operators are always on the driver’s seat, can
run the whole, any, or no part of Digital One.
When operator disengages expert control or opti-
mization to make any setpoint change and back
into optimization, expert control follows the new
setpoint but limits within 10% change around
new setpoint for 30 minutes, honoring but “cross
checking” operator’s decision.
5. Therefore, always accept operator input: Every
setpoint calculated by the expert system can be
constrained or guided by the operator as needed,
and when the operator changes any of these set-
points the expert control is designed to support the
change.
6. Do not rely on forecasts: No matter how successful
the system has been thus far flotation is considered
a stochastic process for good reason. Any forecasted
variable which is critical to the control system will
inevitably be forecasted critically incorrectly, and
the control system must handle that gracefully.
The ultimate goal of these guidelines is that the system
must be able to stay on production line. Equivalently, it
must be able to survive the operators, in that they should
not find it to be too intrusive or too cumbersome, nor
should they need to be overriding it all the time. It should
handle problematic local events appropriately, typically by
allowing the human operators to address the situation with-
out the expert system’s interference.
For specifics, the expert control implemented at Buick
Mill is particularly focused towards the detection of prob-
lematic marcasite ores and to automatically account for
them when they are detected. The primary indicators here
are the iron assays in the Pb and Zn concentrate streams.
When marcasite (or any unusually high iron) is detected,
the iron depressant is enabled and controlled based on the
iron grade until the iron is reduced to acceptable levels in
the concentrate again.
Similarly, the other hybrid reagents are triggered based
on X-ray assay information. The Pb collector is used to
reduce the amount of Pb in the Zn concentrate, as that is
where it ends being problematic if it does not make it to
the bulk concentrate. The Zn collector is added when the
Zn in final tails stream is unusually high. The Cu collector
is added when the Cu in the Pb concentrate is too high, as
the Cu/Pb separation floats the Cu while sinking the Pb.
Pb depressant is added when the Pb in the Cu concentrate
becomes too high, for the same reasoning. These cutoff val-
ues are based on shippable concentrate targets.
The control system for the hybrid reagents has two
modes: basic and advanced. In basic mode, these controls
are on/off controls with hysteresis. They are enabled when
they pass the upper threshold value and are disabled when
the target variable is reduced below a lower threshold value.
The dosing of the reagents is constant when the reagents are
enabled. In advanced mode, the thresholds remain the same
but the dosing is exponential feedback controlled to rapidly
increase the dosage when the targets are significantly above
the threshold values.
The next step in this part of the control system is to uti-
lize techniques from machine learning to optimize the on/
off thresholds and dosing targets. This will move the system
past simply ensuring concentrate quality is maintained into
directly using these reagents to optimize the overall recover-
ies and plant performance.
The first bulk reagent to be addressed by the expert
control system is copper sulfate. Copper sulfate is used in
the Zn flotation to activate the Zn minerals (overwhelm-
ingly sphalerite), and the presence of sufficient copper sul-
fate can be strongly detected based on the bubble size. This
is a parameter that is detected by the VisioFroth ™ system.
Thus, copper sulfate is added when Zn is being lost to the
tails, but not if the bubble size is already small and indica-
tive of sufficient copper sulfate being present.
The air and MIBC for each part of the flotation cir-
cuit are also controllable. For these the primary param-
eter is bubble texture, another soft property reported by
the VisioFroth ™ system. It is used as a reliable indicator of