Automatic Control of Flotation Columns - Advanced Control and Supervision of Mineral Processing Plants - page 262

tem was submitted to a series of step-changes in wash water, tailings and air flow
rates. Based on continuous conductivity measurements, the ANN provided estimates
of bias and entrainment as a function of time. To verify these predictions, indepen-
dent bias measurements were made through a reference method when the column
was operating in steady-state conditions, before and after a given disturbance was
introduced. ANN predictions for both bias and entrainment were very good.
However, the method has some drawbacks such as the need for extensive ex-
perimentation to constitute the database required for both learning and validation
steps of the ANN and the need for an independent and reliable method of measur-
ing the bias rate necessary for the ANN training database; this can easily be done
on a laboratory scale by a steady-state water balance, but the method has not been
attempted on an industrial scale. As a result of the above mentioned procedure, i.e. ,
ANN training using steady-state bias values, it has not yet been confirmed whether
the method would satisfactorily predict transient bias values.
More recently, Aube [29] used a simple multilinear regression technique to re-
place the ANN algorithm for modeling the bias rate to the conductivity profile. The
steady-state validation results were quite satisfactory, eliminating at least some of
the drawbacks outlined above.
Despite the above mentioned problems, both the ANN and multilinear regres-
sion approaches have been successfully used for automatic control of bias in a
laboratory-scale column on various occasions [21], [30], [31].
In order to solve the remaining concerns of the previously discussed techniques,
Maldonado et al. [32] proposed a method which combines the positive features of
each of them: process phenomenology and some empirical correlations. The pro-
posed method was developed for a water-air mixture only, although work is un-
derway for its extension to three-phase systems. In contrast to methods based on
external (water component) flow rate balances, this one gets information directly
below the interface, i.e. wherebiasisdefined. More precisely, it uses a linear rela-
tionship between the bias rate and the volumetric fraction of wash water (ε ww )below
the interface, estimated by applying the “additivity rule” (Figure 6.4(a)). The con-
ductivity profile is obtained using the same sensor designed for froth-depth sensing
[24]. Validation results are shown in Figure 6.4(b).
One important advantage of this method is the fact that it can dynamically esti-
mate the bias rate independently of the existence of steady-state conditions. As such,
the method has been successfully implemented in different control applications on
a pilot scale [33, 34]. The proposed method improved the stability of the control
system compared with those control strategies using bias rate estimates obtained as
the difference between tails and feed rate water components [35, 36].
Figure 6.5 shows the dynamic bias estimates (smoother sequence), during step-
changes in wash water rate (the variable used to control bias rate), and bias values
obtained from a water balance in the collection zone (noisy sequence). During the
validation test, feed rate was regulated at a constant value using a local PI controller
and the froth depth was maintained at a fixed value also using a PI controller, by
manipulation of tailings flow rate. Despite the fact that bias rate estimation is based
on a steady-state water mass balance in the collection zone, it is intentionally plotted