Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 177

170

160

150

Power/50 (kW)

140

Ore hold−up (t)

130

120

Return flow (t/h)

110

100

90

0 2 4 6 8 10 12 14 16 18

time (h)

Figure 4.9 Model response of ore hold-up, mill power draw and return ore flow from screen when
fresh feed ore flow changes from 1 . 150 to 1 . 265 ton/h at t = 8 h [22]

ables, but also disturbances. In general, control strategies become simpler as more
measurements of plant variables become available. For example, if a fast acting
random disturbance is measured, it may be used in a relatively simple feedforward
control scheme to compensate its effect upon the controlled variable. But some sen-
sors used in mineral processing plants fail, or have to be removed for maintenance
or repairs relatively more frequently than in other industries. The options are:

1. switch to a complex control strategy which does not include the signal of the now
unavailable disturbance measurement;
2. try to somehow continue using a manual or automatic control without the impor-
tant information provided by the failed sensor;
3. use an estimation or prediction of measurements delivered by a soft sensor, based
on other correlated measurements, and thus maintain the satisfactory simple con-
trol scheme [3].
Therefore soft sensors provide a means for graceful degradation of the control
performance when sensors fail. But the partial modeling involved in soft sensor de-
sign may cause unexpected control performance when soft sensors are used in auto-
matic control loops because of two circumstances: (a) change of the plant-controller
model structure introduced because of the soft sensor model; and (b) parameter
coupling between the unmodeled and modeled part of the plant because of (i) in-
herent coupling of input/output model parameters due to changes in parameters of
the state/state-output plant model, and (ii) a disturbance causing changes in param-
eters of both the modeled and unmodeled part of the plant. Among the components
selected by the structure determination method measurements usually appear that
are not necessarily inputs to the plant (controls or measured disturbances), but other