Soft Sensing - Advanced Control and Supervision of Mineral Processing Plants - page 172

4.2.4.4 Soft Sensors Based on the Kalman Filter
Soft sensor models using Kalman filters have been used when the phenomenological
knowledge about the plant allows a convenient state-state/output model to be built
[20, 21, 29, 39]. A soft sensor has been tested using data from a pilot semiautoge-
nous grinding mill in which the ore contents comprise 27 size fractions which are
states [20], with aid from the phenomenological model [20, 55]. In this case a low
dimension extended (nonlinear) Kalman filter (EKF) was used, since a linear state
model had a very small region of validity. The estimated states were water, total
mill ore content and the fine ore (of sizes less than the grate opening) mill content,
grinding rate, water and ore discharge rates, and fine ore discharge flow. The water
discharge rate, the ore discharge rate and the grindability parameters were converted
to states so they could be estimated by the EKF. The fine ore discharge flow was also
converted into a state by turning into the independent variable of a first order equa-
tion with input depending on the mill ore contents. The measurements considered
for the state/state-output equations were total ore, fine ore and water feed flows, total
mill weight, total ore discharge flow, water discharge flow, and mill power draw.
Apelt et al. [21] use a much more elaborate Kalman filter equation with 36 states
of which 27 represent the mill ore contents by size fractions (as in [20]), and the
rest are the ball charge, the SAG mill weight, the water content, and parameters.
Good results are reported when the ore flows per each of the 27 ore size intervals
are measured instead of bulk ore discharge flow measurement. If only the bulk ore
discharge flow measurement is considered, results are acceptable for the most part.
Time [min]
Figure 4.8 Test of soft sensor designed using PLS. Concentrate grade sensor measurement (fine
line) and soft sensor output (broad line)