Minerals Engineering 15 (2002) 10431053 This article is also available online at: www.elsevier.

com/locate/mineng

Inferential measurement of SAG mill parameters II: state estimation

T.A. Apelt
a b

a,c,*

, S.P. Asprey a, N.F. Thornhill

a,b

Centre for Process Systems Engineering, Imperial College, London SW7 2BY, UK Department of Electronic and Electrical Engineering, University College, London WC1E 7JE, UK c Department of Chemical Engineering, University of Sydney, NSW 2006, Australia Received 21 June 2002; accepted 20 September 2002

Abstract This paper discusses the combined state and parameter estimation of SAG mill inventories and model parameters. Recognised simulation models are utilised for the rock and water charge state equations. New models of the ball charge and mill shell lining states are presented. The 36 state system is detectable although not completely observable. Five ore grindability and mill discharge grate parameters augment the state system. One mill weight and two discharge measurement models are presented and utilised in two state estimation formulations. Results indicate that a size by size SAG mill discharge measurement model provides superior state estimates and improved discharge grate parameter estimation compared to a bulk SAG mill discharge measurement model. 2002 Elsevier Science Ltd. All rights reserved.
Keywords: SAG milling; Comminution; Modelling; Simulations

1. Introduction The SAG milling process presents measurement and control problems typical of an industrial process in that the total state vector can seldom be measured and the number of outputs is much less than the number of states. State estimation techniques may be utilised to provide acceptable estimates of all the state variables (even those not directly measured) in the face of measurement error and process disturbances (Ray, 1981). The estimates may be used for process monitoring, operation and control. This paper discusses combined state and parameter estimation (CSPE) for SAG mill charge levels and is a continuation of earlier work (Apelt et al., 2001a) which described inferential models for the mill inventories in detail. The formulation of the CSPE problem will be described following a review of related works. State equations for mill rock, water and grinding ball inventories and the shell protective lining are then presented with a list of ve parameters included in the formulation. Discussion of system observability and detectability follows the presentation of the measurement

models incorporated in the CSPE formulation. Discussion of the requisite Kalman lter tuning precedes the presentation and discussion of the results. To conclude, the major ndings of this paper are summarised. 2. Circuit description The discussion centres on the primary grinding circuit shown in Fig. 1. Ore is fed to the SAG mill for primary grinding. The mill discharge is screened with the oversized material recycling via a gyratory cone crusher, and the screen undersize being diluted with water and fed to the primary cyclones for classication. Primary cyclone underow is split between a small recycle stream to the SAG mill feed chute and a ball mill feed stream. The primary grinding circuit products are subjected to further size reduction (ball mill), classication (cyclones) and separation (ash otation) in the secondary grinding circuit. Further details of the grinding circuit and the other sections of the processing plant may be found elsewhere (Apelt et al., 1998, 2001a,b; Freeman et al., 2000; Apelt, 2002). 3. Related works

Corresponding author. Address: Centre for Process Systems Engineering, Imperial College, London SW7 2BY, UK. E-mail address: t.apelt@ic.ac.uk (T.A. Apelt).

Signicant research has been conducted in the use of Kalman lters for comminution process state estimation.

0892-6875/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 8 9 2 - 6 8 7 5 ( 0 2 ) 0 0 2 2 9 - 7

1044

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

Fig. 1. Primary grinding circuit.

Much of the research has been conducted by J.A. Herbst and colleagues whom, over time, have had association with Utah Comminution Centre, Control International (Inc.), GS Industries, JA Herbst and Associates, Svedala (Optimization Services and CISA) and Metso Minerals. Herbst and Alba (1985) discussed the use of a Kalman lter to estimate the current state, model parameters and the predicted state (in the next time step) for adaptive control. Herbst and Gabardi (1988) used a Kalman lter in conjunction with the use of a lifter-bolt strain gauge (for charge position indication), mill powerdraw measurement, and process models of the grinding media and mill shell lining, to estimate the ball charge level and wear rate, liner thickness and wear rate in a ball mill. The result was a consistent ball charging rate for ball charge control at maximum mill powerdraw. Kalman lters were applied to autogenous/semiautogenous (AG/SAG) mills by Herbst et al. (1989) when estimates of rock (one combined state), ball and water charges, shell liner weight and ore grindability were obtained from a dynamic model. Mill bearing pressure and powerdraw measurements were used. Other work of Herbst and colleagues has documented industrial applications and development of a commercial soft-sensor (Herbst and Pate, 1999; Herbst and Pate, 2001). A Kalman lter has been used on an iron ore pebble mill (Herbst and Pate, 1996). A commercial software product was installed to estimate ve states (mill holdup of two rock states (55 mm), water, grinding balls and shell liner weight) and two parameter estimates (ore grindability and charge angle of repose) (Herbst and Associates, 1996). Herbst and Pate (1999) describe a generic softsensor with examples of ore grindability estimation for a ball mill and estimation of mill lling (ore, balls and water), dynamic angle of repose and ore grindability for a SAG mill. The Svedala Cisa OCS optimising control system package for AG/SAG mills includes a softsensor module (Broussaud et al., 2001). Applications on both minimally and comprehensively instrumented plants are described. The softsensor continuously computes a mass

balance and estimates mill charge (mass of solids, simplied size distribution and percent solids in the mill), cyclone feed, circulating load and cyclone overow (particle size and percent solids). The review shows that the technology of Kalman lters in SAG mill state estimation is well progressed. However, the industrial uptake of state estimation has been slow and isolated to larger mining houses and new installations. The lag between development and uptake is a reection of the level of uptake of model based decision making control technologies. Herbst (2000) suggested that industry reluctance is due to perceived gap between the technology cost and its benets and also the performance record of the technology. Therefore according to Hodouin et al. (2001) the area should be considered active with some key issues to be addressed. This assessment is reinforced by the following points: Broussaud et al. (2001) considered the diculty of determination of the ball load and wear rate in a SAG mill in the absence of a suitable commercial ball addition system. They concluded the future combination of Continuous Charge Monitoring (CCM) instrumentation and good control of ball addition should allow a closer on-line optimization of the ball load and further improve SAG mill performance. The Herbst and Pate application of a Kalman lter for the estimation of mill rock and ball hold-up is recognised in a review of automation in the minerals processing industry. Jamsa-Jounela (2001), for instance, highlighted engineering costs, lack of precedent applications and limited control technology as constraints. Research into state estimation for SAG mills is also currently being progressed at the Julius Kruttschnitt Mineral Research Centre where a Kalman lter is utilised to predict ore hardness, mill total charge and mill discharge factors (Schroder, 2000). An open-loop plant trial gave good correlation and tracking performance. The contribution of the work reported here is to present further examples of state estimation for SAG mills and novel models of the SAG mill discharge, SAG mill ball charge and shell protective lining. State estimation of mill rock charge, ball charge (Jb ), water charge and thus total charge (Jt ) is demonstrated. An assessment of structural properties of the models (their observability and detectability) gave insights into 50 their relative performance when used with a Kalman Filter. CSPE for SAG mills is now discussed further with the presentation of the state equations and parameters. Two formulations are presented which include process measurement equations for mill weight and mill discharge.

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

1045

Both are based on a University of Sydney thesis dissertation (Apelt, 2002).

4. Combined state and parameter estimation model Continuous-time nonlinear systems can be described as follows (Henson and Seborg, 1997): _ xt f x; u; h; t _ h0 yt gx; h; t 1 2 3

b discrete-time system function, P kjk is the propagated b estimated state covariance matrix, P k1jk is the estimated state covariance matrix at the next time step time k 1Dt, Q is the process output (measurement) covariance matrix, R is the state covariance matrix. 5. State equations and parameters This research models the SAG milling process by the following 36 state equations and ve parameter equations: Solids. The size by size solids mass balance developed by the Julius Kruttschnitt Mineral Research Centre (Napier-Munn et al., 1996; Valery and Walter, 1998): Accumulation Inflow Outflow Generation Consumption
i1 X dsi fi pi rj sj aij 1 aii ri si dt j1

_ where x is the state of the system, x is the time derivative of state of the system, u is the system input, y is the system output, t is time, h is the system model para_ meters to be estimated, h is the time derivative of the system model parameters, f is the system function and g is the measurement function. Eqs. (1)(3) is referred to a CSPE model since it is utilised to estimate states x and parameters h. In discrete-time, where measurements are available at equally spaced intervals (sampling periods) Dt, Eqs. (1) (3) can be described as follows (Henson and Seborg, 1997):  xk1 f xk ; uk ; hk ; k 4 hk1 hk  yt gxk ; hk ; k 5 6

i 1; . . . ; 27 14

pi d0 ci si

i 1; . . . ; 27

15

 where k is the discrete time index (time kDt), f is the  is the discrete time discrete time system function and g measurement function. One method of on-line state and parameter estimation is achieved through the utilisation of an the extended Kalman lter on the system described by Eqs. (4)(6) (Henson and Seborg, 1997): x ^ xkjk xkjk1 Lk yk g^kjk1 ; hkjk1 ; k Lk
T T b b P kjk1 Gk Gk P kjk1 Gk

7 8 9 10 11 12

where fi is the SAG mill feed rate of solids in size i (t/h), ri is the SAG mill breakage rate for particles in size i which varies with operating conditions within the mill (Variable Rates) (h1 ), aij is the SAG mill appearance function which describes the distribution that particles in size i will form in the smaller sizes j (dimensionless), pi is the SAG mill discharge rate of solids in size i (t/h), si is the SAG mill rock charge in size i (t), ci is the SAG mill discharge grate classication function value for size i (dimensionless) and d0 is the SAG mill maximum discharge rate constant (h1 ). Water. The water mass balance also developed by the Julius Kruttschnitt Mineral Research Centre (NapierMunn et al., 1996; Valery and Walter, 1998): Accumulation Inflow Outflow dsw fw pw dt pw d0 sw 16 17

b b P kjk I Lk Gk P kjk1  x ^ ^ xk1jk f ^kjk ; hkjk ; uk ; k

T b b P k1jk F k P kjk F k Q  ox; h; k  g  Gk   ox ^ x^kjk1 ;hhkjk1 x    of x; h; u; k  Fk   ox ^

where fw is the SAG mill water feed rate (t/h), pw is the SAG mill water discharge (product) rate (t/h), sw is the SAG mill water charge (t) and d0 is the SAG mill maximum discharge rate constant (h1 ). Grinding balls. The new grinding ball mass balance model (Apelt, 2002) is as follows: Accumulation In Out Generation Consumption dbci bii bei bwi1 bwi i 1; . . . ; 7 18 dt where bci is the SAG mill ball charge for balls in size i (t), bii is the SAG mill ball feed rate for balls in size i

13

x^kjk ;hhkjk ;uuk x

^ where xkjk is the ltered estimate of state x at time kDt, ^ xkjk1 is the prediction of state x at time kDt from the previous time step k 1Dt, Lk is the Kalman lter gain at time kDt, I is the identity matrix, Gk is the linearised, discrete-time measurement function, F k is linearised,

1046

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

(t/h), bei is the SAG mill ball charge ejection rate for balls in size i (t/h) and bwi is the SAG mill ball charge wear rate for balls in size i (t/h). Shell lining. The new SAG mill liner weight mass balance model (Apelt, 2002) is described as follows: Accumulation Wear d SMIW wearate dt 19

bii

p 6

Db 1000

3 SGb bps

3600 bst

22

where bps is balls per stroke, bst is the ball stroke time (s), Db is new ball diameter (mm) and SGb is ball specic gravity (t/m3 ). Stroke refers to the stroke rate of the feeding ram. 5.1.2. Ball wear The overall ball wear rate may be determined from operating data. If the ball charge level is being maintained at a constant level, the ball wear rate is equal to the ball feed rate plus ball ejection. For the larger sizes where there is no ejection the wear rate is equal to the ball feed rate. The overall ball wear rate translates to ball wear rates by size. These ball wear rates by size (bwi ) are proposed here to be proportional to the fractional surface area and the ball mass in size i and inversely proportional to ball hardness, i.e., bwi / fSAi 1 SMBCi HBi 23

where SMIW is the SAG mill protective shell lining installation weight (t) and wearate is the SAG mill shell protective lining wear rate (t/h). Parameters. These state equations are augmented by the following set of ve parameter equations:
_ A0 _ b0 _ ta 0 impact breakage ore parameter in aij impact breakage ore parameter in aij abrasion breakage ore parameter in aij

_ fp 0 relative fraction pebble port open area in pi _ d0 0 maximum mill discharge rate coefficient in pi 20

Ore breakage parameters A, b and ta are included in anticipation of an inferential measurement of ore grindability. The mill discharge grate parameter fp is included due to its close link to the pebble port diameter, xp , which is an inuential to the relative contribution to error in the feed passing sizes (F80 . . . F20 ). The maximum discharge coecient parameter, d0 , is included since it aects not only the mill discharge but also the rock and water charge remaining in the mill. A mill discharge factor parameter has also been used elsewhere (Schroder, 2000). Tallying the number of states and parameters brings the order of the system function, f , to 41. 5.1. Ball charge model The dynamic ball charge model proposed by this research is as follows: Accumulation In Out Generation Consumption dbci bii bei bwi1 bwi 21 dt where bci is the mass of balls in ball charge of size i (t), bii is the feed balls in size i (t/h), bei is the balls of size i ejecting from the mill i (t/h) and bwi is the mass of balls wearing out of size i into size i 1 (t). 5.1.1. Ball feed The ball feed to the SAG mill, bii , can be determined from operating conditions. Assuming that the feed balls are of a single diameter, Db , the ball feed is as follows:

where bwi is ball wear of grinding balls in size i (t/h), fSAi is fractional surface area of grinding balls in size i (fraction), HBi is Brinnell Hardness of grinding balls in size i (N/m) and SMBCi is mass of grinding balls in size i (t). The total surface area of grinding balls in size i, SAi , is the product of the number of balls in size i, Ni , and the surface area of a ball of size i, i.e., SAi Ni pDb2 i 24

The number of balls in size i, Ni , is determined from the mass fraction of the total ball charge in size i and the mass of a ball of size i, i.e.,
smbci BC Ni 100 3 Dbi p SGb 6 1000

25

where BC is total ball charge mass (t), Dbi is diameter of ball of size i (mm) and smbci is mass percent of balls in of size i (%). The fractional surface area of the ball charge in size i, fSAi , is
i SAi fSAi Pn Pn Dbismbci i1 SAi i1 Db i

smbc

26

The ball hardness model proposed here is based on the ndings of Banisi et al. (2000), i.e., that the ball hardness of 80 mm balls drops signicantly when the ball wears to less than 65 mm in size (%81% original size). In this work the original ball diameter is 125 mm and it is assumed that the nominal ball diameter at which hardness decreases markedly is 95 mm (76% original size). Assuming hardness of 450 and 250 Brinnell for the outside and inner layer of the balls, respectively,

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

1047

(estimated from data in Perrys (Perry et al., 1984)) and that the variation of hardness across ball diameter can be described by a Whiten classication model type relationship (Whiten, 1972; Napier-Munn et al., 1996), the ball hardness, HBi , can be described as follows:
for Dbi 6 K1  K3 K2 Dbi HBDbi 450 K2 K1 HBDbi 450 for Dbi P K2 27 HBDbi 250

eciency to undersize since the ejected balls are screen undersize. !k ! Dbi 30 Edci 1 exp N foag 1 xp where N 1 and is a discharge grate eciency parameter, foag is discharge grate fraction open area (fraction), Dbi is ball diameter of size i (mm), xp is discharge grate pebble port size (mm) and k 2 and is a discharge grate eciency parameter. The model is validated against operating data by tting the ball wear model parameters (BWki ) and the ball ejection model parameters (BEki ). 5.2. Shell lining model The dynamic model of the SAG mill shell protective lining proposed by Apelt (2002) is presented here. The SAG mill installation weight, SMIW, can be considered a sum of a number of constituents: SMIW shell lining D=C grate 31 where SMIW is SAG mill installation weight (t), shell is SAG mill shell weight (t), lining is SAG mill shell lining weight (t), D/C grate is SAG mill discharge grate weight (t). The mill shell remains intact throughout the operational life of the mill. Therefore, the mill shell weight (shell) is a constant. The lining is the internal shell protective lining and is subject to wear through direct contact with the mill contents. The SAG mill discharge grate is internal to the mill also and is subject to wear. Periodic change out of the mill protective lining and discharge grate occurs to accommodate the wear of these internal components. From plant experience, the shell lining wear occurs predominantly at discharge end. Therefore, the lining term can be broken up into a feed end and a discharge end term: SMIW shell liningFE liningDE D=C grate 32 where liningDE is the weight of the discharge end of the shell lining (t), liningFE is the weight of the feed end of the shell lining (t). Combining the feed end lining terms and a constant (to accommodate mill weight instrument oset) into a SAG mill weight constant, SMWconst (t), allows the model to we rewritten as follows: SMIW SMWconst liningDE D=C grate 33 The dynamic SAG mill liner wear model may then focus on the grouped mill discharge end terms and may be written as follows:

for K1 < Dbi < K2

where K1 is the ball size below which hardness equals 250 Brinnell (mm), K2 is the ball size above which harness equals 450 Brinnell (mm) and K3 2:3 and is the ball hardness curve shape parameter. The ball hardness model is illustrated in Fig. 2. The model can be adjusted to suit a given set of operating conditions by the introduction of a ball wear coecient, BWki , which can be tted to operating data. For mass balance consistency, the units of the ball wear coecient are (Brinnel/h). The ball wear model can now be stated as follows: bwi BWki fSAi 1 SMBCi HBi 28

5.1.3. Ball ejection The model of ejection of balls from the SAG mill proposes that the SAG mill discharge grate behaves as a vibrating screen which can be modelled by an eciency to oversize model (Napier-Munn et al., 1996). Ball ejection of size i, bei , can then be stated as follows: bei BEki Edci SMBCi 29 where Edci is ball ejection eciency to discharge of size i (fraction), bei is ejection rate for balls of of size i (t/h), BEki is ball ejection model tting parameter (h1 ). The eciency model utilised here is taken from Napier-Munn et al. (1996) and is expressed in terms of

Fig. 2. Ball hardness model.

d SMIW wearate dt

34

1048

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

where wearate is the SAG mill shell wear rate (t/h). Integrating Eq. (34) with respect to time, t, yields SMIW SMIW0 wearate t 35 where SMIW0 is the initial SAG mill installation weight (t) and t is time (h). 5.2.1. Wear rate The mill liner wear rate, wearate, can determined from the change-out frequency and the relative change in weight of the discharge grate and discharge end shell lining at change-out time. Based on typical plant experience, the change-out frequency is approximately 6 weeks, or 1008 h (6 7 24), and the change-out liner weight is approximately half the liner installation weight, i.e., liningDE0 D=C grate0 36 1008 where liningDE0 is the installation weight of the discharge end of the shell lining (t) and D=C grate0 is the installation weight of the discharge grate (t). The shell lining is a series of alternating low lifter and high lifter bars separated by shell lining segments. The installation weight of the discharge end shell lining, liningDE0 , is calculated as follows: wearate 2 1 lbw hlt slt liningDE0 Lsm qliner nhl 2 1000 1000 1 lbw llt slt Lsm qliner nll 2 1000 1000 p slt slt Lsm qliner Dsm0 2 1000 1000
1

similar to pieces of pie), foag is the fraction grate open area (fraction) and Dtsm is the SAG mill trunnion diameter (m). In Eq. (38), the rst term represent the weight of the large at disc (that constitutes the discharge grate) of thickness, dgt, less the apertures in the grates (fraction open area foag) and the absent central piece of diameter Dtsm . The second term in Eq. (38) represents the weight of the portion of the discharge end lifter bars that protrude above the surface of the discharge grate. The discharge grate lifter bars are of thickness (height) hlt (mm) and of length (Dsm Dtsm ) (mm). 5.2.2. Shell thickness The assumption that all mill lining wear occurs in the discharge end of the mill allows the mill liner model to be simplied. Subtracting the mill weight constant (SMWconst) from both sides of Eq. (35) yields liningDE D=C grate liningDE D=C grate0 wearate t 39

Assuming that wear is uniform throughout the discharge end of the mill, the wear thickness, wt (the amount of lining component that has been worn away) (mm) can be determined as follows: 1. Take all terms in Eq. (39) to one side of the equation 0 liningDE D=C grate0 wearate t

!2 ! 37

liningDE D=C grate

40

where qliner is the liner density (t/m3 ), nhl is the number of high lifter bars, nll is the number of low lifter bars, lbw is the lifter bar width (mm), hlt is the high lifter bar thickness (height) (mm), llt is the low lifter bar thickness (height) (mm), slt is the shell liner thickness (height) (mm) and Lsm is the SAG mill shell length (m). In Eq. (37), the third term represents the weight of an annular piece of shell lining of thickness slt, dened by the shell inside diameter, Dsm (m), and extending to half the mill length, Lsm =2 (m). The rst and second terms are the weight of portion of high and low lifter bars, respectively, that protrude above the shell lining. The installation weight of the discharge grate, D=C grate0 (t), is calculated as follows:  2  D dgt Dt2 dgt 1 foag p sm D=C grate0 qliner p sm 4 1000 4 1000   lbw hlt slt qliner ndg Dsm Dtsm 1000 1000 38 where dgt is the discharge grate thickness (mm), ndg is the number of discharge grate segments (conceptually

2. Express (liningDE D=C grate) in terms of wear thickness, wt, by the substitution of (slt wt) for slt in Eqs. (37) and (38): liningDE D=C grate 1 lbw hlt slt Lsm qliner nhl 2 1000 1000 1 lbw llt slt Lsm qliner nll 2 1000 1000 !2 ! p slt slt Lsm qliner Dsm0 2 1000 1000  2  Dsm dgt Dt2 dgt sm 1 foag p qliner p 4 1000 4 1000   lbw hlt slt qliner ndg Dsm Dtsm 41 1000 1000 3. Solve Eq. (40) for wear thickness, wtthe root of the equation, e.g., with fzero, the MATLAB scalar nonlinear zero nding function. 4. Determine the current lining thickness, (slt wt) (mm). The state equations required for the combined state and parameter estimation formulation have now been

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

1049

presented, i.e., dynamic mass balance models of the rock (Eqs. (14) and (15)), water (Eqs. (16) and (17)), and ball charges (Eq. (18)) and the shell protective lining (Eq. (19)). These equations form the matrix of state equations that describe the state of the system, Eq. (1). The model parameters of the system, Eq. (2), have also been presented (Eq. (20)). In the following section, the models of the system output (process measurements) (Eq. (3)) are presented. 6. Measurement models The measurement (process output) models typically utilised in CSPE are based on mill powerdraw and mill weight (by way of bearing pressure or load cell measurement) (Herbst and Pate, 1999; Herbst and Pate, 2001). The SAG mill discharge screen oversize conveyor measurement has also been utilised in a formulation that estimates a mill discharge factor parameter (Schroder, 2000) which is assumed to be or related to the maximum mill discharge rate coecient, d0 (h1 ). 6.1. SAG mill weight measurement model The SAG mill weight is measured by load cell. This measurement, y1 t, is the process output that is paired with the mill weight measurement model, g1 x; h; t, which is a sum of the mass of ore particles, grinding balls and water in the mill charge kidney (the charge material that is not in free-fall from the charge shoulder to the charge toe), Mkidney , mass of the mill shell discharge end lining and the discharge grate, (liningDE D=C grate), as presented above, a calibration term, tare, that allows for the dierence between the actual total mill weight and the load cell measurement (includes the mill shell weight and the feed end shell lining weight), g1 x; h; t Mkidney liningDE D=C grate tare 42 The state functionality of the SAG mill weight measurement model, g1 x; h; t is via the rock, ball and water charge components in the charge kidney and the discharge-end shell lining weight. The latter is, in fact, one of the states considered. The kidney mass is a state function, Mkidney x; h; tit is the sum of the rock, ball and water states in the charge kidney. 6.2. SAG mill discharge measurement model 6.2.1. Model 1bulk ow model Model 1 is a bulk ow model which estimates the total volumetric discharge rate of the SAG mill. The

SAG mill discharge stream volumetric owrate can be reconstructed from four available plant measurements and then utilised in the CSPE formulation as the second process output, y2 t as described by Apelt et al. (2001b), y2 t MVOSCF %solsOSCF 100SGs MVPCFDm3 MVPCFD%sols SGl MVPCFD%sols SGl 100 MVPCFD%sols SGs MVPCFDm3 100 MVPCFD%sols SGs MVPCFD%sols SGl 100 MVPCFD%sols SGs MVOSCF 100 %solsOSCF MVPCFWm3 100SGl 43 where MVOSCF is the oversize crusher total feedrate (t/h), MVPCFDm3 is the primary cyclone feed owrate (m3 /h), MVPCFD%sols is the primary cyclone feed density (% solids w/w), MVPCFWm3 is the primary cyclone feed water addition owrate (m3 /h), %solsOSCF is the oversize crusher feed density (% solids w/w), SGs is the ore specic gravity (t/m3 ) and SGl is the process water specic gravity (t/m3 ). The plant measurement, Eq. (43), requires pairing with a measurement model, according to Eq. (3). The corresponding measurement model, g2 x; h; t, is g2 x; h; t SMDCs SMDCl SGs SGl 44

The state functionality of the SAG discharge measurement model, g2 x; h; t, is through the constituents of the terms in Eq. (44). The solids mass owrate, SMDCs (t/h), is the summation of the size by size mill product stream and is thus state dependent. SMDCs
n X i1

pi

n X i1

d0 ci si

45

where pi is the SAG mill discharge rate of solids in size i (t/h) si is the SAG mill rock charge in size i (t) and ci is the SAG mill discharge grate classication function value for size i (dimensionless). The liquid mass owrate, SMDCl (t/h), is a function of SMDCs and thus is also state dependent:   SMDCs SMDCl kg Qm 46 SGl SGs where Qm is the SAG mill volumetric discharge rate through the grinding media (m3 /h) and kg is a coarse material adjustment factor (dimensionless). On account of the SAG mill discharge grate characteristics, i.e., high fractional open area (foag), high relative radial position of the open area (c), and high relative radial position of the outermost aperture (rn ), the mill discharge ow is assumed to be only through the grinding media and that no slurry pool exists at the toe

1050

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

of the charge. Therefore, the mill discharge owrate may be calculated as follows: Qm
2 6100Jpm c2:5 A/1:38 D0:5

^ xt Ax Bu yx Cx

49 50

47

where A is the total discharge grate open area (m2 ), D is the mill inside diameter (m), c is the mean relative radial position of open area (fraction) and / is the fraction critical mill speed (fraction). The nett fractional holdup of slurry in mill that is contained within the grinding charge (live area) interstices, Jpm , is calculated as follows: P16 mm si sw in SGl SGs 48 Jpm Vmill where sw is the mill water charge (t) and Vmill is the mill volume (m3 ). The parameter functionality of the SAG discharge measurement model, g2 x; h; t, is also through the constituents of the terms in Eq. (44). The solids mass owrate (SMDCs ), Eq. (45), is a function of the maximum mill discharge rate constant (d0 ), explicitly, and also the relative fraction pebble port open area (fp ), implicitly (via the grate classication function, ci ). The maximum mill discharge rate constant (d0 ) aects not only the mill discharge but also the rock and water charge remaining in the mill. The relative fraction pebble port open area (fp ) is linked by the classication function to the pebble port aperture size (xp ) which is an inuential parameter in the feed passing size estimates. This reinforces the inclusion of the relative fraction pebble port open area (fp ) in the list of parameters in Eq. (20). The ore breakage parameters A, b and ta aect the rock breakage occurring within the mill and the mill rock charge fractions (si ) and are therefore implicitly present in the state equations, Eqs. (14) and (15), and the measurement model function, Eq. (45). As mentioned, these parameters are included in the formulation in anticipation of an inferential measurement of ore grindability which inuences mill performance. The mill weight measurement y1 t and measurement model g1 x; h; t, Eq. (42), and the reconstructed mill discharge measurement y2 t, Eq. (43), and measurement model g2 x; h; t, Eq. (44), were paired together to form the system output and measurement functions (Eq. (3)). 6.2.2. System observability and detectability Initial diculties in generating results utilising this formulation (Model 1) prompted an investigation of the observability and detectability of the system. This was achieved by: 1. Linearising the CSPE at steady-state conditions to generate a state-space model of the form

where A; B; C are the system matrices, x is the system states (36 states), u is the process inputs and y is the process outputs (measurements). 2. Generating the observability matrix, L0 , (Ray, 1981; Henson and Seborg, 1997) L0 C T jAT C T jAT C T j jAT
2 n1

CT

51

Matrix A is n n where n is the number of states and matrix C is l n where l is the number of measurements. The observability matrix, L0 , has dimension n nl. 3. Determining the rank of the observability matrix, L0 . If the rank of L0 is n then the system is completely observable and each initial state x0 can be determined from knowledge of the process inputs, u, and process outputs, y, over a nite time period. If the rank of L0 is less than n then the system is only partially observable. If the system modes that cannot be observed or reconstructed from the output measurements are stable then the system is detectable (Henson and Seborg, 1997). The rank of the observability matrix, L0 , was determined to be 8 which is less than the dimension of the system (n 36). Therefore, the system is not completely observable. The model was therefore further enhanced as described in the following subsection. 6.2.3. Model 2size by size model The issue of primary concern in utilising the bulk ow model was the low number of plant measurements and measurement models (2) relative to the number of states in the system (36). Therefore, the number of plant measurements was increased through the extension of the SAG mill discharge bulk ow model to a size by size model. Twenty eight new plant measurements, y2...29 x; h; t, are generated by as follows:   oscf i y228 x; h; t OSCFtph s 100   pcfdi PCFDtph s i 1; . . . ; 27 100 52 y29 x; h; t OSCFtph l PCFDtph l PCFW SGl 53 where oscf i is the oversize crusher feed retained in size i (% w/w), pcfdi is the primary cyclone feed retained in size i (% w/w), OSCFtph s is the oversize crusher solids feedrate (t/h), OSCFtph l is the oversize crusher liquid feedrate (t/h), PCFDtph s is the primary cyclone solids

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

1051

feedrate (t/h), PCFWtph l is the primary cyclone liquid feedrate (t/h), PCFW is the primary cyclone feedwater addition (m3 /h) and SGl is the process water specic gravity (t/m3 ). Eq. (52) is a size by size SAG mill discharge process measurement and represents the solids discharge rate for size i (t/h). Eq. (53) is SAG mill discharge process measurement for process water and the liquid discharge rate (t/h). The corresponding measurement models, g2...29 x; h; t, are the right hand sides of Eqs. (15) and (17), i.e., g228 x; h; t d0 ci si g29 x; h; t d0 sw i 1; . . . ; 27 54 55

Table 1 CSPE model resultsstates and parameters State Unit SS value Model 1 estimate Error (%) Model 2 estimate 0 0.23 3.54 8.91 9.33 7.66 5.26 2.43 1.39 0.83 0.60 0.53 0.51 0.48 0.42 0.36 0.30 0.27 0.26 0.27 0.27 0.27 0.25 0.22 0.20 0.17 0.72 2.06 0 0.9 46.7 32.7 13.1 0 0 Error (%)

where si is the mill rock charge in size i (t), sw is the mill water charge (t), d0 is the maximum discharge rate constant (h1 ) and ci is the discharge grate classication function (fraction). The system matrix A for the size by size model (Model 2) is the same as that for the bulk ow model (Model 1). Since the measurement (process output) matrix C has changed with the utilisation of 29 measurements (instead of two), the observability matrix, L0 , has changed accordingly. The rank of the observability matrix, L0 , for Model 2 is 20 which, though a signicant improvement on eight, is (still) less than the dimension of the system (n 36). Therefore, the system remains not completely observable for the new CSPE formulation. However, improved state estimates were possible using the size by size model as described in the following section. 7. Results and discussion The system of states was shown to be detectable for both CSPE formulations (Model 1 and Model 2) despite the incomplete observability of the system. Detectability was established through achieving satisfactory results by appropriate selection of Kalman lter tuning parameters. The Kalman lter is described by Eqs. (7)(13). Suitable initial values for the tuning parameters P , Q and R were based on discussion points raised by Henson and Seborg (1997), Cheng et al. (1997) and Welch and Bishop (2001). The results of the CSPE models are given in Tables 1 and 2. The state and parameter estimates for both CSPE formulations are shown in Table 1 with the steady-state information and the respective errors relative to the steady-state values (initial state values, x0 ). Table 2 contains the corresponding volumetric total (Jt ) and ball (Jb ) charge estimates. 7.1. Model 1bulk ow model Table 1 contains the state and parameter estimates for Model 1the formulation incorporating the bulk

Rock charge by size (t) s1 (t) s2 s3 (t) s4 (t) s5 (t) (t) s6 s7 (t) s8 (t) s9 (t) (t) s10 s11 (t) s12 (t) (t) s13 s14 (t) s15 (t) s16 (t) (t) s17 s18 (t) s19 (t) s20 (t) (t) s21 s22 (t) s23 (t) (t) s24 s25 (t) s26 (t) s27 (t) Water charge (t) sw Ball charge by size (t) bc1 bc2 (t) bc3 (t) bc4 (t) (t) bc5 bc6 (t) bc7 (t) Shell lining SMIW (t) Parameters A (dimensionless) b (dimensionless) (dimenta sionless) fp (fraction) D0 (h)

0 0.23 3.54 8.91 9.33 7.66 5.26 2.43 1.39 0.83 0.60 0.53 0.51 0.48 0.42 0.36 0.30 0.27 0.26 0.27 0.27 0.27 0.25 0.22 0.20 0.17 0.72 2.06 0 0.9 46.7 32.7 13.1 0 0 337.7 75.8 0.45 0.13 0.011 29.85

0 0.23 3.51 8.68 9.43 7.63 5.13 2.46 1.40 0.84 0.63 0.55 0.53 0.50 0.44 0.37 0.31 0.29 0.27 0.28 0.28 0.28 0.26 0.23 0.11 0.17 0.50 2.08 0 0.9 46.7 32.7 13.1 0 0 337.7 75.8 0.45 0.13 0.011 29.85

0 0.5 1.0 2.6 1.1 0.4 2.5 1.3 0.7 1.7 5.3 4.0 4.5 4.6 4.3 4.2 4.0 4.2 3.6 3.4 2.9 3.3 3.0 2.8 44.5 0.8 30.5 1.2 0 0 0.0004 0.0003 0.0008 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0.0004 337.7 0 0.1 0 0 0.001 75.8 0.45 0.13

0.034 208 28.79 3.6

ow SAG mill discharge measurement model (Eqs. (43) and (44)). Referring to the fourth and fth columns of Table 1, the results show that, overall, good state and parameter estimates are possible through the utilisation of this CSPE formulation using two plant measurements. The relative error present in the state estimates is

1052 Table 2 CSPE model resultscharge fractions Volumetric fraction Total charge, Jt Ball charge, Jb SS value 0.2298 0.1420 Model 1 estimate 0.2284 0.1420

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

Error (%) 0.6 0

Model 2 estimate 0.2298 0.1420

Error (%) 0 0

generally <5%. The relative error present in the parameter estimates is either exactly equal or approximately equal to zero. The poor observability of this model can be inferred from the two poor state estimate results that stand out in Table 1, the estimates of the 25th and 27th rock charge states, s25 and s27 , respectively. These relatively high levels of error (44.5% and 30.5%, respectively) are attributed to the small amounts of rock in these size fractions where a small deviation from the steady-state value therefore represents a large relative error. 7.2. Model 2size by size model Table 1 also contains the state and parameter estimates for Model 2the formulation incorporating the size by size SAG mill discharge measurement model (Eqs. (52)(55)). The results in the sixth and seventh columns show that improved state and parameter estimates are possible through the utilisation of this CSPE formulation with 29 plant measurements, which incorporates the size by size mill discharge models and possesses improved observability properties. The results show the extension of the bulk ow SAG mill discharge measurement, Eqs. (43) and (44), to a size by size throughput measurement, Eqs. (52)(55), improved the resulting state estimatesall state estimates displayed exact agreement with the steady-state values, see the seventh column of Table 1. Exact agreement was also displayed in the estimates of the ore impact breakage and abrasion parameters (A, b and ta ). Relative error is present in the estimates of the mill discharge parameters (fp and d0 ) in the seventh column of Table 1. The high level of relative error in the notional fraction pebble port open area relative to total grate open area, fp , is clearly evident. This is due to the CSPE model adjusting the mill discharge parameters (fp and d0 ) to achieve the exact agreement for the rock and water charge state estimates via Eqs. (54) and (55). The state estimates of Table 1 translate to the volumetric charge fraction estimates results in Table 2. Overall, all of the estimates show good agreement with the steady-state conditions and reinforce that both CSPE formulations (Model 1 and Model 2) can be utilised for state estimation. The volumetric ball charge (Jb ) estimates from both CSPE formulations exhibit exact agreement with the

steady-state conditions. The total volumetric charge (Jt ) estimates from both CSPE formulations exhibit good agreement with the steady-state conditions (to within %0.6%). The exact agreement between the Model 2 formulation estimates of the mill charge levels (Jt and Jb ) and the steady-state conditions is expected following the analysis of Tables 1 and 2 above. That analysis also leads to the expectation of the relative error present in the Model 1 total volumetric charge (Jt ) estimate. Although both CSPE formulations are capable of good estimates, the Model 2 formulation provides superior state estimates through adjustment (estimation) of the mill discharge parameters (fp and d0 ). On account of these two characteristics the Model 2 CSPE formulation is considered superior. In either case, further model validation would be required before entering an implementation phase. The results arising from a suite of inferential modelsmodels that infer unmeasured conditions from other measurementsto be presented in a future paper, indicate the error in the charge estimates from a mill weight-based inferential model is 1.3% and 1.8% for the estimates of the total volumetric charge (Jt ) and ball charge (Jb ), respectively. Comparison of these gures with the results in Table 2 (00.6% error) suggests that the CSPE formulations yield superior results.

8. Conclusions Two formulations of combined state and parameter estimation (CSPE) for SAG mills have been presented. They incorporated novel dynamic models of SAG mill ball charge and protective shell lining. Two novel measurement models of the SAG mill discharge, a bulk ow model and a size by size throughput model, and a novel mill weight measurement model were also presented. Assessment of system observability and detectability showed that observability was poor for formulation utilising the mill discharge bulk ow model. The subsequent development of the size by size throughput SAG mill discharge measurement model yielded better observability properties and improved state estimates. Both of the CSPE formulations, whilst not completely observable, are detectable. Both formulations would require further validation prior to implementation. The size by size SAG mill discharge measurement model (Model 2) yielded superior state estimate results and increased capacity to adjust the important discharge parameters (fp and d0 ).

Acknowledgements Acknowledgements go to Northparkes Mines for their assistance with and permission to publish circuit

T.A. Apelt et al. / Minerals Engineering 15 (2002) 10431053

1053

information, the Centre for Process Systems Engineering for part hosting and the University of Sydney for providing Australian Postgraduate Award funding for this research.

References
Apelt, T.A., Galan, O., Romagnoli, J.A., 1998. Dynamic environment for comminution circuit operation and control. In: CHEMECA 98, 26th Australasian Chemical Engineering Conference, CHEMECA, Port Douglas Qld, Australia. Apelt, T.A., 2002. Inferential measurement models for semi-autogenous grinding mills. PhD thesis, Department of Chemical Engineering, University of Sydney, Australia, submitted February 2002. See http://geocities.com/thomasapelt. Apelt, T.A., Asprey, S.P., Thornhill, N.F., 2001a. Inferential measurement of SAG mill parameters. Minerals Engineering 14 (6), 575591. Apelt, T.A., Asprey, S.P., Thornhill, N.F., 2001b. SAG mill discharge measurement model for combined state and parameter estimation. In: SAG 2001, Third International Conference on Autogenous and Semiautogenous Grinding Technology, vol. IV. UBC, Vancouver, BC, Canada, pp. 138149. Banisi, S., Langari-Zadeh, G., Pourkani, M., Kargar, M., Laplante, A.R., 2000. Measurement for ball size distribution and wear kinetics in an 8 m by 5 m primary mill of Sarcheshmeh copper mine. CIM Bulletin. Broussaud, A., Guyot, O., McKay, J., Hope, R., 2001. Advanced control of SAG and FAG mills with comprehensive or limited instrumentation. In: SAG 2001, Third International Conference on Autogenous and Semiautogenous Grinding Technology, vol. II. UBC, Vancouver, BC, Canada, pp. 358372. Cheng, Y.S., Mongkhonsi, T., Kershenbaum, L.S., 1997. Sequential estimation for nonlinear dierential and algebraic systemstheoretical development and application. Computers in Chemical Engineering 21 (9), 10511067. Freeman, W.A., Newell, R., Quast, K.B., 2000. Eect of grinding media and NaHS on copper recovery at Northparkes Mine. Minerals Engineering 13 (13), 13951403. Henson, M.A., Seborg, D.E. (Eds.), 1997. Nonlinear Process Control. Prentice-Hall/PTR. Herbst and Associates, J.A., 1996. Semi autogenous mill lling using a model-based methodology. Proposal submitted to North LimitedNorthparkes Mines. Herbst, J.A., 2000. Model-based decision making for mineral processinga maturing technology. In: Herbst, J.A. (Ed.), CONTROL 2000, Mineral and Metallurgical Processing. SME, Littleton, CO, USA, pp. 187200.

Herbst, J.A., Alba, F.J., 1985. An approach to adaptive optimal control of mineral processing operations. In: XV International Mineral Processing Congress, vol. 3. France, pp. 7587. Herbst, J.A., Gabardi, T., 1988. Closed loop media charging of mills based on a smart sensor system. In: Sommer, G. (Ed.), IFAC Applied Measurements in Mineral and Metallurgical Processing. Pergamon Press, Transvaal, South Africa, pp. 1721. Herbst, J.A., Pate, W.T., 1996. On-line estimation of charge volumes in semiautogenous and autogenous grinding mills. In: SAG 1996, Second International Conference on Autogenous and Semiautogenous Grinding Technology, vol. 2. Vancouver, BC, Canada, pp. 817827. Herbst, J.A., Pate, W.T., 1999. Object components for comminution system softsensor design. Powder Technology 105, 424429. Herbst, J.A., Pate, W.T., 2001. Dynamic modeling and simulation of SAG/AG circuits with MinOOcad: o-line and on-line applications. In: SAG 2001, Third International Conference on Autogenous and Semiautogenous Grinding Technology, vol. IV. UBC, Vancouver, BC, Canada, pp. 5870. Herbst, J.A., Pate, W.T., Oblad, A.E., 1989. Experience in the use of model based expert control systems in autogenous and semi autogenous grinding circuits. In: SAG 1989. First International Conference on Autogenous and Semiautogenous Grinding Technology. UBC, Vancouver, BC, Canada, pp. 669686. Hodouin, D., Jamsa-Jounela, S.-L., Carvalho, M.T., Bergh, L., 2001. State of the art and challenges in mineral processing control. Control Engineering Practice 9 (9), 9951005. Jamsa-Jounela, S.-L., 2001. Current status and future trends in automation of mineral and metal processing. Control Engineering Practice 9 (9), 10211035. Napier-Munn, T.J., Morrell, S., Morrison, R.D., Kojovic, T., 1996. Mineral Comminution CircuitsTheir Operation and Operation. Julius Kruttschnitt Mineral Research Centre, Australia. Perry, T.H., Green, D.W., Maloney, J.O. (Eds.), 1984. Perrys Chemical Engineers Handbook, sixth ed. McGraw-Hill. Ray, W.H., 1981. Advanced Process Control. McGraw-Hill. Schroder, A.J., 2000. Towards automated control of milling operations. In: IIR ConferenceCrushing and Grinding Technologies for Mineral Extraction, Perth, Australia. Valery Jr., Walter, 1998. A model for dynamic and steady-state simulation of autogenous and semi-autogenous mills. PhD thesis, Department of Mining, Minerals and Materials Engineering, University of Queensland. St Lucia, Qld, 1997. Welch, G., Bishop, G., 2001. An Introduction to the Kalman Filter. http://www.cs.unc.edu, Department of Computer Science, University of North Carolina at Chapel Hill. Whiten, W.J., 1972. The simulation of crushing plants with models developed using multiple spline regression. In: Salamon, M.D.G., Lancaster, F.H. (Eds.), Proceedings of the Tenth International Symposium on the Application of Computer Methods in the Mineral Industry. SAIMM, Johannesburg, South Africa, pp. 317323.