Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

C()p y ri ~ hl © IF:\( : :\lItolll atioll ill \Iilling .

\tille ral and \t elal Prolt:·...... ill g-. T"b o .


.1 "1''''' 1 ', I ~li

MODELLING OF GRINDING SYSTEMS


FOR CONTROL
A. J. Niemi* and R. Ylinen**
" flr/ ,illki [ ' lIi; 'l'l'Ii/), r1 Tn//I/tdlJgy, SF-IJ2/51J E'j }(JlJ, F i ll/ll llt!
'''' 1"('(h llil'll/ H nl'ol"ch CI'II /I"I' IJj FIII /WII /, SF-1J2/51J /:"' j}(/{I, F i ll/ollt!

Abstract. Steady state modeLs of open and cLosed grinding Loop are derived,
starting from the dynamic modeL of the rotary miLL and the modeLs of the other
components. The parameters have been evaLuated by sampLing and anaLyses, and by
computat10n ba s ed on the modeLs. Use of the modeLs for controL and optimization
of the steady state is expLained.

The dynamic modeLs are aLso studied in order to deveLop recursive on-Line
methods for parameter and state estimation. Both a generaL, infinite
dimensionaL state space modeL and a sLightLy simpLer finite dimensionaL modeL
are presented and their use for estimation and controL is discussed.

Keywords. Grinding; modeLLing; muLtivariabLe controL methods; parameter


estimation; recycling; state estimation; state space methods; steady state.

1 INTRODUCTION system and used for estimation of proces s


variabLes by KaLman fiLter and for muLtivariabLe
The finaL stages of comminution of the ore take feedback controL (Niemi & al. 1982). An
pLace in rotary milLs. A fine grain size has to experimentaL model can also be used without such a
be reached in order that the vaLuabLe mineral~ r.an modification in frequency domain, for development
be recovered by further proce ss ing. A coarse of compensation of interactions by inverse NYQuist
grind resuLts in a Low yieLd or a Low grade of the diagrams (Hulbert & aL. 1981).
concentrate. On the other hand, a too fine grind
indicates an excessive consumption of energy and The muLtivariable controL has produced a
may aLso Lower the metaLLurgicaL resuLts. considerabLy better control of the finaL particLe
size or its binary distribution, and of the pulp
AccordingLy, the controL of grinding systems has density. However, due to the inertia of the pLant
been the subject of numerous studies, and various to changes of ore feed, a corresponding
controL methods are being used in the industry. retardation has been required in the water feed
SingLe-input singLe-output feedback Loops for control resuLting in a sLuggish overaLL controL
constant vaLue controL of individuaL variabLes are (YLinen & aL. 1984). Therefore a muLtivariabLe
commonLy used and often severaL inputs of a system modeL composed of modeLs of the individuaL
may be controLLed by a number of such Loops. components is indicated as an alternative basis of
However, controL of the numerous variabLes by cant roL.
independent controLLers is made difficuLt by the
interactions of these variabLes in the miLLs and Multidimensional model of the particLe size at the
due to the recycLing stream of sLurry. This has outlet of a grinding miLL is availabLe in the
resuLted in different choices of controLLed weighting function form which incLudes an
variabLes in different pLants, and no combination arbitrary residence time distribution and the
of such feedba c k Loops has received a generaL kinetics of grinding. A corresponding steady
acceptance. state modeL has been derived for the closed
grinding Loop. The residence time distribution
MuLtivariabLe controLs have therefore been and the values of the parameters of the latter
deveLoped in order to take into account the Bare determined by sampling under operation of the
interactions of the process variabLes. The miLL industrial system, by Laboratory anaLysis of the
has usuaLLy been modeLLed as a singLe perfect sampLes and off-Line simulation. Such modeL
mixer which, even after incLusion of kinetic yieLds the grain size distribution of the final
terms, is described by a set of first order output and characteristics of the recycling
differentiaL equations. Thus a division of the materiaL (Niemi & aL. 1986). It can be used for
materiaL in two size cLasses was appLied by simulation, optimization and controL of the steady
Rajamani (1979) who used a muLtivariabLe, OPtimaL state of the system whereby iterative computation
aLgorithm for controL of cLosed Loop grinding in is needed for solution of the closed Loop.
piLot pLant. The state space modeLs have aLso
been used for state estimation by Wyatt-Ma i r & aL. The stated model has been derived by reduction of
(1980), for hardness identification by Herbst & the dynamic models of the Loop components to
al. (1980) and for adaptive optimal control by steady state. A dynamic modeL of the Loop can
Herbst & al. (1985). ALL the Latter authors have also be constructed, as a combination of the
applied an extended Kalman filter approach. modeLs of the Loop components. The present paper
discusses the structure of such dynamic and steady
A state space model has aLso been derived from an state modeLs of the miLL and the grinding loop,
experimental input-output model of the grinding and their reduction to forms, which are more
suitabLe to development of controls.

187
.'I. . .J. \:iemi and R. Ylinen

2 STEADY STATE MODEL AND CONTROL Here 1 is a diagonaL matrix obtained by


OF ROTARY MILL muLtipLication of the matrix (2) ~ith the
residence time distribution get) and by
Th~ r~sid~nc~ tim~ distributions and kinetic integration of the product. The residence time
modeLs of grinding are Linear functions of distribution g can be determined experimentaLLy by
fractionaL amounts of soLids in miLL feed, and may means of tracers. The number of kinetic
therefore be superimposed ~ith each other. If an parameters is usuaLLy so high that they cannot be
ore of variabLe size distribution is considered determined independentLy. ALgebraic formuLae have
~hich enters continuousLy a rotary miLL, the been set up on experimentaL basis ~hich yieLd the
fractions of materiaL of the various size cLasses vaLues of k j and b ij in terms of a fe~ more
at the miLL outLet are obtained at any time t by generaL parameters onLy. The vaLues of these
means of a convoLution integraL; for derivation parameters are evaLuated by means of an extremum
see e.g. Niemi & aL. (1986) and references seeking method, so that the computed output fits
Listed there. the experimentaLLy determined output distribution
for the same input which requires an iterative
(11 computation.

The vaLues of parameters have been extracted at


~here Set) vector of feed size fractions practicaL tests and simuLations have sho~n a good
S, (t) vector of outLet size fractions agreement of the computed output with the measured
get) residence time distribution size distribution at the steady state. The modeL
(5) may no~ be used for simuLation of the steady
(2) states for other vaLues of the inputs. If the
parameters may be assumed constant for some Length
The eLements of the matrix T are obtained from the of time, aLso the dynamic modeL (11 may be used
foLLo~ing recursive formuLae; !-1 is its inverse: for computation of the variabLe output, after the
parameters have been evaLuated in steady state.

tij 0
f;;'0
kLbiLt Lj
i<j
i=j
(3)
After the parameter vaLues and the residence time
distribution have been determined in a basic
steady state, the outputs of a miLL have be~n

1 L=j ki-k j

The soLid materiaL in the size cLass j is above


i>j
computed for other constant inputs. A quite good
agreement of the computed and experimentaL output
distributions has been reached, even ~hen the
considered to break ~ith the rate kjC j b· . throughput and the feed size distribution have
gives the fraction ~hich thereby faLLs to t~J differed considerabLy from the originaL vaLues
cLass i (j=1,2, ••• ,i-1). (Niemi & aL. 1986). Thus the modeL can be used
for a fast simuLation for various vaLues of the
By means of Eq.(1) the size distribution of the inputs and such way for controL of the steady
miLL output can be computed, if the size state, and for optimization assuming a performance
di s tribution of feed i s kno~n as a function of the criterion has been defined.
time and if the residence time distribution and
the vaLues of grinding parameters are aLso kno~n.
The operation of the miLL may be simuLated by 3 STEADY STATE MODEL AND CONTROL
means of this modeL for other feed distributions OF CLOSED GRINDING LOOP
and for other vaLues of the parameters, if e.g.
the vaLues of the kin etic parameters for ores of In a cLosed grinding Loop, a part of the milL
differ e nt hardne s s ar e kno~n. ALso the changes of output is recycLed to the miLL feed (Fig.1I.
fLo~ through the miLL may be introduced in the FEED TO FL OTATION
modeL through their effect on the function g, 9f = ~ f '1f
since this can be considered invariant in terms of
the variabLe z introduced by Niemi (1977).
CLASSIFIER

z = .fO(t) dt (4)
v
T'lTAL FRESH .3 a
Here Q is the voLumetric fLo~ of the sLurry and V SOLIDS FEE9 ~------~=C , =SC Q
the sLurry voLume of the miLL, considered Q = ~Q -a a- - c c
constant. In the te s ts made by severaL schoLars
~ith fuLL size miLLs, the residence time
distributions of the soLids and Liquid have been
found equaL ~hich indicates their dependence on
~
the totaL voLumetric fLo~ onLy.

In the steady state, C and S, are constant q


- c= C
-c'1 c
vectors in Eq.(1), and ~e time-depend~nt part of
the integrand is reduced to the product Dg. Their Figure 1. CLosed grinding Loop.
integraL can be computed numericaLLy, Tf get) is
kno~n as experimentaLLy obtained data, or The Linear miLL modeL presented previousLy can be
anaLyticaLLy if it is approximated by a cLosed appLied to both the fresh and recycL~d feed. The
mathematicaL expression. The steady state modeL totaL soLids feed to th~ miLL Q consists of the
of the mi LL is thus: sum of aLL such partiaL fLo~s of each size cLass
which have been recycLed an arbitrary number of
T- 1 C
_T -J _.:.a = _G .:.a
C (5) times (Ni~mi 1984).

Here C is the vector of the size fractions of the


fresh- feed and S th~ diagonaL matrix of the
Modelling of Grinding Svstems for COil! rol IH9

cLassifier parameters reLating the cLassifier (1976) formuLae were used for the cLassifier. The
underfLow to cLassifier feed for each size cLass. computations converged reLativeLy fast and LogicaL
- A corresponding modeL of a different form has resuLts were obtained both for the output size
been derived by Herbst & aL. (1971). distributions and gross fLows (Niemi & aL. 1986).
At Later tests at which the experimentaL resuLts
Assuming the recycLing soLids fLow is known by were compared with those simuLated using parameter
size fractions through Eq.(6), a number of other data determined in a first test, a good agreement
quantities can be obtained. One of them is the of the experimentaL and simuLated data was
size distribution of cLassifier overfLow If and
the fLow ~r' since the totaL fLow Leaving the
Loop equa s the fresh feed fLow in the steady
observed.

SimuLation of the Loop for various steady states


state. ALso the sLurry density in the return Line serves as an aid to the controL and optimization
is obtained assuming the factor of division of of the operation. Feeds of soLids and water can
Liquid by the cLassifier has been determined. be used as controL inputs in order to counteract
the effects of constant disturbances. The
Determination of the residence time distribution hardness of the ore described by the vaLues of the
of the miLL in cLosed Loop is compLicated by the grinding parameters is usuaLLy an independent
recircuLation of the tracer, because the tracer disturbance, whiLe the size distribution of the
impuLse is foLLowed by an extended secondary puLse fresh feed can be controLLed through manipuLation
at the point of feed. A method based on Fourier of the crushers preceding the miLLs.
transformations has been derived and appLied
succesfuLLy, using tracer information from the The simuLation can LogicaLLy be extended to more
mi LL out Let onLy (Niemi & aL. 1985). The method compLex systems, by combination of the open Loop
appLies to both Liquid and soLid tracers, whiLe in and cLosed Loop modeLs. The range of vaLidity of
the Latter case the different cLassification of the modeL is restricted by the Limited vaLidity of
the tracer particLes of different sizes has to be some of its components in the case of Large
observed. The cLassifier factors are easiLy excursions of the variabLes from the vaLues for
obtained through sampLing of its feed and outLet which the parameters were evaLuated. At present,
streams and screen anaLysis of the sampLes. the range of simuLations seems to be mainLy
determined by the cycLone modeL.
After the residence time distribution of the miLL
aLone has thus been obtained, the vaLues of the
kinetic grinding parameters of the normaL, neutraL
process materiaL can be extracted. SampLes are 4 DYNAMIC MODELS OF ROTARY MILL
first taken of the feed and output of the miLL and
their size distributions determined. The vaLues The dynamic input-output modeL of the rotary miLL
of parameters which provide a soLution of Eq.(5) was presented in Chapter 2 in the weiqhting
are then determined in the same manner as for an function form (1)
open Loop miLL. Since the distribution of the
feed to the miLL is compLeteLy known by sampLing t
and anaLysis, it is immateriaL at such S,(t) J TO(t-,)!-1 Ia ( , )g(t-,)d,
determination, to which degree it incLudes fresh (7)
or recycLed materiaL. t
= J §.(t-')Ia(,)d,
The modeL has been tested with data obtained from
an industriaL, cLosed grinding Loop. The
residence time distribution of the miLL and the where G(t)
vaLues of the grinding and cLassifier parameters
have been first determined and then the operation This modeL is based on the assumption that the
of the Loop simuLated. A good agreement of the totaL voLumetric fLow changes so sLowLy that
computed and experimentaLLy determined output data its infLuence on the residence time distribution
has shown that the modeL and the computationaL and the kinetic parameters needs not to be taken
procedure are weLL operative. into account continuousLy.

After the originaL steady state of the Loop is GeneraLising the convoLution muLtipLication (*) in
thus known, the modeL can be used for search of a naturaL way the modeL (7) can be written as
other steady states produced by other vaLues of
the inputs. The dependence of the cLassifier S,(t) = (§. * Ia) (t) (8)
factors S on fLow and soLids contents of the
sLurry has then to be observed. ExperimentaL By means of LapLace transformation, this is
cLassifier modeLs are avaiLabLe in the Literature brought to the form
and their parameters can be evaLuated by testing
the cLassifier at different states of operation. (9)
S,(s) = §.(s)Ia(s)
A simuLation for a new steady state is started
with a guess of new vaLues of the fLows of soLids where S,(s), G(s) and C (s) denote the LapLace
and Liquids to the cLassifier which fixes aLso the transforms of-~(t), §.ctr and Ia(t), respectiveLy.
corresponding fLows of recircuLation. The vaLues
of the cLassifier factors S are obtained from the Let g(s) denote the LapLace transform of g(t).
experimentaL formuLae. The residence time Then
distribution of the miLL is modified according to
the new voLumetric fLow through the miLL. The
steady state is then determined according to
Eq.(6). Since the rtesuLting totaL fLows
naturaLLy differ from the originaL guess, new
_ I";k,) 0
~(s+k2)
0
0
l

G(5) =T • T-1 (10)


vaLues are chosen and iteration continued in order
to converge to the finaL vaLues. I q(s+k m)
0 0
i.e. L -
The method has been tested by simuLation for feeds
which differed from the originaL vaLues; PLitt's
190 :'I.. .J. "Jiemi and R. ),linen

The modeLs (7) and (10) present a way for The discrete time counterpart of the modeL (13) is
estimation of the kinetic parameters and the
residence time distribution of the miLL. In ~(k+1,L) = ~(k,L-1)
practice the parameters are varying in time so
that for controL purposes their continuous (17)
updating is usefuL, even necessary. The most
updating methods are recursive and based on
recursive system modeLs (state space modeLs) in )' G(Lllt)x(k,Ll
order to decrease the amount of computation. L=O
Often they are connected to state estimation
which is aLso needed in controL, of strongLy which is more suitabLe for numericaL computations.
disturbed systems in particuLar.
The modeL (13) is in generaL infinite dimensionaL,
A state space modeL corresponding to the modeL (7) but in practice the residence time distribution
can be constructed as foLLows. Let the state can be presented by a function which is equaL to
x(t,.) of the system at time t be defined by zero after some finite time intervaL. In the
discrete time case a finite dimensionaL
(12) representation can then be obtained.

Then the modeL (13) is obtained. Let the residence time distibution be
presented by the impuLse response of an
ox ox n-dimensionaL Linear differentiaL system and a
-=(t,1:) + -=(t,1:) = 0 time deLay T.
at ih
~(t,Q) = S(t) (13) x(t) Fx(t) + Gu(t-T)
(18)
y(t) = Hx(t) + Lu(t-T)
Sb(t) = f .§.(,,)~(t,1:)d1:
o
It is now possibLe to construct an
The transfer of the system from the m x n-dimensionaL Linear state space modeL for the
state ~(to'.) at time toto the state ~(t,.) at time mi LL.
t, when C (1:) is given for the time intervaL
. d
t o < 1: ~ t -'-ij1S carr1e out b y

if T > t-t o
( 14)
if 0 < 1: < t-t o

i.e. by concatenating the new input to the oLd


state.

The modeL can be appLied e.g. to a recursive


estimation of the miLL inLet size distribution
~il' if this is not compLeteLy known due to An open Loop rotary miLL is usuaLLy not controLLed
d1sturbances. Furthermore, the estimation of some by the inLet size distribution but by the soLids
process parameters can be inc Luded. Suppose that fLow, water flow or some other manipuLated
the function G(t) depends on parameters a, i.e. variabLe whose effect on the outLet size
it is of the form-G(t,a). Suppose further that the distribution appears e.g. through the variation
parameters are constant i.e. of the residence time distribution, Therefore the
modeLs are not directLy correspond to the needs of
a(t) = 0 ( 15) standard methods of controL design. They, however
are very suitabLe for simuLation and furthermore,
In this case the following recursive estimation for parameter and state estimation, which are
aLgorithm is appLicabLe. important parts of the modern controL system.

OX (Ix 5 DYNAMIC MODELS OF CLOSED GRINDING LOOP


-=(t'1:)+-=(t'1:)=~1 (t'T)(~(t)
at 0T
Consi der the closed grinding Loop in Fig.1
- r .§.(1:':~(t))!.(t'1:)d1:) which consists of a rotary milL, a sump and a
o (16) hydrocycLone cLassifier. Suppose further that the
2(t,0)=S(t) voLumetric fLows Q, Qa' Qb ' Qc and Qf change so
sLowLy that they can be considered as constants in
i (t ) =!<.2 ( t) (~( t ) - "i .§. <-r,i (t ) )1. (t , T) d T) the modeLs and the system can be assumed to be in
steady state with respect to fLows and voLumes.
n
The modeLs presented in the previous Chapter can
Here x and a are the estimates of ~ and ~, and ~1
be taken for modeLs of the rotary miLL aLso here.
and K2 suitabLe weighting matrices chosen e.g.
The sump is described by a oerfect mixer
via the extended KaLman fiLter equations or some
other recursive estimation method. It shouLd be (20)
noted that in this case the extended state
equation is stiLL Linear which simpLifies the Observing aLso Fig. 1,
computation. On the other hand, the
parameters appear usuaLLy nonLinearLy in the
output equation.

(21)
Modelling of Grinding SI'stems for Control 191

where gs(t) = (Qc/V)e-(Qc/V)t If the rotary mi II is modelled by a finite


dimensional linear state space model (19), the
Equivalently with Eg. (21), closed loop system can also be modelled by a
finite dimensional state space model.
(22) Let the mill model be (c.f. Eq.(19»
The volume of the hydrocyclone classifier is x = ~ +.§".S
considered so small that its static model is (30)
accurate enough
lb = !i rb + .!::.nfa
~QC£c (t) and the sump model (c. f. Eq. (20))
(23)
(.!.-~)QC£c(t) (31)

Connecting the models (7), (21) and (23) the


following is obtained: Then the closed loop system has a state space
model (32)
t
f
[ti k~;;l:;;;;;;i:;;;~;~~;;;;d~t;;;i:i£
gs(t-,)Ql(,)d,

t
+ f (gs*.§.) (t-,)~QC£c(,)d, (24)
If = [.Q : (Qc/Of) (.!.-~) 1 ~t]
The closed loop models contain the kinetic
Provided that the closed loop system is stable parameters and the residence time distribution of
Eq.(21) can be solved for QC£c the rotary mill, and the classifier parameters.
The most parameters are dependent on the flows
QC£c(t) = «.!.+(gs*(~»+(gs*(GS»2+ ••• ) which also are at least partially unknown.
Therefore their on-line estimation all together is
(25) a difficult problem. The use of the extended
state estimators, e.g. the extended Kalman
where (qs *(GS» n=(q *(GS» *(g *(GS)f1- 1 for n=2,3, ••• filter, which in fact is closely related to the
iterative scheme for the steady state
The term «gs*(GS»n*gs*(Ql»(t) in the series (25) parameter estimation presented in Ch.3, seems to
corresponds to the fractional flows circulated be a solution for the problem. However, this
in the system exactly n times i.e. the approach has not so far been tested in dynamic
mathematical solution of the model corresponds situations.
to the model of the closed loop obtained by
physical reasoning (~Iiemi 1984). Like the open loop models presented in Ch.4 the
above closed loop models are not as such
According to Eg. (22) and Fig. 1 applicable to the analytic control system design,
because the main manipulated variables of the
(26) process appear in the model parameters. Therefore
simulation is the only way to study the responses
or of the process to thei r changes. By means of such
studies, it is obviously possible to construct less
complicated, aggregated models for the
relationships of the manipulated and output
The state space models of the closed loop system
variables.
can be cons t ruct ed in the same way as those of
rotary mi ll. Let the state ~(t,.) at time t be
defined as 6 CONCLUDING REMARKS
!(t,,) lc (t-,) (28) The steady state and dynami c mode ls of open and
Then closed loop gri ndi ng system s form a good
i'l x (I x basis for the simulation, estimation, control
-=(t,,) + -=(t,,) = 0 and oPtimization of such systems. The methods
(I t ;'1< based on steady state models have been tested in
practical experiments and they have operated
satisfactorily. The use of the dynamic models for
recursive on-line state and parameter estimation
Q~(t,o> f gs h)Ol(t-,)d,
seems applicable according to the experiences from
0
the iterative steady state estimation. The models
can also be a basis for constructing less
(29) complicated, aggregated models to be used in
analytic control system design.

This model is more complicated than the model (13)


because the closed loop system contains
inherently a state feedback loop. The feedback,
furthermore, includes the rotary mill and the sump
so that the system extended by model parameters
as additional state variables becomes nonlinear
making the state and parameter estimation more
difficult than in the open loop case.
192 A. J. Niemi and R. Ylinen

REFERENCES

Herbst J, Grandy G and Mika T 1971, On the


development and use of lumped parameter models
for continuous open and closed-circuit
grinding systems. Inst. Min. Met. Trans.
C. 80, pp.193-198.

Herbst J, Rajamani K amd Pate W 1980,


Identification of ore hardness disturbances in
a grinding circuit using a Kalman filter.
Proc. 3rd IFAC Symp. Automation in Mining,
Mineral and Metal Processing (Editors J O'Shea
and M Polis) pP. 333-348. Pergamon Press.

Herbst J, Alba J 1985, An approach to adaptive


control of mineral processing operations.
Proc. 15th Int. Mineral Proc. Congress,Part
Ill, Cannes, pp.75-87.

Hulbert D, Koudstaal J, Braae M and Gossman G


1981, Multivariable control of an industrial
grinding circuit. Proc. 3rd IFAC Symp.
Automation in Mining, Mineral and Metal
Processing (Editors J O'Shea and M Polis) pp.
311-322. Pergamon Press.

Niemi A 1977, Residence time distributions of


variable flow processes. Int. J. Applied
Radiation and Isotopes 28, pP. R55-8~n

Niemi A, Heinonen P, Jamsa S-L, Melama H,


Iivarinen T and Paakkinen U 1982 Experiences
in multivariable control of sulphide ore
grinding in Vuonos concentrator. Preprints
14th Int. Mineral Proc. Congress, Paper
111-5. Canadian Inst. of Mining and
Metallurgy.

Niemi A 1984, Steady-state models for optimal


operation of ore dressing plants.
Applications of Computers and Mathematics in
Mineral Industries (Proc. 18th APCOM Symp.)
IMM London, pp. 317-323.

Niemi A, Zuniga R 1985, Modell i ng of grinding


and flotation systems by means of tracers.
Flotation of Sulphide Minerals (Edi tor K
Forssberg) , pp.397-408. Elsevier.

Niemi A, Zuniga R 1986, Evaluation of


parameters and testing of models in ore
grinding plants. 1st World Congress Particle
Technology, ~uremberg, 13p.

Plitt L 1976, Mathematical model of the


hydrocyclone classifier. Canadian Min. Met.
Bull. ~9, pp.114-123.

Rajamani K 1979, Optimal control of closed


circuit ball mill grinding. Ph.D. thesis,
Univ. of Utah.

Wyatt-Mair G, Garner K and King R 1980,


Realtime digital computer state estimator for
a hard rock milling circuit. Proc. 3rd IFAC
Symp. Automation in Mining, Mineral and Metal
Processing (Editors J O'Shea and M Polis), Pp.
363-374. Pergamon Press.

Ylinen R, Iivarinen T and Niemi A 1984, A


linear-quadratic-gaussian control algorithm
for sulphide ore grinding. Automation in
Mining, Mineral and Metal Processing (Editor T
Westerlund), PP. 199-205. Pergamon Press.

You might also like