You are on page 1of 6

G. E.

JOHNSON1
Application of Optimization Techniques to the
Assoc. Mem. ASME
Proluction of Plastic Pellets
IV!. A. TOWNSEND
Mem. ASME In this paper the manufacture of pellets from extruded plastic strands is treated as a
constrained nonlinear programming problem. Complete model development is given.
Department of Mechanical Engineering
and Materials Science, The advantages and disadvantages of various optimization strategies and algorithms
Vanderbilt University, are discussed; a solution is also obtained by Johnson's Method of Optimum Design
Nashville, Tenn. 37235 (MOD). Substantial increases in production rates are indicated with attendant favor-
able changes in the energy consumed in the production of a unit mass of product.

Introduction • the pelletizing machine does not stall,


• pellet dimensions are within a certain range as dictated by
In this paper the problem of maximizing the production of material handling equipment or end use,
pellets from extruded plastic strands is mathematically modeled. • operating speed falls within a certain range as specified by
The resulting nonlinear programming problem (equations (4-9))
the pelletizing machine manufacturer, and
proved difficult to solve using commercially available optimi-
• number of strands does not exceed the maximum number that
zation algorithms. Several schemes were unsuccessful [1, 2, 3]
can be handled effectively by an operator or by the inlet
while others proved to be satisfactory for the numerical ex-
opening to the machine.
amples of the text [4, 5]. The success or failure of any algorithm
may be inherent or due to numerical values of the model pa- (a) Mechanics of the Cutting Process. The production rate is
rameters (since these affect the scaling of the design space); given by the product of the mass of a pellet, the number of
regardless, the solution becomes tedious and expensive if the pellets produced as a blade crosses the bed knife, and the rate
constraint defined by equation (9) is properly treated as an at which blades cross the bed knife, viz.
integer constraint. Other problems associated with standard
nonlinear programming techniques are discussed in the section V = {pAl) (N) (BO), (1)
entitled Method of Solution in the following. time
Because it is desirable to identify a general solution strategy
in its simplest form, an alternate approach to the optimization
is developed based on a method proposed by R. C. Johnson B Blades equally spaced on rotor
[6-11] (subsequently referred to as MOD). The result is a Path followed by blade tips
finite procedure to identify the global optimum for any specific

Problem Formulation
Pellets of materials such as polyester, polyethylene, poly-
propylene, etc. are widely used as the input (raw material) t —i
for the manufacture of molded plastic parts and man-made
fibers. As indicated in Fig. 1, typically extruded strands are
pulled into the pelletizing machine by feed rolls. Rotating
knives cut the strands into pellets (which are easily handled
in bulk) as the strands cross a bed knife.
The general problem is to select pellet dimensions, number
of strands, and operating speed to maximize the production
rate subject to the requirements (constraints) that:
l
N Strands, each
Currently, School of Engineering and Applied Science, University of Virginia, of area A
Charlottesville, Va. 22901.
Contributed by the Design Automation Committee of the Design Engineer-
ing Division at the Winter Annual Meeting, San Francisco, Calif., December Rotor
10-15, 1978 of T H E AMERICAN SOCIETY OF MECHANICAL ENQINEEHS. Manu- Fig. 1 Extruded plastic strands are advanced by feed rolls to bed
script received at ASME headquarters May 30, 1978. Paper No. 78-WA/DE-3. knife where rotor blades cut them into pellets

650 / Vol. 101, OCTOBER 1979 Transactions of the ASME


Copyright © 1979 by ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/


As the number of strands increases, the probability of drop- A simple calculation will show that the power required to
ping a strand increases, which generally results in an "upset" advance the strands into the pelletizing machine is negligible
involving downtime, clearing and subsequent restart. Such an compared to the cutting power requirement. This is corrob-
eventuality can be treated in a number of ways. In the present orated by design practice: independent feed power drives are
study it is assumed that operation with the maximum number on the order of .3 KW, whereas the powers involved in equation
of strands costs one percent of the production rate and that the (3) are typically on the order of 10 to 40 KW. Also an operator
cost (in terms of percentage of production rate) decreases lin- can easily pull the strands from the die to the pelletizer at
early as the number of strands decreases. The adjusted pro- startup, but he cannot chop the full width of strands.
duction rate is then given by
(to) Design Variables and Constraints. Considering the most
V = V ( l ~ ^ \ ^ (2) common case of circular cross-section strands of diameter d,
\ Nmoi / ' time the design variables are d, I, fi, and N. These are to be selected
such that the adjusted production rate is maximized. Accord-
To determine a relation between the design variables and ingly, the mathematical statement of the problem is
the power required to perform the cutting operation, consider
a rotor equipped with B blades parallel to the bed knife cutting
N strands, each of cross sectional area A and thickness in the maximize V = ( 1 - '—— ) 15ir pBd*lQN kg/hr (4)
direction of blade travel t at a rotor speed 0, as illustrated in
Fig. 1. As a blade cuts through the strands, the cutting force as a function of d, I, Q and N
increases from zero to a maximum value of approximately NAT
and then decreases to zero again when the blade has passed subject to
completely through the material. The force as a function of
displacement through the strands will be generally as shown P = 6.6 X 10-« Nd^rBQ, < /(0) kw (5)
in Fig. 2. The energy drawn from the rotor is the area under "•mm _<i d _<i "•max
** dma* (6)
the force versus displacement curve in Fig. 2 and is approxi-
mately equal to NAtr/2. This energy drain occurs each time a < l <I (7)
max
blade cuts through the strands—once every 1/BU minutes— < Q < Q,max (8)
and must be returned to the rotor by the drive before the next
blade makes contact with the strands. The rotor acts as a N < JVraM (9)
flywheel, storing energy when no force is applied. The rate
at which energy must be supplied (power) to the rotor is In equation (5) and as used later, it is assumed that the power
available from the drive /(0) is a concave or linear monoton-
NAT t BQ, ically increasing (or constant) function of rotor speed. This
(3) assumption is valid for most practical applications.

Analysis of the cutting operation for the case of a helical


milling type rotor also leads to equation (3) if helix angles of Method of Solution
less than 6 deg are assumed [12], In present machines, the The problem thus formulated has a nonlinear objective func-
helix angle is around 3 deg. tion to be maximized subject to one nonlinear and seven linear
inequality constraints. General-purpose algorithms for the so-
lution of such problems have been the subject of many investi-
gations and there are several optimization codes available. An
advantage of the algorithmic approach is that a given optimiz-
ation code addresses all problems in the same, presumably
rational, way in making progress toward an optimum; if it fails
to make progress, usually some relevant diagnostic is given.
This general applicability makes the algorithmic approach at-
tractive, especially for digital computer solution.
Nevertheless, there can be disadvantages to the algorithmic
approach—particularly when naively applying a package al-
gorithm. To varying degrees these disadvantages may appear
as follows:
(a) There is no assurance that the search will terminate at
DISPLACEMENT the global optimum: different "optima" will likely be reported
from different starting points for a typical problem, and there
Fig. 2 Cutting force cycles between 0 and NAT each time a blade
passes through the strands may be little clue as to where other, possibly better optima

^Nomenclature-
A —cross sectional area of a strand P power required to cut strands
B =number of blades on rotor t thickness of strand in direction of blade travel
d =diameter of cylindrical pellet, m V production rate, Kg/hr
/(0) =available power as a function of rotor speed for drive, kw V adjusted production rate, Kg/hr
I =
pellet length, m 0 rotor speed, rpm
N =
number of strands p density of plastic, Kg/m 3
N =
value of N for which points 1 and 2 coincide (referring shear strength of plastic for cutting state and temper-
to Fig. 3(a)) ature, Pa
NL = first integer less than N mm lower bound
No = first integer greater than N max upper bound

Journal of Mechanical Design OCTOBER 1979, Vol. 101 / 651

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/


may lie. Obviously, we should always like to find the global N would be anticipated. The value for N (and corresponding
optimum, along with some "guarantee." values for d and 0 ) which results in the best value for the ob-
(b) Insight into the nature of the design space and the jective function specifies the optimal point in the feasible design
interaction of the constraints must be largely intuitive on the space with I = lmaK (unchanged.)
part of the user, and trial and error may be required if an im- Accordingly, the solution proceeds as follows. Equation (5)
provement in the objective function is sought through relax- is solved for d and the result is substituted into equation (4)
ation of the constraint set. The use of penalty functions may giving
further affect this insight, robustness of algorithm behavior
and reliability of results. / MN_\
_ OljV \ P2/3 1™.
V = 1.34 X 105 | 1
(c) Following on (o) and (6) and possibly more seriously, \ ~ Nmm)
scaling of variables and constraints m a y greatly alter the effec- with d defined by
tiveness of different algorithms and the reliability of results.
In many cases, one seeks relatively small changes in design d = 53.3
variables from conventional practices which show desirable im- NTBQ
(11)
provements in the objective function. Poor scaling or poor
choice of step size, convergence criteria, penalty parameters, The feasible design space can now be identified and the vari-
univariate search [13], etc. can greatly affect the success of ation of the objective function within the design space can be
algorithmic approaches. determined. This is illustrated in Fig. 3. The constraint set
(d) The iterative calculations involved do not lend them- is projected upon a plane or planes defined by suitable aux-
selves well to hand calculation should digital computer usage iliary variables and the variations of the objective function
be "inconvenient." with respect to these variables are studied.
(e) Finally, for problems with few variables, the time in- In the present case with reference to Fig. 3(a), the general
volved in wading through and debugging the aforementioned constraint set is projected on to the P-Q plane where /(SJ) is a
considerations may be poorly spent compared with other ap- function of the type defined below equation (9). The actual
proaches. Many practical problems have relatively few vari- feasible design space depends upon the numerical values for
ables and straightforward constraints such as this one, and model parameters; however, it is clear from Figs. 3(6) and 3(c)
accordingly other approaches may be more suitable, reliable t h a t the production rate increases as P and il increase.
and enlightening. There are three cases of practical interest.

Johnson's Method of Optimum Design [7] (MOD) is such (1) If the constraint set is as shown in Fig. 3(a), the opti-
an approach. In this method the solution strategy is posed only mum occurs at point 2 (the intersection of A = fimax and P
after qualitative investigation of the objective function and
constraint set. The problem is reduced to the lowest possible (2) If point C were above point 2, then the optimum would
dimensionality and a finite procedure is developed to find the occur at point 1 (the intersection of d = d m i n and P = / ( 0 ) .
global optimum for any specific case. (3) If point 3 were below point 2, then the optimum would
MOD is often simple to apply, particularly when the problem occur at point 3 (the intersection of d = d mol and Q = Slm>1).
has few variables or the constraint set can be represented by A fourth possibility occurs if point A is above point B, in which
simple functional relations. However these are not require- case there is no feasible design space.
ments. M O D is especially attractive when design variables A procedure to sort these points and identify the optimum
are subject to discrete value constraints (typical of standard for any numerical example is given in the flow chart of Fig. 4
stock sizes, integer values, material properties, etc.). A dis- which defines an efficient digital computer code program to
advantage of M O D is t h a t a different code is necessary for each solve specific cases. I t should also be noted t h a t if N were fixed
problem. Finally, the authors have found t h a t the M O D phi- (as would likely be the case for an existing installation), the
losophy can be applied to simplify many complex problems optimal values for d, I, and 0 are obtained by a single pas3
making algorithmic solutions more tractable when they are
necessary.
The problem as formulated in equations (4-9) can be effec-
tively treated by this technique. To some extent explanation
of the technique is implicit in the solution to the problem.
As a further stimulus and as mentioned in the Introduc-
tion, the problem of equations (4-9) was not solved using eq ( 1 0 )
algorithms [1, 2, 3]. Generally termination or no progress with il constant
were indicated at distinctly non-optimal points; this may be
attributable to any of a, b or c above. Algorithms [4, 5] solved
the problem [4] rather slowly and to a slightly suboptimal
point due to the penalty functions.

General Solution by MOD. With reference to equations (4-9)


the variable I does not appear in the functional inequality con-
straint (5) and hence its optimal value may be determined by
consideration of the objective function (4) and the trivial in-
equality (7). Clearly Zmax should be chosen as the best value
eq (10) with
for pellet length regardless of what values the other variables P constant
take on.
As a practical matter the number of strands will be an in-
teger value between 1 and Nmal. Such variables are often con-
veniently treated as parameters of the solution. Then, optimal
values for other variables (here d and 0 ) and the corresponding
Fig. 3 Projection of constraint set and variations of objective func-
optimal value for the objective function are found for each tion: (a) typical constraint set projected onto PS plane; (b) varia-
integer N = 1, 2, 3, . . ., Nmal. Intuitively, a large value of tion of production rate with P; (c) variation of production rate with Si

652 / Vol. 101, O C T O B E R 1979 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/


C Read T , p, d m - j n , d m a x . N m a x , ^ n , a x . ^ n i i r B
' f(%ax)> f
' t y n i n ) • *uiax
D
r
5, = '-max
VSAVE = 0
N= 0

IF
I d = d
N - Nmax J^ H N = N+l a =
min
fyirin
~i r P=f(%in)
No
Feasible
Design
Space
6
T
« Q l f 6.6xlQ- N d^ ax i B i k r f("max

I >0 • ' ?u

n
- ^nax

d by eq (11]
d
= d min
n by solution of eq (11)
with P = f(n) and d = dm\n
P = f(n)

I
j.._____-p-_ -^^
£0
I
<T"if^vsAVE_^>
> 0

VSAVE = V
dSAVE =d
J2SAVE =a
PSAVE = P
NSAVE = N

Fig. 4 MOD flow chart to identify the global optimum for numerical
cases. This procedure is best used if N is fixed or if a comparison of
V* for various N is desired. Detailed information about the feasible
design space is obtainable through intermediate output.

Table 1 Material properties


material shear strength density
A (similar to 4.83 X 107 Pa 1.3 X 103 Kg/m3
polyester)
B (similar to 6.21 X 10 7 Pa 1.3 X 10 3 Kg/m3
glass f i l l e d
polyester)
C (similar to 2.76 X 107 Pa 1 X 10 3 Kg/m3
polypropylene)

Table 2 Optimal solutions


, design point
kg/hr N* d xl03,oi n*,rpm £*X10 3 ,H P ,kw (ref. Pig. 3a)

6720 120 3.00 400.0 3.56 29.8


NUMBER OF STRANDS, II
5622 2.79 397.4 3.56 29.8
Fig. S Maximum (optional) production rate as a function of the nun
7270 3.56 4O0.0 3.56 28.4 ber of strands

Journal of Mechanical Design OCTOBER 1979, Vol. 101 / 653

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/


^ Read T , p, d m j n , d m a x , i^^, N m a x , ^ , ) a x , »mi n , B, f(flm a x)» f ( n miri) J

— — ^ I If f ^A^min)
" IV
6.6x10-6 d3.n T B fr,1n
>

No
I < 0

design f^in)
<N<N
S
££e 6.6x10-6 d3. n x B fl^ max

raJlim

N = f( min)
max Integer "
6.6x10-6 d^.n T B f^ 1 n

I
<0
If ^(^itiax)
6.6x10-6 d 3 . n T B {^|BX
<
I > 0

< If 53.3 [1(fyiax)


5
'max ^ " m a ) <

n =
tynax
P • f("max)
3/3 -d.

>il d = d,max
n -
tynax
N • Nmax
P by (5)
V by eq (4)
N = N max
d by (11)
V by eq (4)

Fig. 6(a)

O-t N by eq (5) N
n
=
=
NL
^max
P = fttVnax)
d by eq (5)
Vj_ by eq (4)

N = NG
d = d
min
« by eq (5)
P = f(n)
VG by eq (4)

i v - VG 7POINT 1) r f ~^<^iFv G ^v L ^ ^

ST0P»» V = VL (POINT 2)
I < 0

Fig. 6(b)
Fig. 6 Simplified optimization procedure requiring the evaluation of
at most seven algebraic expressions and three comparisons

654 / V o l . 191, O C T O B E R 1979 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/


through the flow chart. With N not fixed, it would still be cedure is given in the flow chart of Fig. 6. The nonlinear pro-
possible (though tedious) to perform the calculations by hand by gramming problem described by equations (4-9) has thus been
incrementing N by some integer I > 1 to bracket the optimum. reduced to a single string of calculations involving the eval-
Actually this is not necessary, as is shown in the next section. uation of at most seven algebraic expressions and three com-
At this point it is worthwhile to remark that the selection of parisons.
P and 0 as the variables to define the plane of projection is not
arbitrary. The other choices are d and U or d and P . The present Summary
choice allows solution with only minimal knowledge of f(£l).
As long as /(fi) satisfies the requirements given below equation I t has been shown t h a t an optimization approach can indicate
(9), the constraint set will define a convex region; this does substantial increases in the production rate with attendant favor-
not occur with either of the other two combinations. I t is able changes in the energy consumed per unit mass of product.
simple to show that equation (10) is concave in the feasible The optimization procedures presented (developed after M O D
region (the Hessian is negative semi-definite). Hence any local [7]) are simple and direct making them attractive for repeated
maximum found is the global maximum. If a feasible design use with a wide range of materials and constraints. Computer
space exists, then a local maximum must occur at one of the codes developed from either flow chart will find the global
three points defined above. optimum after a finite number of comparisons and success is
never dependent on any user supplied "tuning" parameters.
Numerical Examples and Discussion However, straightforward application of package algorithms may
not provide the type of result sought.
A typical example is a pelletizing machine with a rated 40 hp Virtually all design problems benefit by a systematic con-
(29.8 K W ) drive, 72 blade rotor, speed range of 125 rpm to sideration of the model with the purposes of simplifying the
400 rpm, and inlet opening to accommodate up to 120 strands. solution and obtaining greater insight into the critical design
The acceptable range for both pellet diameter and length is parameters and constraints. For the pelletizing problem such
2.79 mm to 3.56 mm. Some representative values for material leads to a closed form solution. In other problems certain
properties are presented in Table 1. Typical values for the difficulties of an algorithmic approach may be reduced or avoided.
design variables prior to optimization are N = 100, d = 3 mm, For example if the dimensionality of the design space is de-
0 = 325 rpm, and I — 3 mm. The production rate for materials creased and one obtains convexity in the feasible region, then
A and B is 3840 K g / h r and the production rate for material the likelihood of finding the global optimum is enhanced for
C is 2950 K g / h r . virtually all algorithms.
Values of the pellet and operating parameters that maximize Significant advances in design methodology are possible if
production rate for the three materials are given in Table 2 the most desirable model manipulations can be identified in
using M O D . The right-most column identifies the type of advance for complex real problems. Such considerations are
design point in Fig. 3(a) per the previous discussion. By com- under study and will be discussed further in a later paper.
parison, optimization can result in a 75 percent increase in
production rate and a 15 percent decrease in the energy con-
sumed to produce a unit mass of product for material A. For
References
material B the production rate is increased by 47 percent, and 1 Bates, H. T., "Nelder and Mead ( N E L D E R ALGO-
the energy consumed to produce a unit mass of product is de- R I T H M ) , " Optimization Techniques with Fortran, by Kuester,
J. L., and Mize, J. H., McGraw-Hill, N . Y., 1973.
creased by 21 percent. For material C the production rate is 2 Siddall, J. N., "Subroutine R A N D O M , " OPTISEP De-
increased by 146 percent, while the energy consumed per unit signers' Optimization Subroutines, M E / 7 4 / D S N / R E P 2 , Faculty
mass of product remains virtually unchanged. It is easily shown of Engineering, McMaster University, Hamilton, Ontario,
that energy consumption is directly proportional to pellet di- Canada, 1974.
3 Siddall, J. N., "Subroutine D A V I D , " OPTISEP De-
ameter and inversely proportional to pellet length [12]. Thus signer's Optimization Subroutines, M E / 7 4 / D S N / R E P 2 , Faculty
the change in this ratio from its original value of unity is re- of Engineering, McMaster Univ., Hamilton, Ontario, Canada,
sponsible for the energy savings reported in the foregoing. 1974.
4 Siddall, J. N., "Subroutine S E E K 3 , " OPTISEP De-
Note that several of the optimal values are very close to, signer's Optimization Subroutines, M E / 7 4 / R E P 2 , Faculty of
but not exactly at their limiting values due to the integer con- Engineering, McMaster Univ., Hamilton, Ontario, Canada, 1974.
straint. In this design technique, this is not attributed to a 5 Townsend, M. A., and Johnson, G. E., "In Favor of Con-
penalty function, as it might be using an algorithmic approach. jugate Directions: A Generalized Acceptable Point Algorithm
for Function Minimization," J. Franklin Inst., in press.
Fig. 5 shows the general relationship between the optimal
6 Johnson, R. C., Optimum Design of Mechanical Elements,
production rate and the number of strands. As N increases, John Wiley and Sons, New York, 1921.
the optimum moves from point 3 to point 2 to point 1 (of Fig. 7 Johnson, R. C., Mechanical Design Synthesis - Creative
3(a)) with the maximum value of the objective function oc- Design and Optimization, 2nd ed., Krieger Publishing Co., Inc.,
curring at the value of N for which point 2 and point 1 coincide. New York, 1978.
8 Johnson, R. C , "A Method of Optimum Design," in-
I t is apparent t h a t if a feasible design space exists and there vited paper presented at ASME Design Automation Conference,
is some value of N < N'max for which point 1 and point 2 coin- Chicago, Sept. 27, 1977.
cide, then the production rate will be a maximum for this value, 9 Johnson, R. C , "Three Dimensional Variation Diagrams
N. Since N can only take on integer values, it is necessary to for Control of Calculations in Optimum Design," Trans. ASME,
Journal of Engineering for Industry, Aug. 1967, pp. 391-398.
compare the production rate defined by point 2 and NL to the 10 Ellis, J., "Johnson's Method of Optimum Design Applied
production rate defined by point 1 and A^G. Here NL is the first to a Problem with Simple Functional Relationships," Comp.
integer less than N and No is the first integer greater than N. Aided Des., Vol. 8, No. 1, Jan. 1976, pp. 9-12.
If no such value N exists, then the maximum production 11 Ellis, J., "A Note on Small-Scale Design Optimization,'
ASME Journal of Engineering for Industry, Vol. 99, No. 4, 1977,
rate occurs at Ar = Nma* and the other variables are determined pp. 915-916.
by point 2 if point 2 is a member of the set of points bounding 12 Johnson, G. E., "Energy Cosnumption and Drive Selec-
the feasible design space. If point 2 is not a member of the set tion for Plastics Pelletizing Machines: An Analytical Study,"
of points bounding the feasible design space then the other Soc. of Plast. Eng. Tech. Papers, Vol. 22, April 1976, pp. 616-617.
13 Johnson, G. E., and Townsend, M. A., "Nonoptimal
variables will be specified by point 3. Based on these obser- Termination Properties of Quadratic Approximation Interpola-
vations a very direct procedure can be developed to determine tion Univariate Searches," J. Franklin Inst., Vol. 306, No. 3,
the opthnal values for the design variables. This revised pro- 1978, p p . 257-266.

Journal of Mechanical Design OCTOBER 1979, V o l . 101 / 655

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27975/ on 07/25/2017 Terms of Use: http://www.asme.org/

You might also like