You are on page 1of 6

EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
European Journal of Operational Research 92 (1996) 63-68

Theory and Methodology

Optimal solutions for the machining economics problem


with stochastically distributed tool lives
Eleftherios Iakovou aY* , Chi M. Ip a, Christos Koulamas b
a Department of Industrial Engineering, University of Miami, P.O. Box 248294, Coral Gables, FL 33 124, USA
b Department of Decision Sciences and Information Systems, Florida International University, Miami, FL 33 199, USA

Received June 1994; revised April 1995

Abstract

This paper proposes analytical models and numerical procedures for simultaneously determining the optimal cutting
speed and tool replacement policy in machining economics problems with stochastic tool lives when the objective is the
minimization of the machining cost per part. It is shown that the objective function is separable for certain phase type tool
life distributions, including Gamma, which leads to an efficient solution procedure.

Keywords: Machining economics; Manufacturing; Maintenance; Tool replacement

1. Introduction ing cost per part. The optimal cutting speed and tool
replacement policy are derived for both Markovian
Tool life variability has long been recognized as
and non-Markovian tool failure processes. The rela-
an important aspect of machining economics (Rama-
tionship between the optimal cutting speed and the
lingam and Watson [7], Ramalingam [8]). When
optimal tool replacement policy is shown for various
stochastic tool lives are introduced to machining
tool failure processes. In general, these problems are
economics, the possibility of preventive tool replace-
solved by two-dimensional search procedures along
ment arises, since unforeseen tool failures disrupt
with numerical integration; however, we show that in
production resulting in penalty costs. In that case, an
the case of Gamma distributed tool lives (and certain
optimal solution to machining economics problems
phase type distributions) the problem has a structure
requires the simultaneous determination of the opti-
that allows the computation of the optimal replace-
mal machining conditions (e.g. cutting speed, feed
ment policy using a one-dimensional search. Hence
rate, depth of cut> and the optimal tool replacement
the computational effort for finding the optimal solu-
policy in order to minimize the machining cost per
tion is reduced considerably.
part.
The rest of the paper is organized as follows. The
This paper develops solution methodologies for
related literature is presented in Section 2. The pro-
the single operation, single tool machining eco-
posed formulation is presented in Section 3. Solution
nomics problem with stochastically distributed tool
procedures for Gamma distributed tool lives are pre-
lives when the objective is to minimize the machin-
sented in Section 4. A numerical example is pre-
sented in Section 5 and the findings of this research
* Corresponding author. are summarized in Section 6.

0377.2217/96/$1.5.00 0 1996 Elsevier Science B.V. All rights reserved


SSDI 0377-2217(95)00131-X
64 E. Iakauou et al./ European Journal of Operational Research 92 (1996) 63-68

2. Related literature V : Cutting speed.


t,,, = k/u : Machining time per part.
Tool replacement policies have been studied ex- T : Tool life (a nonnegative random vari-
tensively in the operations research literature. In able).
most cases, however, the tool replacement problem f(x) : Probability density function of T.
is studied in isolation of the machining economics F(x) : Cumulative distribution function of T.
problem under the assumption of a predetermined E(X) : Expected value of random variable X.
processing rate (machining speed). In those cases the t” : Scheduled tool replacement time.
tool replacement policy is considered as a mainte- ? : Actual tool replacement time.
nance problem in reliability theory (Barlow and P : Penalty cost for unforeseen tool failure.
Proshan [2]). Many papers have dealt with the opti- C : Constant in the Taylor tool life equation.

mal maintenance policy of a single unit system. We n : Exponent in the Taylor tool life equation
mention Kao [3] and Anderson [ 11. Taha [lo] pre- (0 < n < 1).
sented a similar model explicitly mentioning cutting t, : Machining time per part.
tools. Recently, Zhou et al. [12] presented a tool MC : Total machining cost per part.
replacement model based on tool wear status. Additional notation will be introduced later as
Only a few papers have dealt with the simultane- needed.
ous determination of the optimal cutting speed and The problem under study is a single machine,
tool replacement policy. La Commare et al. [4] pre- single tool machining economics problem with
sented a formulation for the above problem with the stochastic tool lives. The objective is to determine
restriction that the preventive tool replacement time the optimal u and t * values in order to minimize the
be an integer multiple of the unit machining time. No total machining cost per part MC given as
analytical solution was presented and the optimal
cutting speed and tool replacement policy were ob- MC = cmt,
tained by applying a two-dimensional search to the
E(cost per tool)
feasible solution space. +
Sheikh et al. [9] determined analytically the opti- E(number of parts machined per tool)
mal cutting speed and tool replacement policy under (1)
the restrictive assumption that the Taylor tool life
equation applies not only to the expected tool life but subject to Taylor’s equation (Taylor [ 121):
also to any tool life fractile. This assumption enabled
the authors to determine the optimal tool replace- u{ E( T)}” = c. (2)
ment policy independently of the cutting variables,
resulting in the well known age replacement policy The above equation was originally developed by
in reliability theory (Barlow and Proshan [2]). Other Frederick Taylor at the beginning of the century
earlier treatments to the problem are mentioned in (1907) and is still in use in industry today. In his
the above references. original presentation Taylor defined the constant c as
the cutting speed u which results in a tool life of one
minute. The Taylor tool life equation is still the basis
3. Model formulation of determining economic metal cutting and has been
used in adaptive control machining. The equation is
The following notation is used throughout the empirical in nature and the constant c and exponent
paper: n are determined experimentally. They depend on
cln : Cost of operating the machine per unit the material to be cut, tool material, surface struc-
time. ture, cutting geometry, and other factors. The stan-
cl : Tool cost. dard approach is to determine n, c using regression
t, : Tool change time. via a logarithmic transformation. Actual experiments
k : Machining constant. have revealed that this relation holds for a wide
E. Iakovou et al./ European Journal of Operational Research 92 (1996) 63-68 65

range of cutting speed, known as the ‘Taylorized’ halt in production due to the unforeseen tool failure.
region. Recommended values of n, c based on work In other words,
material/tool material are published extensively in p=c,t,,
handbooks, Rembold et al. [9]. Extended versions of
where t, is the time (in addition to the tool change
Taylor’s equation which include the feed rate and the
time t,)‘the production process is halted due to the
depth of cut have also been developed experimen-
tool failure and c, is the corresponding cost coeffi-
tally. However, in our study we utilize the basic
cient. If a preventive tool replacement policy is in
Taylor equation due to the following assumptions:
effect (replacing the tool at f * or upon failure,
1. As in most machining problems, the maximum
whichever comes first), the expected tooling cost per
allowable feed rate and depth of cut values are
tool cycle is
restricted due to technological constraints. Conse-
quently, the feed rate and depth of cut are set to their c,=c,t,+c,+pF(r*). (4)
maximum values leaving the cutting speed as the The expected duration of a tool cycle is the actual
only cutting variable to be optimized. Eq. (2) then tool replacement time which is equal to either the
becomes the applicable Taylor tool life equation. tool failure time T or the preventive tool replacement
2. The tool life distribution depends on the cause time t*, whichever comes first.
of tool failure. Ramalingam and Watson [7] show
that the tool life distribution can be assumed expo- E(f) =/ mmin( x, f*)f( x) dx
0
nential when the tool fails catastrophically (i.e.
crashes) the first and only time it sustains a signifi-
cant injury in the form of an impact (assuming that
the hazard of such an injury is finite but constant). In
that case a Markovian tool failure process is in = ot’,$(x) dx+t*[l -F(t*)]
/
effect. Ramalingam [8] shows that the tool life distri-
bution is Gamma in the case of tool failure from =/b’(l -F(x)) dx, (5)
multiple injuries due to constant, time-independent
stochastic hazards. In that case a non Markovian tool where the last equality follows from integration by
failure process is in effect. In this paper we develop parts and f = min(T, t * ).
solution methodologies for both cases. The expected number of parts machined per tool
3. The overall machining cost is the sum of the is E(f)/t,. Substituting (3), (41, (5) and t, = k/u
processing cost (c,t,) and the tooling cost per part. in cl), the following constrained optimization prob-
Since cutting tools deteriorate independently of each lem is derived:
other, the tool replacement instances constitute re- min MC = c, k/u
newal epochs of the tool replacement process. Using
the renewal-reward theorem, the expected tooling + kb, +M’(t* )I
cost per part is the expected cost per tool cycle (6)
u/b*‘1 -F(x)) dx
divided by the expected number of parts machined
per tool cycle.
We develop expressions now for both expected subject to u{ E( T)) ” = c, (7)
values. where
When a failure or scheduled tool replacement
takes place there is a charge equal to the tool cost in E(T) =/aw$(x) dx
addition to the replacement cost, given as
c3 = c,t, + c,. =c(l-F(X)) dx.
(3)
In the case of tool replacement due to failure, Notice that in the Taylor equation the untruncated
there is an additional penalty cost p. This penalty expected tool life is used since the relationship be-
quantifies all problems associated with the temporary tween cutting speed and expected tool life is deter-
66 E. Iakovou et d/European Journal of Operational Research 92 (1996) 63-68

mined only by the combination of workpiece and Note that (12) can be expressed as
tool material.
min MC = g,(A) f g2( A)h( Y), (13)
For a given LJ (and hence E(T)) the problem
reduces to choosing t * to minimize where

g,(A) = k%‘!A-“, (14


(8) C
/b’(l -F(x)) dx’

which is a classical problem in replacement theory g2( A) = frYi’-, (15)


(McCall [5]). In general the minimization of (6)
subject to (7) requires two-dimensional search tech-
niques as well as numerical integration. In the next
c,t,+c,+p
I 1 -eey 5 (Y/j!)]
section we demonstrate that in the case of a Gamma h(y) =
L j=O
‘9 (16)
(Erlang) distribution of fixed order and of certain
other phase type distributions, the objective function
is separable and a solution methodology is devel-
1- eeyi (f/j!)
j=O I
and y = At *. Since the objective function in (13) is
oped.
separable in A and y the problem can be solved by
first minimizing (13) with respect to y and then with
4. Solution analysis in the Gamma case respect to A and it is well known in the replacement
literature (McCall [5]) that a unique minimizer y * of
Suppose that the tool life follows a Gamma
(16) exists and can be obtained by solving the fol-
(Erlang) distribution of fixed (and known) order r,
lowing equation:
i.e.,
A
f(x) = tr_ l)! (Ax)r-‘e-Ax, x20. (9)
Since the mean of a Gamma distribution is r/A, Eq.
(7) yields
1 1 r”
-=- -
(10)
V CA’
( 1
Note that r is assumed fixed and hence changing LJ (17)
will only affect A. It is well known (Neuts [6]) that
the expected actual tool replacement time, E(f), is The above equation involves emy and powers of y
given by and therefore can be easily solved using bisection or
Newton’s method. Once the optimal y’ has been
(11) obtained then the optimal A* can be computed by
differentiating (13) with respect to A. It can be easily
Substituting (9) and (10) into (6) yields the following shown that
equivalent problem:
A* =
n
(l-:&y*)
t (18)
min MC=k:
so that the optimal cutting speed can be obtained by
n
+ ncl?2
r(1 -n)h(y*) (19)
1 ’
and the optimal scheduled replacement time is
j=o J’ t* =y=/A*.
JI (20)
E. Iakovou et al./ European Journal of Operational Research 92 (1996) 63-68 67

In the special case of Markovian tool life, i.e., If the matrix A can be expressed as
r = 1, h( y> becomes
A=Az,
c,t, + c, +p(l -e-‘)
h( Y> = where x is a matrix with known elements that do
1-e-Y ’ (21)
not involve A, then it can be easily seen that the
which is a decreasing function in y and hence is objective function in (6) is separable in the variables
minimized at y * = cc with a value of h and y = hr * , similar to (12).

h(y*) =c,t,+c,+p.

Substituting into (19) we obtain 5. A numerical example

1
n

v
* =c[ (1 -n)(cT;+c,+p) (22) In this section we present a numerical example
along with the optimal cutting speed and the replace-
and t * = 00,which is consistent with the memoryless ment time for the case where the tool life is Gamma
property of the exponentially distributed tool life. with order 2 distributed. Comparisons are also made
We remark that the above analysis can be ex- to the case where the tool life is Markovian.
tended to the case of a mixed Gamma distribution of The data for the example are as follows:
the following form: k = 900 ft.
4
rc = 2 min.
A
fCx)= gaqi_
( hx)i- le-,iX,
x> 0, ~=4OOft~min”-~.
l)!
c, = $5.
n = 0.3.
(23)
p = $105.
where ai > 0, 1 < i < q and
c, = 15 $/min.
Eq. (16) reduces to
Cai= 1.
i= 1
35 + 105[1 -ewY(l +y)]
h(y) =
Since the distribution function of a positive ran- (2-2emY -yemY) *
dom variable can be approximated arbitrarily closely This can easily be solved using bisection; h(y) is
by a mixture of Gamma distributions with the same minimized at y* = 1.735 with a minimum value
scale parameters (Tijms [ 1111,this methodology can h( y * ) = 66.6164.
be used to find the optimal tool replacement policy By Eq. (18),
for arbitrary tool lifes.
More generally, the separability property of the 0.3( 15)
objective function in (6) can be extended to a certain ‘* = (1 - 0.3)(66.6164) = o-o965’
class of phase type distributions. The distribution
function of a phase type random variable can be and hence from (19),
written as
V * =4c)o = 161.1057.
F(x)= 1 --cy exp(Ax)e,
where A is an m X m matrix with negative diagonal In comparison under the assumption that the tool
elements and non-negative off-diagonal elements, cx life is Markovian the optimal cutting speed, as given
is a non-negative m vector, e is a unit m-vector and by (22) is
exp< . ) denotes matrix exponentiation (Neuts [6]). It
0.3
is shown in Neuts [6], that
V
(0.3)( 15)
(0.7)(15*2+5+ 105) 1
[*(l-F(x)) dx=aA-i(exp(Ax-Z)e.
= 158.7288.
68 E. lakovou et al/European Journal of Operational Research 92 (1996) 63-68

As expected, the optimal cutting speed when the El Barlow, R.E., and Proshan, F., Mathematical Theory of
tool lives are exponentially distributed is lower than Reliability, Wiley, New York, 1965.
[31 Kao, E.P.C., “Optimal replacement rules when changes of
the optimal cutting speed when the tool lives are
state are semi-Markovian”, Operations Research 21 (1973)
Gamma distributed. This is due to the fact that a 1231-1249.
more conservative operational strategy (cutting [41 La Commare, U., La Diega, N., and Passannanti, A., “Opti-
speed) is needed when the underlying stochastic mum tool replacement policies with penalty cost for unfore-
process (tool life distribution) exhibits higher vari- seen tool failure”, International Journal of Machine Tool
Design Research 23 (1983) 237-243.
ability.
[51 McCall, J.J., “Maintenance policies for stochastically failing
equipment: A survey”, Management Science 11/5 (1965)
493-524.
6. Conclusions @I Neuts, F.M., Matrix-Geometric Solutions in Stochastic Mod-
els - An Algorithmic Approach, The Johns Hopkins Univer-
sity Press, Baltimore, MD, 1981.
Models are presented for simultaneously deter- [71 Ramalingam, S., and Watson, J.D., “Tool-life distributions,
mining the optimal cutting speed and the optimal Part 1: Single-injury tool-life model”, Journal of Engineer-
tool replacement policy in cost minimization machin- ing for Industry 99 ( 1977) 5 19-522.
ing economics. 181 Ramalingam, S., “Tool-life distributions, Part 2: Multiple-in-
jury tool-life model”, Journal of Engineering for Industry
We show that when the tool lives follow a certain 102 (1977) 523-528.
class of phase type distributions, including Gamma, 191 Rembold, U., Blume, C., and Dillmann, R., Computer-In-
the objective function is separable. We exploit this tegrated Manufacturing Technology and Systems, Dekker,
structure to develop efficient solution procedures. New York, 1985.
Our findings reaffirm the need for a more conser- [lOI She&h, A.K., Kendall, L.A., and Pandit, S.M., “Probabilis-
tic optimization of multitool machining operations”, Journal
vative operational policy (lower cutting speed) when of Engineering for Industry 102 (1980) 239-246.
the underlying stochastic process (tool life distribu- [Ill Taha, H., “A policy for determining the optimal cycle length
tion) exhibits higher variability. for a cutting tool”, Journal of Industrial Engineering 17
(1966) 157-162.
1121 Taylor, F.W., “On the art of cutting metals”, Transactions
of the ASME 28 (190713 1O-350.
References [131Tijms, H.C., Stochastic Modelling and Analysis: A Computa-
tional Approach, Wiley, New York, 1986.
[l] Anderson, M.Q., “Monotone optimal preventive mainte- 041 Zhou, C., Chandra, J., and Wysk, R., “Optimal cutting tool
nance policies for stochastically failing equipment”, Naval replacement based on tool wear status”, International Jour-
Research Logistics Quarterly 28 (1981) 347-358. nal of Production Research 28 (1990) 1357- 1367.

You might also like