Professional Documents
Culture Documents
1046-1051,
1990
0895-7177/90 S3.00+0.00
Pergamon Press plc
Printed in GreatBritain
DETERMINATION
AS A MFlRKOVIFIN
Carlos
Ford-Livene
and
Tadao
Mukaihata
Electra-Optical & Data Systems Group Technology Support Division Primary Standards Laborartory Hughes Aircraft Co., El Segundo, California
The problem of determining the optimum adjustment to Abstract. a calibrat on renewal interval of a device under test is treated of a Markovian decision process. The model within the context the total expected adjustment to the calibration calculates the device as it goes from one state to the next. interval 0 of the device are: within specification The possib le states , out-of-specification, and non-significant.
INTRODUCTION Calibration interval adjustment programs (CIcIPs) characterized in terms of probability density functions present the following difficulties: 1. Deriving an accurate mathematical model of complex joint probability density functions involved in d particular CIAP. Relating interactive, multiple parameters to varying probabilities associated with the model. Accounting for probabilities significantly affected by diversified equipment utilization and applications of the end user.
1.
or
in-
2.
or out-
3.
2.
3.
These difficulties are resolved by developing a model based on the Markovian process and empirical data to provide a-priori probabilities for three basic calibration states. A significant feature is the inclusion of the effects of equipment utilization, application and aging which occur These effects between calibrations. alter the probabilities associated with the three states, thus, emphasizing the importance of the utilization/applicaThis latter set of tion/aging factors. data is not available unless the end user interfaces with the system to Another beneficial aspect provide it. is that this model is adaptable to most systems by providing the pertinent athe model Also, priori probabilities. allows for varying the size of interval based on sound naturally, changes, rationale.
Within the context of a simple Markov Process (MP), the probability p+, (i,j = 1,2,3) of a transition to state j during the next calibration interval, given that the system previously occuof i and j pied state i, is a function and not of any history of the device before its arrival in i. Thus, the P+J represent a set of conditional probabilities that a device which previously was in state i will occupy state j after the next transition. Indeed must be in i P+J = 1, since the device 'Gome state after the next transition. The matrix P~[p+,l will enable us to answer certain questions concerning the we may wish to know For example, CIAP. the probability that the device will be in state 2 (out-of-tolerance) after n stages given that it was in state 2 at the beginning of the n-stage process.
Fig.
1 of
MODEL Let us suppose there are 3 states in the CICIP in which a device may be found, namely 1046
Proc. 7th Inr. Conf. on Mathematical To answer this and similar questions we define a state probability o,(n) as the probability that the device occupies state j after n stages if its state at
n-0 is known. Then
and Computer
Modelling
a
1047
3 2 o+ (n) i=l
and
as the device makes transitions from state to state. A problem of interest is to determine the total expected change or adjustment to the calibration recall interval, given that the device goes from state i to state j in n-stages. A stage, for example, could mean each time the device is used by the enduser. To do this, let hi(n) denote the expected total changes in the next n stages if the system is known to be previously in state i. Following Howard (19601, we express 3
h+(n) = Z pLISIl
(2)
O,l,Z,... of corn--
If we now consider a row vector state probabilities, o(n), with pot-tents o+(n), then
j=l o(n+ll = o(nlP, " = 0,1,2,... (31 i = 1,2,3 and recursively, and o(n) = O(OlP", n = 0,1,2,... (41 3 92 = 2 PrJ6+J j=l define the
3 Z p,,hr(n-11 j=l
(51
n = 1,2,3,...
quantity
q,
by
i = 1,2,3
(61
We suppose next that our TIP is a completely ergodic process, i.e., one whose limiting state probability distribution o(n) is independent of the starting state of the device. We can then define a quantity (rL as the probability that the device occupies the ith state after a large number of stages. The row vector os(ol oz ~~1 is thus the CJ is referred limit of o(n) as I-I+=. to as the vector of limiting or absolute state probability (Howard, 19601. Since o(n+l) = o(nlP for all n, we have that
o(n)P = 0P
makes a transition Now, as the device from state i to state j, it induces a change &+, in the calibration recall cycle plus the amount of change it expects to induce if it starts in state j with one fewer stage remaining. These transitional changes from i to j must be weighted by the probability of such a transition, pLlr to obtain the total expected (net) changes. Equation (5) can now be expressed in the form
lim n+m
0 n+ll
= (I = lim
hL(nl
= q,
n+m
3 + X pi,hj(n-11 j=l
(71
x
i=l
02
1 always.
When the device under test makes a transition from state i to state j, our 3-state Markov process induces a change 6+, in the calibration recall interval. We call 6&, the "change" associated with a transition from state i (i=l,Z,J) to state j (j=1,2,31. Here, either er, > 0, (an increase in the length of the interval) in the 6+1 < 0, (a decrease length of the interval) in the 6+, = 0, (no change length of the interval1
The q, may be interpreted as the (net) change in the interval (e.g., change in number of days, weeks or months) to be expected in the next transition out of state i; it is referred to as the expected immediate change associated In vector form, (71 may with state i. be written as h(n) = s + Ph (n-11 (81
n = 1,2,3,... where h(n) is a column vector with comcalled the net or total ponents h+(n), Consider, for examadjustment vector. ple, the following CICIP chart which may be used to decide initially whether to decrease or leave unchanged increase, the calibration recall interval of a To focus attention device under test. on a particular device we will assume the device is a thermistor mount.
1048
PRESENT
OF EOUIPMENT
CHANGE CALIBRATION
IN It INTERVPlL
WITHIN WITHIN
SPECIFICATION SPECIFICfiTION
WITHIN
SPECIFICATION OR
ADD NO
2 MONTHS CHANGE
WITHIN
SPECIFICfiTION
NO
CHANGE 1 MONTH*
OUT OF SPECIFICATION RISK DEFECTIVE OUT OF SPECIFICATION RISK DEFECTIVE OUT OF SPECIFICATION RISK DEFECTIVE NON-SIGNIFICANT NON-SIGNIFICANT
SUBTRXT
OR
OR
SUBTRACT
2 MONTHSI:
OR
SUBTRACT
1 MONTHX
WITHIN
SPECIFICfiTION OR
NO CHANGE NO CHANGE
NON
SIGNIFIIXNT
NO CHANGE
$ One
month
as
21 manufacturing
Minimum
is to be 043
t The values for the increments/decrements circumstances due to enduser utilization well as device characteristics.
accordingly, the 3 states device in the CIAP are: State State State and 6,, 611 63x
of
the
1 - Within specification 2 - Out of specification 3 - Non-significant = +2, = 0, = 0, 6+= = -1, 6zr = -2, 6 xz = -1, refer 6+= = 0 6zirr= 0 653 = 0
0.50, 0.251 and CS,,l = CO, -2, 01, If it is in State 3 respectively. they are Cp3,1 = C0.4, 0.4, 0.21 and Then S = C1.0, C63,l = CO, -1, 01. and (7) enables us to -1.0, -0.4lT, recursively calculate the total expected adjustments as the device transitions from State i (i = 1,2,3) convergence State j (j = l,Z,J) until as illustrated in Table is attained,
to 1.
subscripts
Define
D = [6r,l
= [
0 0
and
take
-1 0 1
to the -20 -1 , 0 l/5 l/2 2/S l/5 l/4 l/5 ]
Table
n
0 1
P = Cp+,l
s;
based
on
historical
records.
If the thermistor mount is in State 1 (in spec) previously, it makes transitions to State 1, State 2 (out-ofspec), or State 3 (non-sig) according to a probability distribution Cp+~l = CO.6, 0.2, 0.21 and induces a corresponding change C61~1 = C+2, -1, 01. if it is in State 2, the Similarly, transition probability and corresponding change are C~r+~l = C0.25,
2 3 4 5 6 7 B 9 10 11 12 13 14 15 16 17 18 19 20
hi(n) 0.0000000 1.0000000 1.3200001 1.4260001 1.4610002 1.4725602 1.4763782 1.4776392 1.4780557 1.47B1933 1.4782387 1.4702537 1.4782586 1.4782603 1.4782609 1.4782610 1.4782611 1.4782612 1.4782612 1.4782612 1.4782612
hZ(n) 0.0000000 -1.0000000 -1.3500000 -1.4650000 -1.5030000 -1.5155499 -1.5196948 -1.5210638 -1.5215160 -1.5216653 -1.5217147 -1.5217309 -1.5217363 -1.5217381 -1.5217386 -1.5217388 -1.5217389 -1.5217389 -1.5217389 -1.5217389 -1.5217389
h3(n) 0.0000000 -0.4000000 -0.4000000 -0.5000000 -0.5172000 -0.5202399 -0.5212439 -0.5215754 -0.5216849 -0.5217211 -0.5217330 -0.5217370 -0.5217383 -0.5217387 -0.5217309 -0.5217309 -0.5217389 -0.5217389 -0.5217389 -0.5217389 -0.5217389
1049
CIAP
age t (in years) of a device under test is likely to affect the probabilities and changes which govern its CI&P. We will consider three age categories, hereafter referred to as "cases" namely: l(t(4, 4<t<a, 8 < ti12
Corresponding to each Case k, k = 1,2,3 are transitional probabilities, pjr, based on historical records, which take the device from State i to State j, The associated adjustment i,j = 1,2,3. matrix D=[&+,J will be the same as before and is normally independent of the device's age. In carrying out the calculations, the following table of values will be used for the thermistor mount.
TABLE
F\d justments
6LL 2 2 2 0 0 0 0 0 0
6zzz -1 -1 -1 -2 -2 -2 -1 -1 -1
6i1 0 0 0 0 0 0 0 0 0
Expected Immediate Adjustment qr 1.25 1.00 0.40 -1.00 -1.00 -1.60 -0.20 -0.50 -0.20
For each case k, k=1,2,3, we wish to determine the total expected adjustment to the calibration interval over an n-stage process as a function of the present state of the thermistor mount. Equation (7) becomes 3 + aXp$,h,(n), j=l
# The results, for a=0.75, lit<4 and a=0.55, a<t<l2, are given in Tables 3 and 4, respectively.
3 = 0.750
alpha
n = 0,1,2,...
k=
1,2,3
Total expected adjustments (+, -, 01 to calibration interval of device under test as a function of device condition and number of periods remaining.
where we have introduced a factor a, O<a<l to reflect uncertainty concerning duration of the process. Here a is defined as the probability that the process will continue to accumulate changes after the next transition. The factor a may instead represent a discounting of future adjustments to the CIAP to reflect device fatigue and stress for a process of indefinite duration. Thus, (9) tells us what the expected future adjustments at each stage of the process will be corresponding to each age category of the device. To start the iteration, we must assign a value to the boundary condition h$COl. We shall take h,(O)=0 as before*.
n
0
hi(n)
0.0000000
h2Ct-t)
0.0000000
-1.0000000
1 2 3 4 5 6 7 e 9 10 11 12 13 14 15 16
1.2500000 1.7712500 2.0740156 2.2036559 2.4377525 2.5528464 2.6391109 2.7030274 2.7523630 2.7887666 2.8160698 2.0365474 2.0519056 2.8634243 2.0720632 2.8785424
-0.9062500 -0.6756250 -0.4756035 -0.3219004 -0.2065055 -0.1201521 -0.0553075 -0.0060350 0.0295730 0.0568773 0.0773552 0.0927134 0.1042321 0.1128711 0.1193504
h3(nl 0.0000000 -0.2000000 -0.0350000 0.1849375 0.3763422 0.5266589 0.6410264 0.7271740 0.7910651 0.8403996 0.8768036 0.9041070 0.9245047 0.9399430 0.9514615 0.9601005 0.9665798
1050 Table n 17 10 19 20 21 22 23 24 25 26 27 20 29 30 31 32 33 34 35 36 37 30 39 40 41 42 43 44 45 46 47 40 49 50 hi(n) 2.0034019 2.0070466 2.0097000 2.0910300 2.0933670 2.0945210 2.0953057 2.0960342 2.0965206 2.0960054 2.0971591 2.0973644 2.0975104 2.0976330 2.0977206 2.0977057 2.0970343 2.0970700 2.0970902 2.8979107 2.0979340 2.0979456 2.0979542 2.0979609 2.0979657 2.0979692 2.0979721 2.0979740 2.0979757 2.0979769 2.0979776 2.0979703 2.0979700 2.0979793 3
Proc. 7th Int. Conf on Mathematical and Computer Modelling (Cont'd) h3(n) 0.9714393 0.9750030 0.9770172 0.9790673 0.9014049 0.9025501 0.9034231 0.9840717 0.9045502 0.9049229 0.9051965 0.9054017 0.9055550 0.9056712 0.9057570 0.9050227 0.9050715 0.9059002 0.9059355 0.9059561 0.9059715 0.9059029 0.9059917 0.9059901 0.9060030 0.9060066 0.9060094 0.9060114 0.9060129 0.9060141 0.9060150 0.9060156 0.9060162 0.9060165
The total interval
CONCLUDING
REMARKS
the
h2(n) 0.1242090 0.1270544 0.1305070 0.1326379 0.1341755 0.1353207 0.1361937 0.1360423 0.1373207 0.1376935 0.1379671 0.1301723 0.1303263 0.1304410 0.1305203 0.1305933 0.1306422 0.1306706 0.1307060 0.1307266 0.1307419 0.1307535 0.1387621 0.1307607 0.1307736 0.1307773 0.1307799 0.1307020 0.1307035 0.1307047 0.1307056 0.1387062 0.1307067 0.1307071
of
stages n in the iterativ@ process; this quantity generally grows without bound unless convergence is attained. In Table 1 convergence of h(n) occur-s for n=lO, without discounting. In Tables 3 and 4, it converges for n=33 and n=19, respectively, when discounting is applied. We say convergence is attained after n iterations in case the difference Ih*(n) - h'(n-l)l is less than some prescribed small quantity E>O; here we have taken s=5x10-5. example, according to Table 3, if the thermistor mount is previously in State 1 (in-spec) the expected adjustment is to increase the calibration renewal interval by approximately 2.9 months. If previously in State 2 (out-of-spec) the interval should be increased by only about 3 days. And, if it is previously in State 3 (nonsignificant) the interval should be increased by about 1 month. It is important to note that the amount of adjustment is based on the p+,'s, the discounting or fatigue factor Q, and the age level of the particular device under consideration. For processes requiring a large number of iterations (slow convergence), a mot-e practical approach is to take the average adjustment to the interval per unit time. This quantity, called the "gain" G, i s meaningful if the process is allowed to go through many stages n. The gain o f the process is defined as
For
4 = 0.550
alpha
Total expected adjustments (+, -, 0) to calibration interval of device undertest as a function of device condition
and number of periods remaining. hi(n) 0 .ooooooo hZ(n) 0.0000000
3 G = Z oiq+ i=l
, where
3 q, = x j=l
P136L3.
(10)
n
0
;
4 5 6 7 a 9 10 11 12 13 14 15 16 17 10 19 20 21 22 23 24 25
0.4000000 0.1140000 -0.1231600 -0.2673073 -0.3492383 -0.3940059 -0.4201374 -0.4340640 -0.4417361 -0.4459504 -0.4402017 -0.4495597 -0.4502627 -0.4506495 -0.4500621 -0.4509791 -0.4510434 -0.4510700 -0.4510903 -0.4511090 -0.4511149 -0.4511101 -0.4511199 -0.4511209 -0.4511214
-1.6000000 -2.2930000 -2.6245401 -2.7929700 -2.0016001 -2.9291630 -2.9549737 -2.9690664 -2.9767064 -2.9010233 -2.9033500 -2.9046301 -2.9053337 -2.9057204 -2.9059331 -2.9060501 -2.9061145 -2.9061500 -2.9061693 -2.9061001 -2.9061060 -2.9061091 -2.9061910 -2.9061920 -2.9861925
h3(n) 0.0000000 -0.2000000 -0.3900000 -0.5710300 -0.6906069 -0.7645501 -0.0076964 -0.0321052 -0.0450834 -0.0534060 -0.0576079 -0.0600050 -0.0612013 -0.0619037 -0.0623703 -0.0625029 -0.0626999 -0.0627642 -0.0627995 -0.0620190 -0.0620297 -0.0620356 -0.0620309 -0.0620407 -0.0620416 -0.0620421
It can ergodic
sesses
be shown MP with
a gain
that
every
completely
6L, pos-
above It is particularly (Howard, 1960). convenient to use the gain when several processes at-e involved, as occurs in metrology initially when a brand new device ot- type of device arrives and a calibration recall interval has to be Normally, no perestablished for it. formance history is available on the new model to aid in determining an However, there initial recall cycle. are usually several options or processes available for starting a CIAP. the gain G To do this, we calculate corresponding to each process then select the one with highest gain to determine which process yields the largest calibration interval over a (Ford-Livene, long sequence of stages. Mukaihata, 1909).
adjustments G as expressed
Proc. 7th Int. Conf. on Mathematical and Computer Modelling REFERENCES John Caldwell, Delbert H. and Ferling, A Analytic Modeling for Elect;&ic Test Equipment Adjustment Policies, Proceedinas of the 1989 National Conference of Standards Laboratories, Denver, CC. Castrup, Howard T., A Calibration Interval Analysis System Care Proceedinas of the 1988 Study, National Conference of Standards Laboratories, Washington, D.C. Cyrus, Finite State Markovian Derman, Decision Processes, Academic Press, 1970
1051
Ford-Livene, Carlos and Mukaihata, Tadao, Dynamic Programming Approach to Initial Calibration Interval Determination, Proceedinqs of the Seventh International Conference on Mathematical and Cornouter Modellinq, Chicago, IL, 1909 Howard, Ronald CI., Dynamic Proaramminq and Markov Processes, MIT Press and John Wiley & Sons, jointly, 1960 Larsen, John C., The Calibration Interval System of the U.S. Navy, Proceedinas of the 1988 National Conference of Standards Laboratories, Washington, D.C.