Sei sulla pagina 1di 8

Journal of Systems Engineering and Electronics

Vol. 19, No. 2, 2008, pp.329-336

Adaptive SPC monitoring scheme for DOE-based APC∗

Ye Liang, Pan Ershun & Xi Lifeng


Dept. of Industrial Engineering and Logistics, School of Management Engineering,
Shanghai Jiaotong Univ., Shanghai 200240, P. R. China
(Received August 3, 2006)

Abstract: Automatic process control (APC) based on design of experiment (DOE) is a cost-efficient approach
for variation reduction. The process changes both in mean and variance owing to online parameter adjustment
make it hard to apply traditional SPC charts in such DOE-based APC applied process. An adaptive SPC scheme is
developed, which can better track the process transitions and achieve the possible SPC run cost reduction when the
process is stable. The control law of SPC parameters is designed by fully utilizing the estimation properties of the
process model instead of traditionally using the data collected from the production line. An example is provided to
illustrate the proposed adaptive SPC design approach.

Keywords: automatic process control, statistical process control, robust parameter design, process transitions.

RPD-like problems. Another major contribution of


1. Introduction that article is to use regression models for automatic
Robust parameter design (RPD), pioneered by process control purpose. It is noted that most of the
Taguchi[1−2] , has been widely used in complex automatic control applications are at the machine
systems to select levels of the controllable factors level, rather than at the process level. The major
such that the system responses are robust to the obstacle to implementing the control theory at the
noise variations. It is an offline methodology based process level is the unavailability of appropriate
on design of experiment (DOE) with a priori un- process models for control design. It is extremely
derstanding of the noise factors[3] . Once the factor difficult or impossible to use differential or differ-
levels are selected, they can be used in processes ence equations to describe the relationship among
without further online adjustments. However, when the response variables, control variables, and noise
the process is running, some of the noise factors may variables for complex manufacturing processes (e.g.
be observable or measurable. Thus, it is desirable to stamping, forging, welding, etc.). In such cases, DOE
fully utilize this additional information to compen- has often been considered as an effective approach
sate the changes online by manipulating the online to get a regression model of the process. The study
controllable factors to further reduce the process provides engineers with alternative methods to realize
variance. Pledger[4] discussed an alternative approach the complex process control purpose.
of parameter design taking observable uncontrollable Statistical process control (SPC) refers to statistical
factor into consideration. Jin[5] presented a new class methods used extensively to monitor and improve the
of automatic process control (APC) approach based quality and productivity of manufacturing processes
on the DOE regression model using observable noise and service operation. SPC primarily involves the
factors for discrete-part manufacturing. One major implementation of control charts, which are used to
contribution of that article is expanding the DOE detect any change in the process that may affect the
and RPD from offline design tool to a new class of on- quality of output. The SPC methods are reasonable
line APC technique, which is capable of better solving under the assumptions that the process mean is con-
* This project was supported by the National Natural Science Foundation of China (50405016; 70671065).
330 Ye Liang, Pan Ershun & Xi Lifeng

stant and the observations are independent. How- with adaptive sample size and showed that signifi-
ever, in a DOE-based APC controlled process, the cant improvements have been achieved particularly
output mean and variance will change owing to the for small shifts. Prabhu[13] and Costa[14] suggested
shift of observable noise factors and the adjustment a combined adaptive chart with variable sample size
of controllable factors. The concept of compensating and sampling interval based on the latest obtained
in real-time will cause a number of unscheduled man- data. Costa[15] and Magalhães[16] respectively dis-
ufacturing state transition events. In this case, the cussed how to adjust all the chart parameters in X̄
traditional SPC scheme cannot be applied to moni- chart with variable parameters. In summary, almost
tor such a process with time-varying statistical pa- all kinds of adaptive SPC schemes are designed for two
rameter changes. The reason being that the control main purposes: one is to improve the detection per-
chart designed for the previous manufacturing state formance of traditional Shewart charts for small mean
will become non-effective for continuous monitoring of shifts, while the other is to reduce the sampling cost
the subsequent process. Updating the parameters of when the monitored process is greatly stable. Simi-
control charts by collecting sufficient process output larly, the proposed SPC scheme in this article aims to
data seems to be impossible. Hence, it is necessary better track the process transitions in real time, and
to develop a SPC scheme integrated with the process to achieve the possible reduction of the sampling cost.
model and the process control strategy. The informa- In this article, we intend to generate a SPC scheme
tion from the process model and the process control integrated with DOE-based APC approach, which is
law should be fully utilized to adjust the SPC param- capable of tracking the process transitions automat-
eters, which are quite different from the traditional ically by integrating the process model and applied
strategy, that is, the determination of SPC parameters control strategy, and reducing the sampling cost as
only depends on the output data from the production well.
line. We point out that traditional methods for SPC
2. Review and discussion of DOE-based
parameter adjustment are basically corrective actions.
APC
Here, we develop a real-time approach to update the
SPC parameters directly based on the manufacturing In this section, a general but necessary introduction
process model as soon as a process state transition oc- of the DOE-based APC method is given since the pro-
curs. The shift detection using the in-process sensor posed SPC monitoring strategy is designed for better
can indicate when to conduct the adjustment for SPC monitoring the DOE-based APC controlled processes.
parameters. In DOE-based APC, an expanded regression model is
In our proposed SPC scheme, the SPC parame- used to describe a discrete-part manufacturing process
ters, i.e., control limits, sample size, and sampling in- as
tervals, are designed to be adjusted online according Y = f (x, e, n) + ε (1)
to the changes of process state, which is enlightened where, Y is the process response variable, x are the
by the adaptive control charts literature. Reynolds[6] controllable factors that can be online adjusted during
and Runger and Pignatiello[7] explored the method of each production/operation cycle, e represent online
varying the time between subgroups while keeping all observable factors measured or estimated from online
other control chart parameters constant. Amin and sensing signals, n represent the remaining noise factors
Miller[8] and Reynolds[9] respectively presented a ro- that cannot be estimated or measured, and ε is the
bustness study on X̄ chart with variable sampling in- modeling error with ε ∼ N (0, σε2 ).
tervals and its monitor performance for non-normal Generally, the following second-order regression
distributed data or correlated process. Daudin[10] model is used in real case and can be expressed as
suggested a double-sampling X̄ chart similar to the
double-acceptance sampling plans. Prabhu[11] and Y =β0 + β1T x + β2T e + β3T n + xT B1 e + xT B2 n+
Costa[12] provided the design procedure of the X̄ chart xT B3 x + eT B4 n + ε (2)
Adaptive SPC monitoring scheme for DOE-based APC 331

where, x ∈ n×1 , e ∈ m×1 , n ∈ p×1 , and other and the corresponding minimum quality loss is J(x∗ ).
vectors and matrices are of appropriate dimensions. Next, the special feature of the DOE-based APC pro-
The model parameters (β’s and B’s) do not change cess output is discussed and then the reason why an
over time. alternative design of adaptive control chart should be
Then, the quadratic loss function is selected as the developed to monitor such process can be more com-
control objective function since DOE-based APC fo- prehensible. From Eqs.(6), (7), it is seen that x∗ does
cuses on the nominal-the-best problems, i.e., not hold constant and will change according to the
changes of ê during the production cycle. Thus, the
J(x) = var(Y ) + (E(Y ) − t)2 (3)
shifts both in mean and variance will occur conse-
where, var(Y ) and E(Y ) are the variance and expec- quentially when the process parameters (i.e., x and e)
tation of Y , respectively, and t is the target specified change in real time. This can be clearly observed when
by engineering design. the quality characteristic variables of the DOE-based
When the observation ê is obtained through an on- APC applied process are plotted in time sequence for
line observer, the corresponding var(Y ) and E(Y ) are SPC monitor purpose. Figure 1 follows to illustrate
the special feature of DOE-based APC applied process
Ee,n,ε (Y |x, ê) = β0 +β1T x+β2T ê+xT B1 ê+xT B3 x (4) response, which was initially presented in Jin[5] .

vare,n,ε (Y |x, ê) = (β2 + B1T x)T σê2 (β2 + B1T x)+
(β3 + B2T x + B4 ê)T σn2 (β3 + B2T x+
B4 ê) + tr(B4 σê2 B4T σn2 ) + σε2 (5)

The subscripts on E(·) and var(·) indicate the random


variables on which E(·) or var(·) operates. Therefore,
the objective function can be further expressed as Fig. 1 DOE-based APC applied stamping process

J(x) = (β0 − t + β1T x + β2T ê + xT B1 ê + xT B3 x)2 + In Fig. 1, the first 12 points correspond to the case
of thick raw material and the next 12 points corre-
(β2 + B1T x)T σê2 (β2 + B1T x)+
spond to thin raw material. In the stamping process,
(β3 + B2T x + B4 ê)T σn2 (β3 + B2T x + B4 ê)+
the thickness of the raw material is treated as the ob-
tr(B4 σê2 B4T σn2 ) + σε2 (6) servable noise factor.
where, ê is the observation of e, σê2 represents the ob- It can be found in Fig. 1 that the system response
servation uncertainty, σn2 is the variance of remaining plotted is not globally subject to identical normal dis-
noise factors, and tr(·) is the trace of a matrix. tribution but piecewise subject to different normal dis-
The optimal setting of controllable factor x∗ is de- tributions respectively. Each observable noise factor
fined as that which minimizes the quality loss. Since shift and the corresponding regulation of x will change
the regression model in Eq.(2) is only validated within process both in mean and variance. We point out that
the experimental region, any setting outside the exper- the collection of samples used to estimate the changed
imental region should be treated with caution. Hence, output distribution parameters can lead to a period of
a constraint on x is constructed to ensure that the downtime and reduce the SPC monitor continuous ef-
proposed optimal solution is within the experimental ficiency. In some cases, updating the control chart by
region. Since x is coded as a value in [−1, 1] dur- means of data collection is almost infeasible since dur-
ing the experimental design, the unit hypercube of ing the transition period, there may not be sufficient
{x : x∗ ∞  1} constrains the coded values of the response data, which are available for establishing the
control settings. Mathematically new control chart. A class of control chart, which can
automatically update its parameters to consecutively
x∗ = arg min{J(x)}, s.t. x∗ ∞  1 (7) monitor the changing process, is desired to be devel-
x
332 Ye Liang, Pan Ershun & Xi Lifeng

oped. are fixed while control limits (i.e., center line (CL), up-
per control line (UCL), and lower control line (LCL))
3. Design of adaptive control charts for
are allowed to be regulated in real time. We suggest
DOE-based APC applied process
using the conventional 3 sigma limits as a good bal-
Generally, the adaptive control chart is designed based ance of α risk (type I error) and β mistake (type II
on the last sample points, that is, if the last sample error). Practice over the years has shown that plac-
point is near the centerline, relative longer sampling ing the limits in such a way gains a better trade-off
interval, smaller sample size, and looser control limits between the economic consequences of failing to de-
are to be employed. On the contrary, if the last value tect a special cause when it does occur and wrongly
plotted is near the control limits, relative shorter sam- identifying the presence of a special cause when it has
pling interval, larger sample size, and tighter control really not occurred.
limits are to be used. All chart parameter adjustments Based on Eq. (4), the center line of control chart
are conducted mainly to achieve two goals: better de- can be easily set at the estimated value of process
tection performance and SPC cost reduction. In this mean Ŷ , which is calculated by
segment, a new class of adaptive control charts is de-
signed for the DOE-based APC applied process. The Ŷ (x∗ , ê) = β0 + β1T x∗ + β2T ê + x∗T B1 ê + x∗T B3 x∗ (8)
proposed chart design procedure quite differs from the
Based on Eq. (5), the process variance can be effec-
previous research in the adaptive control charts lit-
tively estimated by
erature in two aspects. Firstly, the adjustment law
of adaptive control chart parameters is generated di- σ̂Ŷ2 = V (x∗ , ê) + σ̂ε2 (9)
rectly based on the DOE-based APC applied process
model, rather than on the locations of last sampling V (x∗ , ê) = (β2 + B1T x∗ )T σê2 (β + B1T x∗ )+
points. Secondly, the proposed chart adaptively ad-
tr(B4 σê2 B4T σn2 ) + (β3 + B2T x∗ +
justs its parameters to eliminate the influence of pre-
B4 ê)T σn2 (β3 + B2T x∗ + B4 ê) (10)
dictable causes rather than improve the power of de-
tecting small shifts in the process mean. A general where, σ̂Ŷ2 is the estimation value of DOE-based APC
framework of the proposed adaptive control chart de- applied process variance, and V (x∗ , ê) represents the
sign is shown in Fig. 2. transmitted variance model.
As random term variance σε2 in Eq. (5) cannot be
3.1 Control limits adjustment for better tra-
estimated by just using the fitted regression model,
cking process changes
a “two-phase” method is used to build the adaptive
In this segment, an adaptive control chart is designed control chart with variable control limits. In the first
in the sense that sample size and sampling interval phase (i.e., the σε2 -estimation stage), some samples are

Fig. 2 The framework of the proposed adaptive control chart design for DOE-based APC applied process
Adaptive SPC monitoring scheme for DOE-based APC 333

collected from the production line to estimate σε2 . The soon as the current process state (x∗λ , êλ ) changes. As
calculation procedure for estimation is illustrated as a result, for a given DOE-based APC applied process,
below. the associated adaptive control chart can then be es-
(1) Calculate the mean-squared error (MSE) using tablished and applied.
collected samples
3.2 Sampling parameters adjustment for pos-
n

1 sible cost reduction
MSE = (Yi − Ŷ1 )2 (11)
n−p−1 i=1
In this section, the adaptive control chart is designed
where, n is the number of samples being used to esti- in the sense that the width of control limits is fixed
mate σε2 , p is the number of factors appearing in the while the sampling parameters (i.e., sample size and
regression model. Ŷ1 is the estimation of process mean sampling interval) are allowed to be regulated for eco-
at the σε2 -estimation stage. Note that the current level nomic design purpose. This scheme is motivated by
of regression factors in the first phase x∗1 and ê1 should the fact that if the process is more stable and in good
be substituted into Eq. (8), since x∗ and ê may change control, relatively smaller sample size and longer sam-
over time. pling interval should be used to reduce the SPC run
(2) Calculate the value of the transmitted variance cost. In this article, the sample size is adjusted so that
model at the σε2 -estimation stage, which is denoted by when the process is in the in-control state, the average
V1 = V (x∗1 , ê∗1 ). run length (ARL) of the proposed chart is the same
(3) Obtain the estimation of σε2 by under all process conditions. ARL is the expected
number of samples taken until the chart signals an
σ̂ε2 = MSE − V1 (12)
out-of-control condition, which can be calculated by
In the second phase (i.e., control chart application
phase), the adjustment law of floating control limits is ARL = 1/(1 − β)
   
UCL − µ LCL − µ (15)
to be determined and a concrete calculation procedure β=Φ √ −Φ √
σ/ n σ/ n
is presented to set up the proposed adaptive control
chart. where, β is the probability of the type II error, Φ (·)
(1) Obtain the current observation of noise factors is the cumulative probability function of the standard
denoted by êλ in real time. Then, the current optimal normal distribution, and µ and σ represent the process
settings of controllable factors x∗λ are determined using mean and variance, respectively. By substituting Eq.
the DOE-based APC approach. (14) into Eq. (15), we can further derive
(2) Calculate the current value of the transmitted
variance model Vλ by substituting x∗λ and êλ into Eq. ARL =
1
(10).  √   √ 
(3) Estimate the current process variance online by Ŷ +3(σ̂Ŷ / n)−µ Ŷ −3(σ̂Ŷ / n)−µ
1−Φ √ +Φ √
σ̂Ŷ / n σ̂Ŷ / n
σ̂Ŷ2 = Vλ + σ̂ε2 (13) (16)
In Eq. (16), the difference between Ŷ and µ is the re-
(4) The adaptive control chart can then be set up
gression model error, which is normal distributed with
using the following adjustment law for control limits.
expectation of zero. Hence, it can be seen that if ARL
CL = Ŷ (x∗λ , êλ ) is designed to be the same for various process states,

UCL = CL + 3σ̂Ŷ (14) (σ̂λ / n) should hold constant. Based on this fact, the
adjustment law for sampling parameters is developed
LCL = CL − 3σ̂Ŷ
and its design procedure is presented as below
From the above procedure, it is seen that CL, UCL, (1) Pre-calculate all the possible estimated values
and LCL will automatically update their settings as of process deviation σ̂λ under various observations of
334 Ye Liang, Pan Ershun & Xi Lifeng

noise factors. It is affordable because e has limited lev- Table 1 The factors in the injection molding
els pre-determined according to experimental design. experiment
Then, the maximal one can be found via a simple com- Controllable Observable noise Remaining noise
parison and is denoted by σ̂max . factors factors factors
(2) Determine the fixed width of control limits by A, Cycle time N, Moisture M, Percentage
±3σ̂max . Content regrind
(3) Estimate the current process variance σ̂λ2 follow- B, Mold temperature O, Ambient
ing the same procedures (1)∼(3) introduced in Section temperature

3.1. C, Cavity thickness

(4) Optimize the current sample size by D, Holding pressure


E, Injection speed
τ × σ̂ 2
nλ = 2 λ (17) F, Holding time
σ̂max G, Gate size

where, τ is the maximum allowable sample size de-


The regression model used in that article is
cided according to the available resources.
(5) If n = τ is used and holds constant, the optimal ŷ = 2.25 + 0.425A + 0.062 5C − 0.281 3D+
sampling interval can be obtained by 0.143 8E − 0.231 3G − 0.05M + 0.106 3M E−
τ ×υ 0.125M C − 0.093 8M D + 0.587 5N C−
Iλ = (18)
nλ (20)
0.556 3N E + ε
where, υ is the minimum sampling interval specified To work out the variance, a couple of assumptions
by engineer according to the production rate. have to be made. Assume that the remaining noise
(6) Set the control limits by factors M and O are subject to normal distribution
with expectations of zero and standard deviation of
CL = Ŷ (x∗λ , êλ )
0.5. Also, ε ∼ N(0, σε2 ), where, σ̂ε2 = 0.002 5 is ob-
UCL = CL + 3σ̂max (19) tained by analyzing the experimental data.
LCL = CL − 3σ̂max Therefore

From the above design procedure, it can be seen that E(y|A, C, D, E, G, Ĥ) =
if a smaller process variance (σλ2 < σmax
2
) is detected, 2.25 + 0.425A + 0.062 5C−
a smaller sample size or a longer sampling interval can 0.281 3D + 0.143 7E − 0.231 3G+
be adaptively determined and applied, and vice versa.
0.587 5N̂ C − 0.556 25N̂E (21)
4. Performance analysis of the proposed
adaptive SPC scheme var(y|A, C, D, E, G, N ) =
(−0.05 − 0.125C + 0.106 3E−
In this section, a case study is presented to demon-
strate the applicability of the proposed SPC chart and 0.093 8D)2 σM
2
+ σε2 (22)
evaluate its performance by means of ARL analysis. where, N̂ is the observation of noise factor N . Ac-
Engel[17] reports an experiment to improve an injec- cording to the DOE-based APC strategy, the control
tion molding process by minimizing the shrinkage of law for the injection molding process is given as below.
the product. Pledger[4] re-analyzed this experiment Mathematically
and changed the aim of the experiment by setting ⎧

⎪ (−1,−1,+1,−1,+1), if N̂ = +1
shrinkage to a low target (i.e., 1.6% shrinkage) rather ⎨
than minimizing it. The experiment consists of seven (A, C, D, E, G) = (−1,+1,−1,−1,+1), if N̂ = 0



controllable factors and three noise factors listed in (−1,−1,−1,−1,+1), if N̂ = −1
Table 1. All factors have two levels (i.e., +1 and −1). (23)
Adaptive SPC monitoring scheme for DOE-based APC 335

Then, estimations of process mean and variance can Markov chain. The second method is to solve an in-
be calculated for various process conditions. tegral equation to gain the ARL. The last approach
⎧ is to estimate the ARL via simulation. In this article,

⎪ [1.075, 0.006 4], if N̂ = +1
⎨ for each process condition, the injection molding pro-
[E(y), var(y)] = [1.669, 0.003 5], if N̂ = 0 cess is simulated 10 000 times using random signals



if N̂ = −1
[0.650, 0.011 3], to simulate the disturbance of noise factors to esti-
(24) mate the ARL. Both the disturbance process and the
According to the proposed adaptive control chart ARL-computing program are generated by means of
strategy, each detected shift of N̂ will trigger the ad- the software Matlab. The ARL estimation results are
justment of the control chart parameters. The ad- listed in Table 3.
justment law is obtained by the design procedure pre-
Table 3 ARL estimation results for the injection
sented in section 3 and is shown in Table 2.
molding process
Table 2 Adjustment law for adaptive control ARL
Mean shift
chart parameters Ĥ = +1 Ĥ = 0 Ĥ = −1 Average

Ĥ = +1 Ĥ = 0 Ĥ = −1 0.5σy 143 144 143 143

Adaptive control chart with variable control limits 1.0σy 46 44 43 45

and fixed sampling parameters 2.0σy 6 5 6 6

CL 1.075 1.669 0.650


The data in Table 3 lead to the following conclu-
UCL 1.315 1.846 0.969
LCL 0.835 1.491 0.331
sions:
Width of control limits 0.480 0.355 0.638 (1) For large mean shifts (i.e., 2.0σy ), ARL is equal
Sample size 5 5 5 to 6, which shows that the proposed adaptive control
Sampling interval 2 2 2 charts perform quite well (ARL<10);
Adaptive control chart with variable sampling (2) For medium mean shifts (i.e., 1.0σy ), ARL is
parameters and fixed width of control limits equal to 45, which shows that the performance of con-
CL 1.075 1.669 0.650 trol charts is well and acceptable (ARL<50).
UCL 1.394 1.988 0.969 (3) For small mean shifts (i.e., 0.5σy ), ARL is equal
LCL 0.756 1.350 0.331 to 143, which indicates that the control charts do not
Width of control limits 0.638 0.638 0.638 perform well (ARL>100).
Sample size 5 3 9 It can be seen that the performance of the proposed
Sampling interval 3.6 6.0 2.0 adaptive control charts is almost the same as that of
Shewart-type X̄ control charts. This is because the
In this case study, the fiducial sample size is set as aim of the proposed charts design is to better track
5, τ is equal to 9, and υ is 2 units of time (arbitrary the process changes that arise from predictable causes
assumptions for illustration). From Table 2, it can for application in DOE-based APC applied process
be seen that variable control limits are regulated for rather than to improve the power of detecting process
better tracking the process changes. It can also be mean shifts. If the detection of a smaller mean shift
seen that relatively smaller sample size (n = 3) and is critical, some advanced monitoring schemes such as
longer sampling interval (I = 6) are used when the EWMA charts or CUSUM charts should be integrated
process is detected to be more stable ((var(y)|N̂ = with the proposed SPC design for joint monitoring the
0) < (var(y)|N̂ = +1)) to monitor cost reduction. DOE-based APC applied process.
Generally, measures of performance for control
charts are often expressed in terms of the ARL. There
5. Conclusions
are three approaches widely used to obtain the ARL. DOE-based APC can better solve RPD-like problems
The first method is to calculate the ARL using a in the complex manufacturing process. However, on-
336 Ye Liang, Pan Ershun & Xi Lifeng

line process parameter control and observation shifts with variable sampling intervals. Journal of Quality Tech
of noise factors will change the process condition so nology, 1993, 25(1): 36–44.
that the system output is not subject to independent [9] Reynolds M R, Arnold J C, Balk J W. Variable sampling
identical distribution any more. The proposed adap- interval X charts in the presence of correlation. Journal of
tive SPC scheme can be used to solve this problem. Quality Technology, 1996, 28(1): 12–30.
The control law for UCL, LCL, and CL is designed to [10] Daudin J J. Double sampling X charts. Journal of Quality
be the function of the process state parameters, which Technology, 1992, 24(1): 78–87.
ensures the automatic control chart online update for [11] Prabhu S S, Runger G C, Keats J B. X chart with adap-
better tracking the process transitions. By use of the tive sample sizes. International Journal of Production Re-
optimal design for sample size and sampling intervals, search, 1993, 31(12): 2895–2909.
the SPC cost is reduced when the estimated process [12] Costa A F B. X charts with variable sample size. Journal
variance indicates that the process is relatively more of Quality Technology, 1994, 26(3): 155–163.
stable. In a broad sense, this research is another good [13] Prabhu S S, Montgomery D C, Runger G C. A combined
attempt to generate an optimization design for SPC adaptive sample size and sampling interval X chart scheme.
based on the process model. More effect is required Journal of Quality Technology, 1994, 26(3): 164–176.
to introduce other advanced chart techniques into the [14] Costa A F B. X chart with variable sample size and sam-
proposed chart design for better detection power of pling intervals. Journal of Quality Technology, 1997, 29(2):
mean shifts in the process. 197–204.
References [15] Costa A F B. X charts with variable parameters. Journal
of Quality Technology, 1999, 31(4): 408–416.
[1] Taguchi G. Introduction to quality engineering. NJ: Qual-
[16] Magalhães M S De, Costa A F B, Epprecht E K. Con-
ity Resources, White Plains, 1986.
strained optimization model for the design of an adaptive
[2] Taguchi G. System of experimental design: engineering
X chart. International Journal of Production Research,
methods to optimize quality and minimize cost. NJ: Qual-
2002, 40(4): 3199–3218.
ity Resources, White Plains, 1987.
[17] Engel J. Modeling variation in industrial experiments.
[3] Montgomery D C. Design and analysis of experiments. (3rd
Applied Statistics, 1992, 41(3): 579–593.
ed). NY: John Wiley & Sons, 1991.
[4] Pledger M. Observable uncontrollable factor in parameter
Ye Liang was born in 1981. He is a Ph. D. student
design. Journal of Quality Technology, 1996, 28(2): 153–
at the Department of Industrial Engineering and Lo-
162.
gistics of Shanghai Jiaotong University. His research
[5] Jin J H, Ding Y. Online automatic process control using
focuses on quality control, SPC, DOE, and robust de-
observable noise factors for discrete-part manufacturing,
sign. E-mail: schneideye@sjtu.edu.cn
IIE Transactions, 2004, 36(9): 899–911.
[6] Reynolds M R, Amin R W, Arnold J C, et al. X-charts Pan Ershun is an associate professor at the Depart-
with variable sampling intervals. Technometrics, 1988(2), ment of Industrial Engineering and Logistics of Shang-
30: 181–192. hai Jiaotong University. E-mail: pes@sjtu.edu.cn
[7] Runger G C, Pignatiello J J. Adaptive sampling for process
control. Journal of Quality Technology, 1991, 23(2): 135– Xi Lifeng is a professor at the Department of Indus-
155. trial Engineering and Logistics of Shanghai Jiaotong
[8] Amin R W, Miller R W. A robustness study of X charts University. E-mail: lfxi@sjtu.edu.cn

Potrebbero piacerti anche