132 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO.

1, FEBRUARY 2013

VM-Based Baseline Predictive Maintenance Scheme

Yao-Sheng Hsieh, Student Member, IEEE, Fan-Tien Cheng, Fellow, IEEE, Hsien-Cheng Huang, Chung-Ren Wang,
Saint-Chi Wang, and Haw-Ching Yang, Member, IEEE

Abstract—Most conventional FDC approaches are used to find C&H Concise and healthy.
the TDs required for monitoring and the TDs’ related key CbM Condition-based maintenance.
parameters that need to be monitored, and then apply the SPC
DHI Device health index.
approach to detect the faults. However, in a practical situation,
an abnormal key-parameter value may not be caused solely by DHIT Threshold of DHI.
its own TD; it may result from the other related parameters. DMW Dynamic moving window.
Therefore, accurate fault classification or diagnosis may not be ECF Exponential curve fitting.
achieved. Moreover, most conventional PdM methods require FDC Fault detection and classification.
a library of degradation patterns from previous run-to-failure
GSI Global similarity index.
data sets. Without those massive historical failure data, the PdM
methods may not function properly. In this paper, we propose a GSIT Threshold of GSI.
virtual-metrology- (VM) based BPM scheme that possesses the ISIB Baseline individual similarity index.
capabilities of FDC and PdM. The BPM scheme contains the TD ISIB− T Threshold of ISIB .
baseline model, FDC logic, and a RUL predictive module. The KIS Keep important sample.
TD baseline model generated by the VM technique is applied
LB Lower bound.
to serve as the reference for detecting the fault. By applying the
BPM scheme, fault diagnosis and prognosis can be accomplished, LCL Lower control limit.
the problem of the conventional SPC method mentioned above LSL Lower spec limit.
can be resolved, and the requirement of massive historical failure MTBF Mean time between failure.
data can also be released. NH3 Ammonia.
Index Terms—Automatic virtual metrology (AVM), baseline OOC Out of control.
predictive maintenance (BPM) scheme, dynamic-moving-window PdM Predictive maintenance.
(DMW) scheme, fault detection and classification (FDC), keep PI Prediction interval.
important sample (KIS) scheme, predictive maintenance (PdM).
PECVD Plasma-enhanced chemical vapor deposi-
tion.
PM Preventive maintenance.
Nomenclature PPM Predictive and preventive maintenance.
Abbreviation List RF Radio frequency.
RUL Remaining useful life.
AVM Automatic virtual metrology. SiH4 Silane.
BDM Breakdown maintenance. SPC Statistical process control.
BEI Baseline error index. TD Target device.
BEIT Threshold of BEI. UB Upper bound.
BPM Baseline predictive maintenance. UCL Upper control limit.
Manuscript received April 28, 2012; revised August 3, 2012; accepted Au- USL Upper spec limit.
gust 27, 2012. Date of publication September 17, 2012; date of current version
January 30, 2013. This work was supported in part by the National Science
Council, Taiwan, under Contracts NSC100-2622-E-006-011-CC2, NSC100-
2221-E-006-127-MY2, NSC101-2218-E-006-019, and NSC101-2622-E-006- List of Symbols in Equations
002-CC2, and in part by the Ministry of Education of Taiwan under Project
AIM-HI. There are currently patents pending for the work presented in this
paper in Taiwan, the U.S., China, Japan, and Korea with application no. yT Actual value of TD.
101126242. yT Mean of yT .
Y.-S. Hsieh, F.-T. Cheng, H.-C. Huang, C.-R. Wang, and S.-C. ŷB Baseline predictive value of TD.
Wang are with the Institute of Manufacturing Information and Sys-
tems, National Cheng Kung University, Tainan 701, Taiwan (e-mail: john- yE Absolute value of “ŷB minus yT .”
[email protected]; [email protected]; wilson@ yE− S Sick value of yE .
super.ime.ncku.edu.tw; [email protected]; saint− [email protected]. yE Mean of yE .
edu.tw).
H.-C. Yang is with Institute of System Information and Control, National σyE Standard deviation of yE .
Kaohsiung First University of Science and Technology, Kaohsiung 811, yE− B Begin value of yE .
Taiwan (e-mail: [email protected]). yE− D Dead value of yE .
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. ki ith sample number.
Digital Object Identifier 10.1109/TSM.2012.2218837 ŷEi ith yE predictive value corresponding to ki .
0278-0070/$31.00 
c 2012 IEEE
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 133

yE i ith yE actual value corresponding to ki .

kB Bth sample number corresponding to yE− B .
kS Sth sample number corresponding to yE− S .
A Interception value of the ECF formula.
C Slope value of the ECF formula.
ŷEi− UBS+j Predictive S+jth UB value.
ŷEi− LBS+j Predictive S+jth LB value.
k̂D− RUL D− RULth sample number corresponding to
yE− D .
S+j 95% PI value corresponding to the S+jth
sample.
t α2 (kS −kB +1−p) t-distribution with kS −kB +1−p degrees of
freedom with p being the number of param- Fig. 1. SPC control chart of throttle-valve angles in a PECVD tool.
eters and α = 0.05.
k̂D− UB D− UBth sample number corresponding to model. However, it is usually difficult to develop accurate
yE− D . fault-growth models in most real-world applications [6].
k̂D− LB D− LBth sample number corresponding to In many applications, measured input/output data is the
yE− D . major source of knowledge to understand the system degra-
dation behavior. The data-driven approaches rely on the as-
sumption that the statistical characteristics of data are rela-
I. Introduction tively consistent unless a malfunctioning event occurs in the

P RODUCTION equipment is an essential part of any man-

ufacturing factory. Failure of a component, module, or
device (e.g., heater, pressure module, and throttle valve) in the
system. Without an understanding of the physics of the fault
progression, data-driven approaches are one of the feasible
options [6].
production equipment may cause production abnormalities that Most conventional FDC approaches are to find out the TDs
lead to poor product quality and/or low production capacity, required for monitoring and the TDs’ related key parameters
and thus cause significant losses. needed to be monitored, and then by applying the SPC
In general, the most commonly utilized approach for rem- approach [7] to detect the faults. As described in [8], several
edying the problems mentioned above is scheduled PM, i.e., researchers utilized model-based algorithms to define the key-
to execute maintenance-related operations in a predetermined parameters’ SPC control limits that detect the faults. The
time interval. This time interval is basically decided according methods proposed in [9] and [10] are two examples of the
to the MTBF of the TD [1]. As such, how to schedule a conventional FDC approaches mentioned above.
proper PM is usually a key issue for the factories. An improper However, in a practical situation, an abnormal key-
scheduled PM may increase the maintenance cost or lower the parameter value may not be solely caused by its own TD;
production capacity [2]. instead, it may also due to the influence of other related
To improve equipment maintenance programs for increasing parameters. Take practical PECVD throttle-valve angles for
fab performance, the International SEMATECH Manufactur- example. As shown in Fig. 1, the center is at 27° and the
ing Initiative (ISMI) proposed an initiative of PPM [1]–[5]. As UCL and LCL are 32° and 22°, respectively, as defined
defined by ISMI, PPM comprises PM, CbM, PdM, and BDM. by the maintenance engineers. A total of 450 samples are
Among them, ISMI claimed that CbM and PdM capabilities monitored. As indicated in Fig. 1, this conventional SPC
should be developed and available as individual or incremental method concludes that those samples in red circles 1, 2, and
modules so that the end user can choose to implement one, 4 are outliers while the sample in green circle 3 is within the
some, or all the capabilities. CbM is the approach to activate control limit.
the maintenance task after one or more indicators show that After careful inspections, the fact is that those samples in
equipment is going to fail or that equipment’s performance red circles 2 and 4 are indeed abnormal and are caused by the
is deteriorating [1]. The technique of FDC is an approach throttle-valve’s malfunction. As for the sample in red circle 1,
related to CbM. PdM, on the other hand, is the technique of this abnormality is not caused by the throttle-valve itself; it
applying a predictive model to relate facility-state information results from the deviation of the related parameter “ammonia
to maintenance information for forecasting the RUL and the (NH3 ).” Also, the deflection shown in the green circle 3 is due
need for maintenance events that will alleviate unscheduled to the deviation of the related parameter “tube pressure.” As
downtime. such, the conventional SPC method cannot detect and diagnose
As mentioned in [6], intelligent fault diagnostic and prog- the faults at the samples in red circle 1 and green circle 3. The
nostic methods can be roughly categorized into two major authors presented a preliminary study [11] of a VM-based
classes: model-based and data-driven methods. Model-based FDC scheme to remedy the problem mentioned above.
approaches require an accurate mathematical model to be Liu et al. [12] used the similarity-based method for the
developed and use residuals as features, where residuals are spindle load prediction and diagnosis. They created plural
the outcomes of consistency checks between the sensed mea- match matrices utilizing a lot of historical information. The
surements of a real system and the outputs of a mathematical central concept is to have a library of degradation patterns from
134 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

Fig. 2. BPM scheme.

previous run-to-failure data sets and estimate the current re-

maining life by matching the current degradation pattern with
the most similar ones in the degradation library. Kim et al. [13]
utilized the support-vector-machine classifier as a tool for
estimating health-state probability of machine degradation
process. Through prior analysis of historical failure data,
discrete failure-degradation stages were employed to estimate
the health-state probability for accurate long-term machine
prognosis. Fig. 3. State diagram of a device.
The RUL predictive methods utilized in [12] and [13]
require a library of degradation patterns from previous run-to-
failure data sets. Without those massive historical failure data, and BEI modules, which are utilized to generate the baseline
these two RUL predictive methods may not function properly. of TD (ŷB ), the indexes of ISIB , DHI, and BEI, respectively.
The purpose of this paper is to propose a VM-based BPM The right-hand part defines FDC logic to convert DHI, BEI,
scheme. The BPM scheme possesses the capabilities of not and ISIB values into the healthy status of TD.
only FDC but also PdM. The BPM scheme is composed of The TD baseline model predicts ŷB through collecting the
the TD baseline model, FDC logic, and a RUL predictive TD-related process data (X), while the TD data (yT ) is applied
module. The TD baseline model is generated by referring for tuning or retraining the TD baseline model. The ISIB
to the technique of AVM [14] to serve as the reference model transforms X into the ISIB value. Moreover, the DHI
for detecting the fault. By applying this BPM scheme, fault module converts |yT −yT | into DHI index where yT represents
diagnosis and prognosis can be accomplished, the problem the mean of fresh samples of yT , and the BEI module converts
of the conventional SPC method mentioned above can be yE (= |yT –ŷB |) into the BEI index. At last, DHI, BEI, and ISIB
resolved, and the requirement of massive historical failure data values are sent to the FDC logic to infer the healthy status of
can also be released. TD.
The remainder of this paper is organized as follows. Sec- Before applying the FDC logic, the thresholds of DHI, BEI,
tion II details the BPM scheme. Section III then presents the and ISIB (denoted by DHIT , BEIT , and ISIB− T , respectively)
illustrative examples. Finally, summary and conclusions are should be assigned. DHIT and BEIT are utilized to judge
drawn in Section IV. whether the TD is in the sick state or not. How to define the
values of DHIT and BEIT will be explained as follows. The
value of ISIB− T will be defined in Section II-E.
II. BPM Scheme The state diagram of a device proposed in [15] is shown in
As shown in Fig. 2, the BPM possesses two portions: FDC Fig. 3, which includes five states: initial, active, inactive, sick,
and PdM. The FDC portion consists of two parts. The left-hand and dead. Usually, a device is in the active state. However,
part generates TD and ISIB baseline models as well as DHI when yE is larger than its threshold (yE− S ), the device enters
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 135

the sick state. The device returns to the active state when yE
becomes less than yE− S again. On the contrary, if the sick
symptom of the device gets worse, such that the available
resources of the device are exhausted, the device enters the
dead state; in other words, the device is down.
For the sake of convenience, the values of DHI and BEI
are normalized to varying from 1 to 0; higher value indicates
healthier TD. The key transitions in Fig. 3 are from “Active to
Sick” and from “Sick to Dead.” In order to match the DHI and
BEI values with those two key transitions, the following rules
(which sets the thresholds from “Active to Sick” and from
“Sick to Dead” to be 0.7 and 0.3, respectively) are proposed:
1) when 1 > DHI/BEI > 0.7, TD is healthy and normally
operated;
2) when 0.7 > DHI/BEI > 0.3, TD is sick, it cannot work
after its RUL is worn out;
3) when 0.3 > DHI/BEI > 0, TD is dead, it needs
immediate maintenance.
According to the rules above, the values of DHIT and
BEIT are defined to be 0.7. Note that it is just an example to
assign the thresholds from “Active to Sick” and from “Sick
to Dead” to be 0.7 and 0.3, respectively, in this paper. Users
may reassign any proper numbers between 0 and 1 (such as
0.8 and 0.2) to be the thresholds from “Active to Sick” and
from “Sick to Dead” according to their requirements.
The details of TD baseline model, ISIB model, DHI module, Fig. 4. Procedure of collecting the important samples needed for creating
and BEI module are presented in Sections II-A to II-D. TD baseline model.
Illustration of the FDC logic and flowchart of the baseline FDC
The online operations are executed just after maintenance
execution procedure are described in Sections II-E and II-F,
and contain three steps: 1) Step 3 in Fig. 4 to collect the
respectively.
fresh samples that just finished the maintenance process;
The PdM portion of the BPM scheme also consists of two
2) Step 4 in Fig. 4 to check whether the number of fresh and
parts. The upper part contains the TD baseline model while
healthy samples collected is enough for modeling or not; and
the lower part includes a RUL predictive module, which will
3) Step 5 in Fig. 4 if enough samples are collected, to add
be delineated in Section II-G.
those fresh samples to the C&H historical samples produced
in Step 2 of Fig. 4.
A. Important Samples Needed for Creating TD Baseline The rule of thumb for determining the number of samples
Model needed for creating the baseline model is about 10 times the
The purpose of the TD baseline model is to generate the number of the related process data, and the ratio between the
healthy baseline of the TD (ŷB ). To begin with, the procedure C&H historical samples and the fresh samples are 3 to 1. All
shown in Fig. 4 for collecting the necessary important samples the samples created in Step 5 of Fig. 4 are utilized to construct
to create the TD baseline model is performed. The TD baseline the TD baseline model. Therefore, this baseline model not only
model-creation procedure consists of two stages: 1) offline contains the C&H samples but also possesses fresh data just
operations by performing the KIS scheme to select the C&H after maintenance such that the healthy baseline of the TD
historical samples, and 2) online operations to collect the fresh can be generated. With the necessary important samples being
samples just after maintenance, with each sample including completely collected, the TD baseline model can be built and it
yT and its corresponding X. The offline KIS scheme includes can function properly. The execution procedure of TD baseline
two steps: 1) Step 1 in Fig. 4 to ensure the appropriateness model will be explained in Section II-F.
of every historical sample collected, which means the sample
was generated under healthy status and its data quality is good; B. Samples Needed for Creating ISI B Model
and 2) Step 2 in Fig. 4 to utilize the DMW scheme [16] for As shown in Fig. 2, except for the TD baseline model,
picking out the C&H samples from all of the healthy historical the ISIB model should also be created in the box of baseline
samples selected in Step 1. The DMW scheme [16] adds a new models. The authors invented the concept of the ISI in [17] to
sample into the model and applies a clustering technology to identify the key-process parameters that cause major deviation
do similarity clustering. Next, the number of elements in each when the GSI of the input set of process parameters exceeds
cluster is checked. If the largest number of the elements is its threshold (GSIT ) in an AVM system [14].
greater than the predefined threshold, then the oldest sample ISI of an individual process parameter is defined as the de-
in the cluster of the largest population is deleted. gree of similarity between this individual process-parameter’s
136 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

standardized process datum of the input set and the same

process-parameter’s standardized process data in all the his-
torical sets that are used for training and tuning the ISI model
[17]. In this paper, the fresh process data (X) collected just
after maintenance (as expressed in Step 3 of Fig. 4) are
adopted to train the ISI model such that the healthy baseline
of ISI in TD of this particular maintenance cycle can be
generated; hence, baseline-ISI (ISIB ) is named in this paper.
As indicated in Fig. 2, the ISIB model transforms X into the
ISIB value, which is then considered in the FDC logic that
will be detailed in Section II-E.

C. DHI Module
In general, online SPC schemes are utilized to monitor the
quality status during manufacturing processes, and necessary
actions are performed if the process is out of statistical control
[18]. The DHI module applies the concept of online SPC to
convert |yT −yT | into DHI index where yT represents the
mean of fresh modeling samples of yT . The fresh yT samples
for modeling are collected just after maintenance (as expressed
in Step 3 of Fig. 4) to construct the DHI module such that the
healthy baseline of DHI in TD of this particular maintenance
cycle can be generated.
Fig. 5. Configurations of SPC control charts of DHI and BEI. (a) Converting
Fig. 5(a) shows the configuration of SPC control chart of yT into DHI. (b) Converting yE into BEI.
yT with the mean of the yT modeling samples (yT ) being the
baseline value and at the center. The Min yT , LSL, LCL, yT ,
UCL, USL, and Max yT are shown on the control chart of yT ; assigned to be 1, 0.7, 0.3, and 0, respectively. The formulas
furthermore, the DHI mapping values of the corresponding yT for converting yE into BEI are expressed as
mentioned above are normalized and assigned to be 0, 0.3, 0.7,  
yE
1, 0.7, 0.3, and 0, respectively. BEI = 1− × 0.3 , when 0 < yE ≤ Spec
The formulas for converting the upper half of yT into DHI Spec
 
can be derived as follows: yE − Spec
BEI = 0.7− × 0.4
  (HardSpec − ȳT ) − Spec
yT − ȳT (2)
DHI=1− ×0.3 , when ȳT < yT ≤ UCL when Spec < yE ≤ (HardSpec − ȳT )
UCL−ȳT
   
yT −UCL yE − (HardSpec − ȳT )
DHI=0.7− ×0.4 , when UCL < yT ≤ USL BEI = 0.3 − × 0.3
USL−UCL Max yE − (HardSpec − ȳT )
  when (HardSpec − ȳT ) < yE .
yT − USL
DHI=0.3− ×0.3 , when USL < yT .
Max yT −USL
(1)
By the same token, the formulas for converting the lower E. Illustration of the FDC Logic
half of yT into DHI can also be derived. In this paper, The right-hand part of the baseline FDC scheme shown in
the UCL/LCL and USL/LSL correspond to the process Spec Fig. 2 defines the FDC logic to convert DHI, BEI, and ISIB
(the border of sickness) and HardSpec (the border of death) values into the healthy status of TD. Before applying the FDC
of the TD, and the associated DHI values are 0.7 and 0.3, logic, the DHIT , BEIT , and ISIB− T should be assigned. Both
respectively. DHIT and BEIT are defined to be 0.7. Six times of standard
deviations of each individual process datum is assigned as the
threshold of the ISIB ; hence, ISIB− T = 6. The FDC logic is
D. BEI Module elaborated as follows.
The purpose of the BEI module is to transform the dif- If DHI > DHIT and all ISIB values are smaller than their
ference between yT and ŷB , i.e., yE (= |yT –ŷB |), into the corresponding ISIB− T , then a green light is shown. The green
BEI index. Fig. 5(b) depicts the configuration of SPC control light indicates that the TD is healthy and all of its related
chart of yE with zero (0) being the baseline value and at the parameters are normal.
bottom. The Spec, HardSpec− yT , and Max yE are shown on If DHI > DHIT and at least one ISIB is larger than its
the control chart of yE ; furthermore, the BEI mapping values corresponding ISIB− T , then a purple light is shown. This
of the corresponding yE mentioned above are normalized and implies that the TD is normal but the related process parameter
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 137

yE > yE− S or not, where yE = |yT –ŷB | and yE− S represents

the threshold of yE for detecting the sick state of the TD. To
avoid a false alarm, two consecutive detections of yE > yE− S
are required to make sure that the device is entering the sick
state. The predictive model will be activated to predict the
RUL once the device is in the sick state.
The conventional data-driven methods for predicting RUL
include simple projection models, such as exponential smooth-
ing, autoregressive model, and regression-based methods. One
major advantage of these techniques is the simplicity of their
calculations [19]. Considering limited amount of run-to-failure
data sets, a regression-based prediction method is a suitable
choice.
Failures of a large population of statistically identical and
independent items often exhibit a typical bathtub curve with
the following three phases: 1) early failures; 2) failures with
a constant (or nearly so) failure rate; and 3) wear-out failures
[20]. Generally, the sick state of a device occurs in the
wear-out phase. A failure in the wear-out phase results from
aging, wear-out, or fatigue, etc., whose failure rate increases
exponentially with time [20]. Therefore, a regression-based
ECF is adopted in this paper to implement the RUL predictive
model.
Fig. 6. Flowchart of baseline FDC execution procedure.
Define yE− B = y+3σyE where yE and σyE represent the mean
and standard deviation of yE of the healthy baseline samples
corresponding to the largest ISIB is abnormal and should be collected just after maintenance (as expressed in Step 3 of
checked. Fig. 4). Also, let yE− S and yE− D stand for the values of yE
If DHI < DHIT and BEI > BEIT , or if DHI < DHIT when the device enters the sick and dead states, respectively.
and at least one ISIB is larger than its corresponding ISIB− T , Referring to Fig. 5(b), we have yE− S = Spec and yE− D =
then a yellow light is shown. This yellow light means that the HardSpec−yT .
TD itself is healthy, while the reason for its OOC is due to The RUL predictive model is shown in Fig. 7 and the
abnormality of the related process parameter corresponding to flowchart for calculating RUL is depicted in Fig. 8, which
the largest ISIB that should then be checked. is explained as follows.
If DHI < DHIT and BEI < BEIT and all ISIB values are Phase I (Offline)
smaller than their corresponding ISIB− T , then a red light is Step 1) Calculate yE− B and find kB , which is the sample
shown. The red light reveals that the TD is abnormal and is number corresponding to yE− B . Then, define yE− D
caused by itself. This TD needs to be maintained. and yE− S .
Phase II (Online)
F. Flowchart of Baseline FDC Execution Procedure Step 2) If the condition of two consecutive detections of yE
The baseline FDC execution procedure is shown in Fig. 6 > yE− S is confirmed, then the TD is entering the sick
and is explained below. state and the flow jumps to Step 3.
Step 1) Collect a yT and its corresponding X of a new Step 3) Collect all the samples between yE− B and yE− S and
workpiece until the collection is successful. apply the ECF formula [21] to obtain the wear-out
Step 2) Add the yT and its corresponding X of the new failure equation
workpiece into the TD baseline modeling set and
ŷEi = A · eCki , i = B, B + 1, . . . , S (3)
retrain the TD baseline model, while the ISIB model
does not need recreation. with
Step 3) Calculate ŷB , BEI, DHI, and ISIB of this new work- 
S
piece. (ki − k̄)[ln(yEi ) − ln(yE )]
i=B
Step 4) Use the DHI, ISIB , and BEI to infer the healthy status C= (4)

S
of TD via the FDC logic. (ki − k̄)2
Step 5) Delete the yT and its corresponding X of the new i=B

workpiece from TD baseline modeling set. A = e(ln(yE )−Ck) (5)

where
G. RUL Predictive Module 
S
ki
The RUL predictive module includes a detection scheme i=B
k̄ =
and a predictive model. The detection scheme checks whether kS − k B + 1
138 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

Fig. 7. RUL predictive model.


S
K = [kB , kB+1 , . . . , kS ]T
ln(yEi )
i=B
ln(yE ) = and ŷEi − UBS+j is the predictive S+jth UB value,
kS − k B + 1
ŷEi − LBS+j is the predictive S + jth LB value, k̂D− RUL is
and ki is the ith sample number, ŷEi is the ith yE
the predictive sample number corresponding to yE− D ,
predictive value corresponding to ki , yEi is the ith
S+j is the 95% PI value corresponding to the S +
yE actual value corresponding to ki , kB is the Bth
jth sample, t α2 (kS − kB + 1 − p) is a t-distribution
sample number corresponding to yE− B , kS is the Sth
with kS –kB +1–p degrees of freedom with p being
sample number corresponding to yE− S .
the number of parameters and α = 0.05.
After obtaining the wear-out failure equation,
After obtaining all the 95% UB and LB values from
ŷEi = A · eCki , the UB equation, ŷEi − UB = AUB · eCUB ki ,
ki = kS+1 until k̂D− RUL , the UB equation
and LB equation, ŷEi − LB = ALB · eCLB ki , are also
derived as follows. ŷEi− UB = AUB · eCUB ki (9)
Step 4) Apply the failure equation, ŷEi = A · eCki , to predict
and the LB equation
the ŷEi values started from ki = kS+1 until the ŷEi
value is equal to or greater than the yE− D value; its ŷEi− LB = ALB · eCLB ki (10)
corresponding sample number is denoted as k̂D− RUL .
Then, utilize the two-sided 95% PI formula [22] to can then be derived by the same ECF formula [21]
calculate all the 95% UB and LB values from ki = and the approaches that were applied to obtain the
kS+1 until k̂D− RUL with failure equation: ŷEi = A · eCki .
Step 5) Apply the failure, UB, and LB equations (3), (9), and
ŷEi− UBS+j = ŷES+j − S+j , j = [1, 2, . . . , (D− RUL − S)] (10) to find RUL, K̂D− UB , and K̂D− LB that correspond
(6) to yE− D with
RUL = K̂D− RUL = k̂D− RUL − kS (11)
ŷEi− LBS+j = ŷES+j + S+j ,j = [1, 2, . . . , (D− RUL − S)] K̂D− UB = k̂D− UB − kS (12)
(7)
K̂D− LB = k̂D− LB − kS (13)

S+j = t α2 (kS − kB + 1 − p) sy2E × [1 + (kS+j )2 (KT K)−1 ] where k̂D− RUL is the D− RULth sample number cor-
(8) responding to yE− D , k̂D− UB is the D− UBth sample
where number corresponding to yE− D , and k̂D− LB is the
S D− LBth sample number corresponding to yE− D .
(ŷEi − ȳE )
syE = i=B
kS − k B + 1 III. Illustrative Examples
S
i=B yEi Three cases of a throttle valve in a vacuum module of a
ȳE =
kS − k B + 1 PECVD tool for solar-cell manufacturing at Motech Industries,
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 139

operation is activated to collect 10 fresh samples. Each sample

contains yT and its corresponding X.
The purpose of this example is to illustrate the necessity
of adopting the C&H samples in the baseline model. The
testing data of an entire PM period containing 390 samples are
examined in this illustration. As shown in Fig. 9(a), 30 C&H
samples and 10 fresh samples collected just after maintenance
are adopted as the modeling samples; while the testing case
shown in Fig. 9(b) adopts 40 fresh samples collected just
after maintenance for modeling without any C&H samples.
The angles of the TD (yT ), together with the baseline values
(ŷB ) and their related process data (NH3 , SiH4 , pressure, and
RF power) of samples 296–380 are displayed. Note that the
TD’s angles and their related process data of the testing runs
in Fig. 9(a) and (b) are exactly the same.
Comparing the baseline values (ŷB ) in the testing runs of
Fig. 9(a) and (b), apparently, the ones in Fig. 9(b) are relatively
rough, which are incorrect. This is because the modeling
samples of Fig. 9(a) contain 30 C&H samples while the ones in
Fig. 9(b) do not. As shown in the left portion of Fig. 9(a) and
(b), the peak-to-peak variations of all the related TD process
data of Fig. 9(a) are much larger than those of Fig. 9(b). As
a result, the representativeness of the model-operating space
of the case Fig. 9(b) is rather weak. Hence, the reason that
the C&H samples are required for modeling is clear.

B. Illustration of the FDC Portion of the BPM Scheme

The purpose of this example is to demonstrate the FDC
functions of the BPM scheme. The 450 testing samples used
in the example illustrated in Section I are reutilized for easy
comparison. Again, the procedure shown in Fig. 4 is followed
to collect the important samples needed for creating the TD
baseline model. Therefore, 30 C&H samples plus 10 fresh
samples collected just after maintenance are adopted as the
baseline modeling samples. Each sample contains yT and its
corresponding X. The 10 fresh samples will be adopted for
Fig. 8. Flowchart for calculating RUL. constructing the ISIB model and DHI module as well as the
yE− B value.
The Min yT , LSL, LCL, yT , UCL, USL, and Max yT of
Inc., Tainan, Taiwan, are adopted as the illustrative examples. this example for converting yT to DHI are 0, 5, 22, 27, 32, 50,
As such, the throttle valve is the TD in these examples. Based and 90, respectively. Also, the Spec, HardSpec−yT , and Max
on the domain knowledge, the angle of the TD may be affected yE for transforming yE to BEI are 5, 23, and 63, respectively.
by the following related parameters: NH3 , SiH4 , pressure, and The maximal allowable deviation of each related process
RF power. Three illustrations are included in this section: datum is 10% of each individual nominal value. Therefore,
1) necessity of including the C&H samples in the baseline by assigning 10% of each individual nominal value to be
model; 2) FDC portion; and 3) RUL portion. equivalent to 6 of the ISIB value, the value of ISIB− T is equal
to 6 [23]. Moreover, as mentioned in Section II-E, both DHIT
A. Illustration of the Necessity of Adopting the C&H Samples and BEIT are assigned to be 0.7, which is the threshold for
in the Baseline Model entering the sick state.
The procedure shown in Fig. 4 is followed to collect The results of running the FDC portion of the BPM scheme
the important samples needed for creating the TD baseline are shown in Fig. 10. Besides yT , the healthy baseline of the
model. In this example, the angle of the TD (throttle valve) TD (ŷB ) is also displayed. Furthermore, DHI, BEI, and ISIB
is designated as yT while its related process data (X) contains values of all the related parameters, NH3 , SiH4 , pressure, and
NH3 , SiH4 , pressure, and RF power. Since the number of the RF power, are also shown below the yT and ŷB results.
related process data is four, about 40 samples are required Observing the sample at yellow-circle 1 in Fig. 10, its
for building the baseline model. Therefore, the KIS scheme is DHI < DHIT and the ISIB (=120) of NH3 is larger than
performed offline on 2602 healthy historical samples to select its corresponding ISIB− T (=6); therefore, a yellow light is on
about 30 C&H ones. Then, just after maintenance, the online according to the FDC logic shown in Fig. 2. This yellow light
140 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

Fig. 9. Illustration of the necessity of adopting the C&H modeling samples. (a) BPM results with C&H modeling samples. (b) BPM results without C&H
modeling samples.
means that the TD itself is healthy, while the reason for its larger than ISIB− T , thus a purple light is shown. This implies
OOC is due to the glitch that occurs in the related NH3 , which that the TD is normal but the tube pressure is abnormal and
should be checked. should be checked.
All the samples within red-circles 2 and 4 satisfy the
conditions that DHI < DHIT and BEI < BEIT as well C. Illustration of the RUL Portion of the BPM Scheme
as all ISIB values of these samples are smaller than their An entire PM period containing 390 samples utilized in
corresponding ISIB− T , hence red lights are displayed. A red Section III-A are reexamined in this example. Thirty C&H
light indicates that the TD is abnormal and caused by itself. samples are collected and included in the modeling samples.
The sample in purple-circle 3 meets the conditions of DHI The first 10 fresh samples are utilized for constructing the
> DHIT as well as the ISIB value (=13.5) of pressure are ISIB model and DHI module as well as the yE− B value. The
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 141

Fig. 10. Results of the FDC portion of the BPM scheme.

Fig. 11. BPM related data and indexes of an entire PM period.

following 380 samples with the BPM related data and indexes Phase I (Offline)
are shown in Fig. 11. As depicted in Fig. 11, the angles of the Step 1) Calculate yE− B (= ȳE + 3σyE ) = 0.28 and find kS =
TD (yT ) increase gradually due to the aging effect, while the 13. Then, define yE− D (=HardSpec−yT ) = 50−37.74
baseline values (ŷB ) are relatively flat due to the fact that their = 12.26, and yE− S (=Spec) = 5.
related process data, NH3 , SiH4 , pressure, and RF power, are Phase II (Online)
stable with little variations. The conditions of DHI < DHIT
Step 2) The condition of two consecutive detections of yE
and BEI < BEIT with all ISIB values being smaller than their
> yE− S is confirmed at kS = 213.
corresponding ISIB− T occur at around sample 213; therefore,
Step 3) Collect all the samples between kB = 13 and kS = 213
sample 213 is the entry point of the sick state. This event
and apply (3)–(5), the failure equation is obtained
activates the RUL predictive process. All the RUL predictive
as
results are shown in Fig. 12 and are explained as follows.
According to the flowchart shown in Fig. 8, we have the ŷEi = 0.5753 × e0.00112ki . (14)
following.
142 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

Fig. 12. RUL predictive results.

Step 4) Apply (14) to predict the ŷEi values started from ki IV. Conclusion
= kS+1 = 214 until the ŷEi value is equal to or greater In this paper, a VM-based BPM scheme was proposed. The
than the yE− D value (12.26); this sample number is BPM scheme possesses FDC and PdM capabilities. The BPM
denoted as k̂D− RUL , which is 273. Then, calculate all scheme is composed of TD baseline model, ISIB model, DHI
the 95% UB and LB values from 214 until 273 with module, BEI module, FDC logic, and RUL-predictive module.
(6)–(8). Finally, the UB equation is derived as The TD baseline model predicted the TD’s baseline (ŷB ) by
ŷEi− UB = 0.5221 × e0.0112ki (15) collecting the TD-related process data (X), while the TD data
(yT ) was applied for retraining the TD baseline model. The
and the LB equation is ISIB model transformed X into the ISIB value. Also, the
ŷEi− LB = 0.6585 × e0.00112ki . (16) DHI module converted |yT − yT | into DHI index, where yT
represents the mean of fresh modeling samples of yT , and the
Step 5) With yE− D = 12.26 and (14)–(16) being applied BEI module converts yE (= |yT − ŷB |) into the BEI index. The
to find k̂D− RUL , k̂D− UB , and k̂D− LB ; they are 273, DHI, BEI, and ISIB values were sent to the FDC logic to infer
281, and 261, respectively. Then, utilize (11)–(13) to the healthy status of TD. Finally, the RUL-predictive module
calculate RUL, K̂D− UB , and K̂D− LB with kS = 213 that considers the residual of the sensed TD data (yT ) and
as follows: the outputs of the TD baseline (ŷB ) was applied to predict the
RUL. The illustrative examples revealed that the BPM scheme
RUL = K̂D− RUL = k̂D− RUL − kS = 273 − 213 = 60
can perform fault detection, achieve fault diagnosis, and also
K̂D− UB = k̂D− UB − kS = 281 − 213 = 68 accomplish fault prognosis.
K̂D− LB = k̂D− LB − kS = 261 − 213 = 48.

Acknowledgment
In fact, the actual sample number of the TD entering the The authors would like to thank Motech Industries, Inc.,
dead state is 289. As a result, the actual RUL is 289−213 = 76 Tainan, Taiwan for providing the raw data for the PECVD
sample periods. The sample period of this example is about process used as the illustrative example.
an hour.
The RUL predictive module of the BPM scheme has also
been applied to four different throttle valves. Each valve References
belongs to different PECVD tool. The predictive RULs are [1] ISMI Predictive and Preventive Maintenance Equipment Implementa-
29, 17, 19, and 60, while their corresponding actual RULs tion Guidelines, Technology Transfer #08064934A-ENG, International
are 32, 19, 21, and 68, respectively. The test results of the SEMATECH Manufacturing Initiative, Aug. 30, 2008.
[2] Predictive Preventive Maintenance Impact on Factory Cost and Cycle
RUL predictive module show that the predictive accuracy is Time, International SEMATECH Manufacturing Initiative, Dec. 31,
reasonably fine. 2009.
HSIEH et al.: VM-BASED BASELINE PREDICTIVE MAINTENANCE SCHEME 143

[3] ISMI Consensus Preventive and Predictive Maintenance Vision Guide- Fan-Tien Cheng (S’87–M’89–SM’98–F’08) re-
line: Version 1.1, Technology Transfer #06114819C-ENG, International ceived the B.S. degree from National Cheng Kung
SEMATECH Manufacturing Initiative, Nov. 20, 2007. University (NCKU), Taiwan, in 1976, and the Mas-
[4] J. Hollister and P. McGuire, Research on the Current Status of Predictive ters degree in 1982 and the Ph.D. degree in 1989
Maintenance (PdM) Algorithms and Applications, ISMI Predictive and from Ohio State University, Columbus, all in elec-
Preventive Maintenance Initiative, Alan Weber and Associates, Inc., Feb. trical engineering.
2009. He is currently the Chair Professor of NCKU.
[5] ISMI Predictive Preventive Maintenance Data Requirements, Interna- From August 1998 to July 2001, he was the Di-
tional SEMATECH Manufacturing Initiative, Sep. 30, 2009. rector of the Institute of Manufacturing Engineering
[6] G. Vachtsevanos, F. L. Lewis, M. Roemer, A. Hess, and B. Wu, (IME), NCKU. He built a Web-enabled experimental
Intelligent Fault Diagnosis and Prognosis for Engineering Systems. manufacturing execution system and a supply chain
Hoboken, NJ: Wiley, 2006. information system for integrated circuit packaging. He also established test
[7] W. A. Shewhart, Economic Control of Quality of Manufactured Product. beds for e-diagnostics, equipment engineering systems, engineering chain
New York: Van Nostrand, 1931. management systems, and automatic virtual metrology frameworks for semi-
[8] D. J. Inman, C. R. Farrar, V. Steffan, and V. Lopes, Damage Prognosis: conductor manufacturing at IME Automation Laboratory for educational and
For Aerospace, Civil and Mechanical Systems. New York: Wiley, 2005. research purposes. He founded the e-Manufacturing Research Center (eMRC)
[9] B. Sun, P. B. Luh, Z. O’Neill, and F. Song, “Building energy doctors: at NCKU in January 2008, and since then has been the Director of eMRC. His
SPC and Kalman filter-based fault detection,” in Proc. IEEE Conf. current research interests include semiconductor manufacturing automation,
Automat. Sci. Eng., Aug. 2011, pp. 333–340. e-manufacturing, virtual metrology, and intelligent predictive maintenance.
[10] C. Letellier, G. Hoblos, and H. Chafouk, “Robust fault detection based Prof. Cheng was a recipient of the Senior Scientist Award from the DoD, Tai-
on multimodel and interval approach: Application to a throttle valve,” wan, in 1994, the Kayamori Best Automation Paper Award at the IEEE ICRA
in Proc. 19th IEEE Mediterranean Conf. Contr. Automat. (MED), Jun. in 1999, the Outstanding Industry-University-Cooperation (IUC) Award from
2011, pp. 479–484. the MoE, Taiwan, in 2003, the NCKU Distinguished IUC-Professor Award in
[11] Y.-S. Hsieh, F.-T. Cheng, and H.-C. Yang, “Virtual-metrology-based 2004 and 2008, the ROC National Science Council (NSC) Outstanding IUC
FDC scheme,” in Proc. IEEE Conf. Automat. Sci. Eng., Aug. 2012, pp. Award (as the only awardee) in 2006, the ROC NSC Outstanding Research
80–85. Award in 2006 and 2009, the University Industry Economy Contribution
[12] J. Liu, D. Djurdjanovic, J. Ni, N. Casoetto, and J. Lee, “Similarity Award, Individual Award from the Ministry of Economic Affairs, Taiwan,
based method for manufacturing process performance prediction and in 2008, the TECO Award from the TECO Technology Foundation, Taiwan,
diagnosis,” Comput. Indust., vol. 58, pp. 558–566, Jan. 2007. in 2010, the 2011 National (Silver) Invention and Creation Award from
[13] H.-E. Kim, A. C. C. Tan, J. Mathew, and B.-K. Choi, “Bearing fault MOEA, Taiwan, the 2011 Award for Outstanding Contributions in Science
prognosis based on health state probability estimation,” Expert Syst. and Technology from the Executive Yuan, Taiwan, and the 2012 National
Applicat., vol. 39, pp. 5200–5213, Apr. 2012. (Gold) Invention and Creation Award from MOEA, Taiwan. From 2000 to
[14] F.-T. Cheng, H.-C. Huang, and C.-A. Kao, “Developing an automatic 2004, he was an Associate Editor of the IEEE Transactions on Robotics
virtual metrology system,” IEEE Trans. Automat. Sci. Eng., vol. 9, no. and Automation. He was the IEEE ICRA Kayamori Best Automation
1, pp. 181–188, Jan. 2012. Paper Award Committee Chair in 2006. He currently is the Senior Program
[15] F.-T. Cheng, C.-L. Wang, H.-C. Yang, S.-L. Wu, and C.-Y. Lo, “A Committee Member of ICRA 2011 and the Program Chair of IEEE WCICA
study on application cluster service scheme and computer performance 2011. From 2007 to 2009, he was the Convener and the Program Director of
evaluator,” J. Chinese Inst. Eng., vol. 31, no. 4, pp. 675–690, Jun. 2008. the NSC Automation Engineering Program, Taiwan.
[16] W.-M. Wu, F.-T. Chen, and F.-W. Kong, “A dynamic-moving-window
scheme for virtual-metrology model refreshing,” IEEE Trans. Semicond.
Manuf., vol. 25, no. 5, pp. 238–246, May 2012.
[17] F.-T. Cheng, Y.-T. Chen, Y.-C. Su, and D.-L. Zeng, “Evaluating reliance
level of a virtual metrology system,” IEEE Trans. Semicond. Manuf.,
vol. 21, no. 1, pp. 92–103, Feb. 2008.
[18] M. P. Groover, Automation, Production Systems, and Computer- Hsien-Cheng Huang was born in Tao-Yuan, Tai-
Integrated Manufacturing, 2nd ed. London, U.K.: Prentice-Hall, 2001. wan, on April 25, 1979. He received the B.S. de-
[19] X.-S. Si, W. Wang, C.-H. Hu, and D.-H. Zhou, “Remaining useful life gree from the Department of Industrial Engineering
estimation: A review on the statistical data driven approaches,” Eur. J. and Technology Management, Da-Yeh University,
Oper. Res., vol. 213, pp. 1–14, Aug. 2011. Chang-Hua, Taiwan, in 2004, and the M.S. and
[20] E. E. Lewis, Introduction to Reliability Engineering, 2nd ed. New York: Ph.D. degrees from the Institute of Manufacturing
Wiley, 1996. Information and Systems, National Cheng Kung
[21] C. C. Steven and C. P. Raymond, Numerical Methods for Engineers: University (NCKU), Tainan, Taiwan, in 2007 and
With Software and Programming Applications, 4th ed. New York: 2011, respectively.
McGraw-Hill, 2002. Since 2010, he has been the Deputy Director
[22] S. D. Gryze, I. Langhans, and M. Vandebroek, “Using the correct of the e-Manufacturing Research Center, NCKU.
intervals for prediction: A tutorial on tolerance intervals for ordinary His current research interests include automatic virtual metrology, factory
least-squares regression,” Chemometrics Intell. Lab. Syst., vol. 87, pp. automation for semiconductor manufacturing, equipment engineering systems,
147–154, Jun. 2007. e-manufacturing, embedded systems, fault detection and classification, and
[23] R. Meli and L. Santillo, “Function point estimation methods: A com- short-range wireless technology.
parative overview,” in Proc. FESMA Conf., Oct. 1999, pp. 271–286.

Yao-Sheng Hsieh (S’12) was born in Peng-Hu, Tai-

wan, on May 27, 1983. He received the B.S. degree
in 2005, and the M.S. degree in 2007, all from Chung-Ren Wang was born in Tainan, Taiwan, on
the Department of Industrial Management and the June 6, 1988. He received the B.S. degree from Na-
Institute of Industrial Engineering and Management, tional Chung Cheng University, Chiayi, Taiwan,
National Formosa University, Yun-Lin, Taiwan. He in 2010, and the M.S. degree from the Institute of
is currently pursuing the Doctoral degree with the Manufacturing Engineering, National Cheng Kung
Institute of Manufacturing Information and Systems, University, Tainan, Taiwan, in 2012.
National Cheng Kung University, Tainan, Taiwan. His current research interests include virtual
His current research interests include automatic metrology and predictive maintenance.
virtual metrology, factory automation for semicon-
ductor manufacturing, and predictive maintenance.
144 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 26, NO. 1, FEBRUARY 2013

Saint-Chi Wang was born in Kaohsiung, Taiwan, Haw-Ching Yang (M’04) was born in Tainan,
on November 30, 1988. He received the B.S. de- Taiwan, in 1966. He received the B.S. degree in
gree from the Department of Computer Science engineering science from National Cheng Kung Uni-
and Information Engineering, National Cheng Kung versity (NCKU), Tainan, in 1989, the M.S. degree
University (NCKU), Tainan, Taiwan, in 2011. He in control engineering from National Chiao Tung
is currently pursuing the Masters degree with the University, Hsinchu, Taiwan, in 1991, and the Ph.D.
Institute of Manufacturing Information and Systems, degree from the Institute of Manufacturing Engineer-
NCKU. ing, NCKU, in 2003.
His current research interests include baseline pre- From 1997 to 2000, he was a Lecturer with the
dictive maintenance. Department of Automation Engineering, Dayeh Uni-
versity, Changhua, Taiwan. Since September 2004,
he has been an Assistant Professor with the Institute of Systems and Control
Engineering, National Kaohsiung First University of Science and Technology,
Kaohsiung, Taiwan. He is experienced in the analysis, modeling, design, and
implementation of manufacturing automation. His current research interests
include production prediction, system optimization, virtual production control,
intelligent systems, and cloud-based applications.