5464 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 8, NO.

9, SEPTEMBER 2023

An Intelligent Factory Automation System With

Multivariate Time Series Algorithm for
Chip Probing Process
Yu-Ming Hsieh , Member, IEEE, Chin-Yi Lin , Member, IEEE, Jan Wilch , Graduate Student Member, IEEE,
Birgit Vogel-Heuser , Fellow, IEEE, Yu-Chen Lin, Yu-Chuan Lin , Member, IEEE,
Min-Hsiung Hung , Senior Member, IEEE, and Fan-Tien Cheng , Life Fellow, IEEE

Abstract—Chip-probing is the key process for IC manufacturing Index Terms—Intelligent factory automation (iFA), multivariate
to its ensure quality. As the number of tests increases, the test version of time series prediction (TSPMVA ), remaining useful life
quality and the test yield will be affected because the needles on the (RUL), chip probing.
probe card of the tester will be contaminated by external objects or
worn out. Whether a needle polish of the probe card is required can
be determined through real-time monitoring on various detection
I. INTRODUCTION
indicators such as resistivity and yield. However, both resistivity NTEGRATED Circuit (IC) refers to the fabrication of tran-
and yield are lagging indicators, and excessively frequent needle
polishes will increase the processing time and reduce the test
throughput. The so-called Intelligent Factory Automation (iFA)
I sistors, resistors and/or capacitors in a single chip. IC has
the characteristics of slim, high reliability, high density, and
low cost; therefore, it has been widely used in various domains
system platform, realized by integrating several intelligent services
including Intelligent Predictive Maintenance (IPM), was proposed
of computers, home appliances, industry, communication, and
to accomplish the goal of Zero-Defect Manufacturing. However, the military, etc. There are five major steps in the IC manufacturing
current remaining useful life (RUL) prediction algorithm in IPM process: circuit design, mask fabrication, wafer fabrication, chip
is a univariate time series prediction. The RUL prediction may testing, and IC packaging. The test process is crucial for IC
not be accurate enough if only one variable is adopted to describe manufacturing to ensure the final quality. Since the IC manu-
the dynamic changes of the time series. A supervisory architecture facturing can never reach the yield of 100%, it is necessary to
for chip probing process based on iFA is proposed in this letter. test before putting an IC on the system to ensure normal and
The Multivariate Version of Time Series Prediction (TSPMVA ) complete IC functions. However, the process of chip probing is
in this architecture can use the vector autoregression model to highly dependent on the reliability of probe card that contains
improve the accuracy of RUL prediction. Experimental results plural needles. Regarding the probing reliability, Liu et al. [1]
reveal that the proposed supervisory framework with TSPMVA
can not only monitor the tester’s health status efficiently but also
suggested an Antenna-on-Chip mechanism that can be aligned
improve the accuracy of needle’s RUL prediction by extracting with the test pads. Cheng et al. [2] proposed a compact fluo-
multiple features. rescent probe based on fluorescence intensity ratio technique.
Experimental results show that the proposed probe offers a
novel and reliable method for real-time chip thermal monitoring.
Manuscript received 9 March 2023; accepted 4 July 2023. Date of publi-
cation 13 July 2023; date of current version 24 July 2023. This letter was
Kim et al. [3] proposed a novel and massively parallel testing
recommended for publication by Associate Editor L. Moench and Editor J. Yi method using multi-root. Observing the experimental results,
upon evaluation of the reviewers’ comments. This work was supported in part the method can not only reduce test costs up to 90% but also
by the “Intelligent Manufacturing Research Center” (iMRC) through Featured achieve nearly 94.84% of test reliability without affecting the
Areas Research Center Program within the framework of the Higher Education yield. The recommended number of contacts for needles of
Sprout Project by the Ministry of Education (MOE) in Taiwan and in part by the
National Science and Technology Council of Taiwan, under Contracts MOST different materials are determined via rigorous experiments in
110-2923-E-006-010-MY3, MOST 111-2218-E-006-012, MOST 111-2221-E- [4]. However, with the dynamic changes of the test status, if the
006-200, NSTC 111-2222-E-006-021, NSTC 112-2221-E-006-115-MY2, and polish timing is only defined by a fixed value, unexpected down
NSTC 110-2221-E-034-009-MY2. (Corresponding authors: Yu-Ming Hsieh; or overly early polish might occur. On top of that, the references
Fan-Tien Cheng.)
Yu-Ming Hsieh is with the Academy of Innovative Semiconductor and
mentioned above do not predict the RUL of each individual
Sustainable Manufacturing, National Cheng Kung University, Tainan 701401, needle.
Taiwan (e-mail: [email protected]). With the number of tests increases continuously, the needles
Chin-Yi Lin, Yu-Chen Lin, Yu-Chuan Lin, and Fan-Tien Cheng are with on the probe card of the tester will be stained by foreign objects
the Institute of Manufacturing Information and Systems, National Cheng Kung (e.g., debris) or worn out that affect the test quality and lead to
University, Tainan 701401, Taiwan (e-mail: [email protected]; yuchen@
imrc.ncku.edu.tw; [email protected]; [email protected]). poor test yield. In general, there are two modes of probe polish:
Jan Wilch and Birgit Vogel-Heuser are with the Institute of Automation Auto-Polish and Smart Polish. Auto-Polish sets a certain number
and Information Systems, Technical University of Munich, 80333 Garching, of tests to be performed after a period of time and determines
Germany (e-mail: [email protected]; [email protected]). the timing for regular maintenance based on experience; while
Min-Hsiung Hung is with the Department of Computer Science and Infor-
mation Engineering, Chinese Culture University, Taipei 11114, Taiwan (e-mail:
Smart Polish monitors several features (e.g., resistivity [5])
[email protected]). in real time, and examines the yield through test results to
Digital Object Identifier 10.1109/LRA.2023.3295237 decide whether a needle polish of the probe card is required.

2377-3766 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
HSIEH et al.: INTELLIGENT FACTORY AUTOMATION SYSTEM WITH MULTIVARIATE TIME SERIES ALGORITHM 5465

requires more data and computing resources than univariate time

series does. In one word, the more data, the better prediction
accuracy.
Sanayha and Vateekul [8] used the convolutional neural net-
works to predict the RUL of turbofan engines in the power plants.
Compared to the univariate time series prediction, the prediction
accuracy was improved by 30% after adding in multivariate
variables. Jing et al. [9] proposed an algorithm based on the long
short-term memory (LSTM) recurrent neural network (RNN) to
Fig. 1. Timeline of chip probing operation scenario.
predict the lifetime of UV LEDs. The results showed that the
algorithm they proposed has a 29.7% lower lifetime prediction
However, both resistivity and yield are lagging indicators; when error compared with the nonlinear least squares.
abnormalities are found via these features, some test losses might Liu et al. [10] built a large vector autoregression (VAR)
have taken place already. Moreover, too many Smart Polishes model to forecast three important weather variables for 61 cities
will increase cycle time and reduce test output; while too few around the United States. The experimental results show that
Smart Polishes may affect the yield test results. the proposed method had a good forecasting performance and
The current chip probing operating scenario is illustrated in more than double the skill scores of other forecasting models
Fig. 1. Observing Fig. 1, the operation of Wafer 11 begins at in the literature. Although the VAR model [11] can predict the
t11 . The loading time (TL ) for loading Wafer 11 on the Tester future trend based on historical multivariate time series data,
is 60 seconds, and the test-operation time (TT ) of a single piece the best combination that VAR(p) requires for model creation is
of wafer is 1250 seconds. When the wafer tests are completed, unclear.
a large amount of log files recording all processing information As mentioned above, there are two major requirements on
(e.g., yield, resistivity, probe test sequence, and wafer/chip map, probe card health status management in the chip probing pro-
etc.) are generated. It takes 600 seconds for the data collection cess:
and analysis functional modules to analyze the data and suggest Requirement 1: Provide probe card maintenance map that
whether to perform Smart Polish; the required data-collection- includes the health status of each needle.
and-analysis time is denoted as TD . If any abnormality occurs at Requirement 2: Real-time monitor and estimate the probe card
this stage (Wafer 11), the decision of performing Smart Polish RUL as well as predict proper needle polish timing.
can only be made after TL +TT + TD . To accomplish the above requirements, this letter develops
As for Wafer 12 in Fig. 1, the loading operation will com- a supervisory architecture with TSPMVA , which deals with the
mence at t12 , which is the timing just after TT of Wafer 11. issues of predictor selection of VAR through Bayesian Infor-
When the test result of Wafer 11 exceeds the threshold set by mation Criterion (BIC) [12] and a series of hypothesis tests as
the factory, this abnormal result requires immediate Smart Polish well as provides probe card maintenance map to help engineers
to avoid excessive waste. Yet because TL (60 seconds) is shorter execute Smart Polish effectively.
than TD (600 seconds), Wafer 12 is already on the Tester for The remainder of this letter is organized as follows.
test operation when smart-polish alert of Wafer 11 is issued. In Section II introduces the chip probing process; Section III briefs
other words, Smart Polish can only be performed before loading the ABPM scheme and time series analysis; Section IV proposes
Wafer 13 onto the Tester, which leads to the loss of Wafer 11 and the developed supervisory architecture; Section V presents the
Wafer 12. Therefore, how to predict the RUL of each needle on illustrative examples. Finally, summary and conclusions are
the probe card and provide an early warning at the right timing stated in Section VI.
is essential. To be exact, if a proper RUL prediction module
is implemented and able to foresee that the Tester will cause
II. CHIP PROBING PROCESS
an abnormality before performing Wafer 11 tests, then Smart
Polish can be performed prior to Wafer 11 processing. As such, A Tester is an inspection device used for electrical tests on
the loss of Wafers 11 & 12 can be waived. chips in the process of IC development and manufacturing.
The novel Intelligent Factory Automation (iFA) system plat- Tester conducts electrical needle tests by transmitting the
form [6], as shown in Fig. 2, is realized by integrating several needle test signals from the needle test machine to each chip
core technologies such as the Advanced Manufacturing Cloud under test on the wafer through the needles on the probe card.
of Things (AMCoT) framework, Cyber-Physical Agent (CPA), Then the signals returning from each chip under test are collected
Intelligent Predictive Maintenance (IPM), Automatic Virtual for determining whether the chips on the wafer are normal. As
Metrology (AVM), and Intelligent Yield Management (IYM), such, the health condition of the probe needles will affect the
etc. As such, the iFA system can provide a total-solution package correctness of the wafer tests. Four-point probe measurement
for implementing intelligent manufacturing. To improve the [5] is a common method for measuring the resistance value of
efficiency of production tools and their availability, tool health semiconductor materials. It consists of four equidistant tungsten
status monitoring and RUL prediction should be conducted metal tips with a limited radius, and each tip is supported by a
through IPM in the iFA system. The IPM system adopts the spring at the other end to minimize sample damage during prob-
Advanced Baseline Predictive Maintenance Scheme (ABPM) ing. Four metal tips move up and down during the measurements,
[7] as the main algorithm. a high impedance current is supplied to the two outer probes.
A target device usually has more than one feature that reflect The relative voltage between the two inner probes is measured
its aging condition. Therefore, in order to improve the accuracy to calculate the sample resistivity. Therefore, the resistivity
of the RUL prediction of the target device, considering multivari- of the wafer can be calculated just by acquiring the distance
ate variables simultaneously has become inevitable. Multivariate between the probes and applying a fixed current. The smaller
time series prediction is closer to the actual situation, yet it the contact resistance of the probe, the better. According to this
Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
5466 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 8, NO. 9, SEPTEMBER 2023

Fig. 2. The iFA system for chip probing process.

measurement method, a probe needle has 4 sets of resistivity for model, with the theoretical model being the AR model. The AR
the data adopted in this work (i.e., DC_CREDQ0, DC_ CREDQ and VAR models are illustrated below.
1, DC_CREDQ 2, and DC_CREDQ 3). 1) Autoregressive (AR) Model: The AR model adopts
the same variable yt and uses the previous spans of yt
III. ADVANCED BASELINE PREDICTIVE MAINTENANCE (yt−1 , yt−2 , . . .) to predict the performance of yt , assuming
SCHEME AND TIME SERIES ANALYSIS that they are linearly dependent. AR(p) represents the auto-
regression of order p, and p is the number of lagging spans
A. ABPM Scheme required by the AR model, as shown in (1) and (2) [7]:
 p
Lin et al. proposed the Advanced Baseline Predictive Main-
tenance (ABPM) [7] Scheme as shown in the upper right corner ŷt = ϕi yt−i (1)
of Fig. 2. The ABPM Scheme is comprised of the Data Quality i=1
 p

Module, Algorithms Pool Module for data preprocessing, De- 
vice Health Index (DHI) Module, and RUL Predictive Module. εt = yt − ŷt = yt − ϕi yt−i (2)
The core of RUL Prediction in the ABPM Scheme is the i=1
Time-Series Prediction (TSP) Scheme, which includes the TSP where
Modeling Module, RUL Estimation Module, Pre-alarm Module, yt : actual aging feature at time t.
and Death Correlation Index (DCI) Module [7]. ŷt : predicted aging feature at time t.
The core of RUL estimation is the TSP algorithm based on ϕi : least square estimation coefficient of AR model, i = 1,
the autoregressive integrated moving average (ARIMA) model 2, …, p.
[13]. TSP applies the information criterion to create a time series yt−i : actual aging feature at time t-i.
analysis model that predicts the RUL of a TD for avoiding εt : white noise error term at time t.
unscheduled down. As such, the goal of nearly zero down time
of the TD can be achieved and the maintenance cost can then be
2) Vector Autoregression (VAR) Model: The VAR model is
reduced.
of multivariate time series. The core concept is that each variable
The resistivity of probe can represent its health status, and
depends not only on its past values, but also on other variables.
there are four sets of resistivity in a probe. However, the algo-
Proposed by Sims [11] in 1980, this was used for empirical
rithm adopted in the original TSP [7] is an ARIMA model, which
research on macroeconomics. As more data collected, the situa-
is of a univariate time series prediction. If there are correlation
tional requirements for multivariate time series increase; which
between those four sets of resistivity, the RUL prediction may
makes it impossible for the original AR model to take all vari-
not be accurate enough.
ables into account. Therefore, the VAR model that expresses the
multivariate time series relationship in matrix form is extended.
B. Time Series Analysis The VAR model is widely used in the research and analysis of
Time series analysis [14] is a statistical method of dynamic multivariate time series.
data processing based on the stochastic process theory and statis- The VAR model changes the original AR model ŷt =
p
tical methods. It uses specific statistical indicators to represent i=1 i yt−i into a matrix representing multivariate autoregres-
ϕ
the value of a certain physical phenomenon on different time sive coefficients. If there are k time series {Yi,t }, i = 1, . . . , k,
spans to solve the problem of practical prediction. Define the the vector can be expressed as (3) cited from [11]:
aging features of the TD in different time spans as: yT = { …,  p

yt-2 , yt-1 , yt , yt+1 , yt+2 , …}, time series analysis can be applied Ŷt = Ai Yt−i + εt (3)
to find out the statistical relationship of TD aging features among i=1
different time spans for RUL prediction. There are mainly three where
basic univariate models of the TSP models, which are the autore- Ŷt : predicted aging feature matrix at time t, t = 1, 2, …,T.
gressive (AR) model [15], the moving average (MA) model, and A0 : constant term of the VAR model.
the mixed model composed of autoregressive moving average Ai : least square estimation coefficient matrix of the VAR
(ARMA) model based on AR and MA. The major model used in model, i = 1,2, …, p.
this study is the VAR model [11] in the multivariate time series Yt−i : actual aging feature matrix at time t − i, i = 1,2, …, p.
Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
HSIEH et al.: INTELLIGENT FACTORY AUTOMATION SYSTEM WITH MULTIVARIATE TIME SERIES ALGORITHM 5467

εt : white noise error matrix at time t.

Convert VAR(p) that contains two time series into a matrix

form as (4) cited from [13]:
p    
A11,i A12,i ε
ŷt = A0 + Y + 1,t (4)
A21,i A22,i t−i ε2,t
i=1
3) Granger Causality Test: The Granger causality test was
demonstrated by Granger in 1969 [15]. It is a statistical method
for hypothesis testing to verify whether one set of time series
is the cause of another set of time series. In the 1980s, many
economists modified the Granger causality test by adding more
than two time series and expanded it into a VAR model [16] Fig. 3. Flowchart of the chip probing process for detecting abnormal needles.
to generate more effective estimation results. Therefore, if a
Granger causality exists between multiple time series, adding
the past information of one time series into another time series
can increase the explanatory capacity of variables and enhance amount of tests on each needle of the probe card varies. The test
the estimation results. results of all the needles should be precisely acquired to generate
First, assume that aging feature 1 (yAF1 ) and aging feature the probe card maintenance map. Table I(a) shows the test results
2 (yAF2 ) are generalized stationary sequences. To examine the of Wafer_01. Observing the red square, the test results of Chip 2
Null Hypothesis that yAF1 is not the Granger causality of yAF2 , and Chip 3 indicate “Pass”, and their positions are x coordinate
create an AR model by adopting the lagging spans of yAF1 , and (XC) 1, y coordinate (YC) 23; and XC 1, YC 24, respectively.
derive (5) from (1) [15]: Four sets of resistivity data are also collected in Table I(a).
AF 1
ŷt = ϕ1 yt−1 AF 1
+ ϕ2 yt−2 AF 1
+ · · · + ϕp yt−p + εt (5) Table I(b) shows the probing map of Wafer_01, where Chip
2 and Chip 3 with XC 1 YC 23 and XC 1 YC 24, respectively,
where ϕi was defined in (1) and p represents the earliest time in are tested by needle no. 286. Table I(c) depicts the probe card
the lagging spans of the yAF1 variable tested to be significant. hardware map that presents the x and y hardware coordinates
Then, expand the auto-regression model by adopting the of each needle on the probe card. Observing Table I(a), (b),
lagging spans of yAF2 as in (6) [15]: and (c), both Chip 2 and Chip 3 are tested by needle no. 286,
AF 1 AF 1 AF 1
ŷt = ϕ1 yt−1 + ϕ2 yt−2 + · · · + ϕq yt−p while the x and y coordinates of needle no. 286 are XHC 1
AF 2 AF 2
and YHC 24, respectively; also, the test results are “Pass” and
+ bq yt−q + · · · + br yt−r their corresponding resistivity values are acquired. As such, via
matching the co-relations among Table I(a), (b), and (c), the test
+ εt (6)
results and resistivity of all the needles on the probe card for
where ϕi was defined in (1) and b represents the least square each wafer during the test process can be acquired.
estimation coefficient of the AR model of yAF2 variable, and Data Denoising: The main purpose of predictive maintenance
q represents the earliest time in the lagging spans of the yAF2 is to predict the RUL of a target device (TD). To prevent too
variable tested to be significant; while r represents the latest time much noise from affecting the correctness of RUL prediction,
in the lagging spans of the yAF2 variable. If no lagging spans of data denoising is necessary. A cumulative average (CMA) [17]
yAF2 are left in the model, the Null Hypothesis of yAF1 not being can indicate the average of a series of resistance values up to a
the Granger causality of yAF2 is verified. certain point. CMA is less susceptible to extreme values and it
can observe the overall average trend, hence it is adopted for data
IV. THE SUPERVISORY ARCHITECTURE WITH TSPMVA denoising in this letter. The formula of accumulative average is
In order to meet the monitoring requirements of the chip used to reduce the noise as in (7):
j
1 AFjk
probing process, this study proposes a monitoring process based
on the iFA system [6] to avoid losses caused by delayed Smart CM Ajk = (7)
j
Polish. The flowchart of the chip probing process of the super-
visory architecture is shown in Fig. 3 and explained as follows. where
AFjk : kth resistivity value at Chip j.
CMAjk : accumulative mean of the kth resistivity value at Chip
A. Step A. Data Preprocessing
j.
Various intelligent services can be implemented as pluggable with k = 1, 2, 3, 4.
application modules and plugged into CPA for edge computing j = 1, 2, …, m.
[6]; thus, as depicted in the lower left corner of Fig. 2, after
completing the tests, data-preprocessing module as well as the B. Step B. Abnormal Needle Detection and Probe Card
proposed abnormal detection scheme can be plugged into the
Maintenance Map Preparation
CPA connected to the designated Tester for data preprocessing
and abnormal needle detection. The purpose of IPM is to monitor the health status of a TD so
Data preprocessing is executed right after data collection. The that it can predict TD’s RUL to prevent unscheduled down [6].
two major functions of data preprocessing, namely Probe map & Therefore, the first step is to find out the aging feature that can
test results preparing and Data denoising, are illustrated below. represent the TD’s health status. After that, the failure history
Probe map and test results preparing: As explained above, the of the aging feature data are collected and the thresholds of
probe card must perform multiple tests on each wafer, while the “Sick” and “Dead” are defined. Once the TD enters the Sick
Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
5468 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 8, NO. 9, SEPTEMBER 2023

TABLE I
AN EXAMPLE OF TEST RESULTS

state, the RUL prediction will be executed to achieve the goal

of preventing unscheduled down. However, since it is difficult FUNCTION CheckAbnormalWafer( Xik , threshold, n).
to define the thresholds of “Sick” and “Dead”, detecting ab- exceedingCount ← CountExceedingChips( Xik ,
normal needles becomes complicated. To resolve this issue, the threshold).
following approach is proposed for detecting abnormal needles abnormalPercentage← 0.01 ∗ n.
automatically. And they are depicted on the right side of Fig. 3 IF exceedingCount > abnormalPercentage THEN.
and explained below. RETURN TRUE.
1) Step B-1: Define Resistivity Threshold of a Wafer: The ELSE.
percentile rank (PR) [18] is utilized in this study to check the RETURN FALSE.
resistivity distribution of each wafer. First, define Xik as the kth END IF.
resistivity value of all chips of the ith wafer, and the number of END FUNCTION.
chips of each wafer as n. Then sort Xik from small to large. Let
P Rα_ik be the α percentile value of the kth resistivity of the ith
wafer, and the pseudocode that defines the resistivity threshold
of a wafer is set as follows in italics:

FUNCTION DefineResistivityThreshold(Xik , alpha).

Sort Xik in ascending order.
n ← length( Xik ).
P Rα_ik ← alpha / 100 ∗ (n + 1).
threshold ← waferData[ P Rα_ik ].
RETURN threshold.
END FUNCTION.
Set alpha as 99.

Fig. 4. Flow chart of the TSPMVA algorithm.

2) Step B-2: Count the Number of Chips That Exceed the
Threshold in a Wafer: Define the resistivity threshold with the
following pseudocode, and calculate the amount of chips that
exceed the threshold within each wafer. C. Step C. Instant Needle Monitoring and Smart Polish
Requirement Prediction
When abnormal needles are detected on the probe card main-
FUNCTION CountExceedingChips( Xik , threshold).
tenance map, IPM can then be used to decide the proper time for
count ← 0.
Smart Polish. Since there are four sets of resistivity in one needle,
FOR EACH resistivityValue IN Xik .
the influence among different features is possibly ignored when
IF resistivityValue > threshold THEN.
adopting the original TSP algorithm [7] as it can only monitor
count <- count + 1.
one sort of features. To tackle this problem, a new architecture
END IF.
based on the original TSP algorithm [7] is proposed in this study
END FOR.
as shown in the bottom right corner of Fig. 2. The univariate TSP
RETURN count.
algorithm is substituted by the multivariate TSPMVA algorithm
END FUNCTION.
in the TSP modeling module. The flow chart of the multivariate
TSPMVA algorithm is shown in Fig. 4. There are four main
3) Step B-3: Check Abnormal Wafer: Then, check if one procedures of TSPMVA as explained below.
wafer is abnormal with the following pseudocode: 1) Step i. Check Time Series Assumptions: Once an abnormal
This step checks whether the number of chips whose resis- needle is detected, the method proposed in [5] can be used to
tivity values exceeding the threshold is greater than 1% of n. If define the modeling data set (YM ) and start establishing the
yes, it is judged as an abnormal wafer. TSPMVA model. Before utilizing TSPMVA for RUL prediction,
4) Step B-4: Provide Probe Card Maintenance Map: After the data should be made sure to fulfill two conditional assump-
finding out the abnormal wafer, the probe card maintenance map tions: 1) mean functions are constants that will not change
can be generated according to the relationship between the probe over time; and 2) covariance function (cov(yt , yt−k ) = γk )
card and the wafer test results to help figure out the abnormal is time-irrelevant, which is called “weak stationarity”.
needles. Also, the RULs of abnormal needles on the probe card 2) Step ii. Perform Granger Causality Test: After confirm-
can be predicted via TSPMVA as illustrated in Step C. ing that the modeling data (YM ) fulfill the two conditional
Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
HSIEH et al.: INTELLIGENT FACTORY AUTOMATION SYSTEM WITH MULTIVARIATE TIME SERIES ALGORITHM 5469

assumptions, perform the Granger causality test on the modeling each wafer has 1824 chips and one chip has 91 test items. A
samples to test if there are correlations between the variables or probe card contains 480 needles in total. As the number of chips
not. If no, go back to the original TSP algorithm. per wafer is greater than the number of needles, each wafer needs
3) Step iii. Conduct Optimal Model Selection: Optimal to be tested up to 4 times. The proportion of a total 480 needles in
model selection can be conducted once the modeling data (YM ) each wafer test is as follows: 3.33% for 1 time, 3.13% for 2 times,
pass the Granger causality test. The operation flow is as follows. 90.41% for 3 times, and 3.13% for 4 times. Per statistically, most
Step iii-1: Calculate the autocorrelation function (ACF) and needles are tested 3 times per wafer.
the partial autocorrelation function (PACF), and mark the most According to the literature [5] and the rule of thumb of the
yt-1 -related ACF time as A in (8) and PACF time as B in (9). chip-probing factory, the decision of probe cleaning is made
A = argmax (ρk ) (8) based on four sets of resistivity: DC_CREDQ0, DC_CREDQ1,
DC_CREDQ2, and DC_CREDQ3. Hence, in this experiment,
B = argmax (ρρk ) . (9) the resistivity value is taken as the aging feature. Fig. 5 shows
Step iii-2: Execute the Ljung-Box test on YM . Ljung-Box the test results of needle no. 279. It can be found in the red square
test is applied to check whether any set of autocorrelation is 0 that the DC_CREDQ3 value oscillates significantly and enlarges
in time series. When H0 is not rejected, YM is white noise. If gradually before regular clean in Fig. 5(a). After discussing with
YM is white noise, go to Step iii-3; else jump to Step iii-4. Null engineers, DC_CREDQ3 is chosen as the main feature of this
hypothesis and alternative hypothesis are calculated as: experiment.
H0 : ρ1 = ρ2 = · · · = ρM = 0 (10)
A. Data Processing
Ha : ρi = 0 where i = 1, 2, . . . , M. (11)
Step iii-3: Add more samples into YM . If YM is white noise, After CPA completes data collection, data pre-processing is
the modeling data are not representative; to enhance the explana- carried out immediately. The probe map and test results prepar-
tory of the model, more modeling samples should be added in ing as well as data denoising are then executed respectively
this step. according to the methods mentioned in Section IV-A. The test
Step iii-4: Create a model with all VAR(p) combinations. For data of needle no. 279 are taken for illustration as displayed in
time series, AR generally checks the PACF values. Take the AR Fig. 5. Needle no. 279 is tested 3 times in this wafer, so there
lag with the highest PACF value as the largest lagging span [17]. are 600 (200∗3) test samples. In total, seven (7) Auto Polishes
Create multiple VAR(p) models, p = 1, 2, . . . , B. according for regular clean and two (2) Smart Polishes occur during this
to the previous VAR formula. There will be p kinds of com- period. The resistivity raw data of each test of needle no. 279 are
binations, namely, VAR(1), VAR(2), …, VAR(p) combination shown in Fig. 5(a). Conduct denoising on four resistivity values
models. Then calculate the Bayesian information criterion (BIC) by applying (10), and re-calculate once Auto Polish or Smart
[12] of each VAR (p) model combination and pick the VAR Polish occurs. Judging from the blue rectangle on the results
model combination with the lowest BIC as the best model. shown in Fig. 5(b), the resistivity value after denosing becomes
4) Step iv. Determine the Final Model: Although BIC can smoother and it possesses an aging tendency, which can avoid the
be used to select the best model through the above processes, it risk of inaccurate RUL prediction due to large data fluctuations.
is necessary to remove the insignificant predictors in the model
through statistical methods. When the estimated coefficient is B. Abnormal Needle Detection and Probe Card Maintenance
greater than the 95% standard error confidence interval, it is an Map Preparation
insignificant predictor that is not crucial for the interpretation Perform abnormal needle detection after completing data
of future spans, hence it should be deleted. The formula of preprocessing, and set α equal to 99 in this example. The results
comparing with the 95% standard error confidence interval is of Section IV-B Step B-1 and Step B-2 are shown in Fig. 5(c) and
shown as (12). (d). As depicted in the blue rectangle in Fig. 5, as it gets closer to
|ϕi | > 1.96 × s.e. (ϕi ) , i = 1, 2, . . . , p (12) the second Smart Polish, the number of chips whose resistivity
where measured by needle no. 279 exceeding P R99_ik becomes more;
s.e.(ϕi ): Standard errors of unstandardized VAR regression which means that the health status of needle no. 279 is probably
coefficients. getting poor and demands immediate clean. Define that 1% of
Re-estimate the model whose predictors have been removed. chips (18 chips) in one wafer exceeding P R99_ik as abnormal.
Then perform the Ljung-Box test on the model residual of the re- The probe card maintenance map can be acquired by further
estimated model. When the model predictors are explanatory, the examining the total 480 needles with the data between the
residual should be the white noise. As such, the null hypothesis interval of Auto Polish {7} and Smart Polish {II}, as displayed in
and alternative hypothesis of the Ljung-Box test are calculated the top portion of Fig. 3. Observing the probe card maintenance
as (13) and (14): map, needles no. 259, 279, 280, and 285 with red highlights are
H0 : ρ1 = ρ2 = · · · = ρn = 0 (13) detected as abnormal. The probe card maintenance map can not
only pre-alert the users, but also determine the proper timing for
Ha : ρj = 0, where j = 1, 2, . . . , n. (14) Smart Polish.
Finally, this model can be determined as the most suitable for
model creation. C. Instant Needle Monitoring and Smart Polish Requirement
Prediction
V. ILLUSTRATIVE EXAMPLES
This study proposes TSPMVA to predict the RUL of abnor-
An example of actual test data from a chip probing factory is mal needles, and assist users in judging when Smart Polish
adopted for demonstration. There are 200 wafers to be tested; is required. Needle no. 279 will be used as an example and

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
5470 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 8, NO. 9, SEPTEMBER 2023

Fig. 5. Test results of needle no. 279.

Fig. 6. TSPMVA prediction results of needle no. 279.

elaborated in the following paragraphs. And LSTM is adopted on (4), the matrix representing the VAR formula is as below:
for comparison here.    
Observing Sample 65 in Fig. 6, 1% of the chips exceeding −0.0014 −0.4857X1 X1 ,1 0.8939X1 X2 ,1
ŷt = +
the threshold (14) set by Section IV-B occurs; then it enters the −0.0179 −0.4161X2 X1 ,1 −0.6432X2 X2 ,1
“Sick” state (the resistivity value corresponding to the Sick state  
0.1974X1 X1 ,2 0.9835X1 X2 ,2
is 58) to start the RUL prediction on needle no. 279. After dis- + + ···
cussing with engineers, the “Dead” spec of resistivity is defined −0.2963X2 X1 ,2 −0.2725X2 X2 ,2
as 68. Then, apply the procedures in Fig. 4 to build the RULMVA  
0.2097X1 X1 ,7 0.2843X1 X2 ,7
prediction model. After obtaining the modeling samples (YM ), + + εt , t = 1, . . . , 7
0.2092X2 X1 ,7 0.25595X2 X2 ,7
perform the Granger causality test on these modeling samples as
depicted in Step ii of Fig. 4. Set M = 30 [19], and let t indicate The results of TSP, TSPMVA and LSTM are shown in Fig. 6.
the time of Sick state; YM = {Yt−30 , Yt−29 , . . . , Yt−2 , Yt−1 }, Since DC_CREDQ3 is chosen as the main feature, LSTM
and Yt−i indicates the resistivity matrix at time t − i and i = 1, takes DC_CREDQ0, DC_CREDQ1, and DC_CREDQ2 as the
2, …, 30. variables to conduct RUL prediction on DC_CREDQ3. And the
The results show that causality relationship exists between modeling samples are shown in the red square of Fig. 5. It can be
DC_CREDQ3 and DC_CREDQ2, thus these two parameters observed from Fig. 6(a) that as the aging feature trend begins to
are used for model creation. Next, after Step iii and Step iv, the rise, the number of chips exceeding the spec starts to go beyond
time series model-VAR(7) is found to have the lowest BIC and 18 (1%) and the aging features then start to enter the Sick state.
is hence a more suitable combination for model creation. Based In the aging feature prediction in Fig. 6(b), the 102nd wafer

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.
HSIEH et al.: INTELLIGENT FACTORY AUTOMATION SYSTEM WITH MULTIVARIATE TIME SERIES ALGORITHM 5471

TABLE II prediction for chip probing process based on iFA is proposed

RUL PREDICTION RESULTS in this letter. The experimental results reveal that the proposed
architecture can provide a maintenance map of the probe card,
predict the needles’ RUL in a probe card, and suggest proper
timing for needle Smart Polish. However, Smart Polish is not
only related to the health condition of the needles but also to the
product yield and the wafer map defect pattern of the probe card.
Therefore, a process that takes product yield and probe card’s
wafer map defect pattern monitoring into consideration should
be developed in the future.

ACKNOWLEDGMENT
The authors would like to thank Nanya Technology Corpora-
test value exceeds the “Dead” spec (68). Fig. 6(c) depicts that tion in Taiwan for providing the raw data used in the illustrative
after DHI becoming less than 0.7, needle no. 279 enters the
examples.
“Sick” state and RUL prediction is activated. In Fig. 6(d), the
RUL prediction of the original TSP [7] shows a red light at the
93rd probing test; hence, TSP suggests that Smart Polish may REFERENCES
be performed 9 more time spans before it enters the “Dead” [1] Q. Liu, A. C. F. Reniers, U. Johannsen, M. C. van Beurden, and A. B. Smol-
state (time span 102); since the average number of tests for each ders, “Improved probing reliability in antenna-on-chip measurements,”
needle is 3, a pre-alarm of (9/3 =) 3 wafers ahead of the death IEEE Antennas Wireless Propag. Lett., vol. 17, no. 9, pp. 1745–1749,
Sep. 2018.
may be issued. On the other hand, Fig. 6(d) shows that with [2] T. Cheng et al., “A compact fluorescent probe for real-time thermal
TSPMVA , a red light is on at the 89th probing test. Therefore, monitoring of chips,” J. Lightw. Technol., vol. 40, no. 18, pp. 6288–6295,
TSPMVA suggests that Smart Polish can be carried out (102-89 Sep. 2022.
=) 13 time spans, namely four wafers, before entering the [3] H. Kim, Y. Lee, and S. Kang, “A novel massively parallel testing method
“Dead” state. It can also be seen from Fig. 6(d) that LSTM using multi-root for high reliability,” IEEE Trans. Rel., vol. 64, no. 1,
pp. 486–496, Mar. 2015.
shows blue light alert at time span 66, which is (102-66 =) 36 [4] D.-S. Liu, M.-K. Shih, and F.-M. Zheng, “An investigation of wafer
time spans, i.e., (36/3 =) 12 wafers prior to the actual failure. probe needles mechanical properties and contact resistance changing under
The experiment is carried out under the hardware specifi- multiprobing process,” IEEE Trans. Compon. Packag. Technol., vol. 31,
cations of CPU: Intel(R) Core (TM) i7-10700 @ 2.90GHz, no. 1, pp. 196–203, Mar. 2008.
[5] D. R. Zrudsky, H. D. Bush, and J. R. Fassett, “Four point sheet resistivity
RAM: 128 GB, and GPU: NVIDIA GeForce RTX 2080 Ti techniques,” Rev. Sci. Instrum., vol. 37, no. 7, pp. 885–890, Jul. 1966.
GAMING OC 11 G. TSPMVA takes 7 seconds and 2 seconds, [6] F.-T. Cheng, Industry 4.1: Intelligent Manufacturing With Zero Defects,
respectively, for 1-span modeling and prediction time on Needle 1st ed. Hoboken, NJ, USA, Wiley, 2021.
no. 279. As for LSTM, it takes 90 seconds to complete model [7] C.-Y. Lin, Y.-M. Hsieh, F.-T. Cheng, H.-C. Huang, and M. Adnan, “Time
creation and 2 seconds for predicting 1 time span. In other words, series prediction algorithm for intelligent predictive maintenance,” IEEE
Robot. Automat. Lett., vol. 4, no. 3, pp. 2807–2814, Jul. 2019.
the time required for creating TSPMVA is much shorter than [8] M. Sanayha and P. Vateekul, “Remaining useful life prediction us-
LSTM. ing enhanced convolutional neural network on multivariate time series
The RUL prediction result comparison of the four needles is sensor data,” Walailak J. Sci. Technol., vol. 16, no. 9, pp. 669–679,
shown in Table II. It can be observed that TSPMVA indeed allows Feb. 2019.
[9] Z. Jing, J. Liu, M. S. Ibrahim, J. Fan, X. Fan, and G. Zhang, “Lifetime
users with more lead time for Smart Polish. In short, if there is prediction of ultraviolet light-emitting diodes using a long short-term
a causality relationship between the features, TSPMVA is more memory recurrent neural network,” IEEE Electron Device Lett., vol. 41,
effective compared to the original TSP. no. 12, pp. 1817–1820, Dec. 2020.
The original TSP [7] uses only one aging feature to describe [10] Y. Liu, M.-C. Roberts, and R. Sioshansi, “A vector autoregression weather
the dynamic variation of time series without considering the model for electricity supply and demand modeling,” J. Modern Power Syst.
Clean Energy, vol. 6, no. 4, pp. 763–776, Jul. 2018.
effect of other variables. This causes late alarms for all predic- [11] C. A. Sims, “Macroeconomics and reality,” Econometrica, vol. 48, no. 1,
tions of the needles and results in wafer loss. Based on the above pp. 1–48, Jan. 1980.
experimental results, it can be concluded that if there is a Granger [12] R. Shibata, “Selection of the order of an autoregressive model by Akaike’s
causality relationship between the features, its inclusion can help information criterion,” Biometrika, vol. 63, no. 1, pp. 117–126, Apr. 1976.
[13] G. E. P. Box and D. A. Pierce, “Distribution of residual autocorrelations
TSPMVA to better describe the dynamic behavior of the data and in autoregressive-integrated moving average time series models,” J. Amer.
provide enhanced prediction results. On the other hand, overly Stat. Assoc., vol. 65, no. 332, pp. 1509–1526, 1970.
early warnings of LSTM can lead to the increase of maintenance [14] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis:
cost, which is not preferable. The reason for the poor prediction Forecasting and Control. Hoboken, NJ, USA: Wiley, 1994.
performance is that LSTM needs more failure records for model [15] C. W. J. Granger, “Investigating causal relations by econometric models
and cross-spectral methods,” J. Econometrica, vol. 37, no. 3, pp. 424–438,
creation, yet most predictive maintenance problems may not Aug. 1969.
have a large number of failure records to improve prediction [16] C. W. J. Granger, “Testing for causality: A personal viewpoint,” J. Econ.
accuracy. As shown in this case, only one complete failure record Dyn. Control, vol. 2, pp. 329–352, Sep. 1980.
can be used for model creation. [17] R. G. Brown, Smoothing, Forecasting and Prediction of Discrete Time
Series. Englewood Cliffs, NJ, USA: Prentice Hall, 1963.
[18] S. W. Huck, Reading Statistics and Research, 5th ed. Boston, MA, USA:
VI. SUMMARY AND CONCLUSION Allyn & Bacon, 2007.
[19] R. V. Hogg and E. A. Tanis, Probability and Statistical Inference. New
A supervisory architecture with TSPMVA that utilizes the VAR York, NY, USA: Macmillan, 1977.
model to improve the accuracy of the original TSP [7] RUL

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on September 13,2023 at 16:41:59 UTC from IEEE Xplore. Restrictions apply.