.. . Expert Sysrems will1 Applications 41 (2014) 7579-7595 Bivariate quality control using two-stage intelligent monitoring scheme CrossMark - Ibrahirn Masood a.'!', Adnan Hassan 'Faculty of Mechanical and Manufacturing Engineering, Universiti Tun Hussein Onn Malaysia, 86400 Pant Raja, Batu Pahat, Johor, Malaysia b ~ a c u l t y ofMechanical Engineering, Universiti Teltnologi Malaysia, 81310 UTM Sltudai, Johor, Malaysia A R T 1 C L E I N F O A B S T R A C T Article history: Available online 5 June 2014 -- Keywords: Balanced monitoring Bivariate quality control Statistical features Synergistic artificial neural network Two-stage monitoring In manufacturing industries, it is well known that process variation is a major source of poor quality products. As such, monitoring and diagnosis of variation is essential towards continuous quality improve- ment. This becomes more challenging when involving two correlated variables (bivariate), whereby selection of statistical process control (SPC) scheme becomes more critical. Nevertl~eless, the existing tra- ditional SPC schemes for bivariate quality control (BQC) were mainly designed for rapid detection of unnatural variation with limited capability in avoiding false alarm, that is, imbalanced monitoring per- formance. Another issue is the difficulty in identibing the source of unnatural variation, that is, lack of diagnosis, especially when dealing with small shifts. In this research, a scheme to address balanced mon- itoring and accurate diagnosis was investigated. Design consideration involved extensive simulation experiments to select input representation based on raw data and statistical features, artificial neural net- work recognizer design based on synergistic model, and monitoring-diagnosis approach based on two- stage technique. The study focused on bivariate process for cross correlation function, p = 0.1-0.9 and mean shifts, p = k0.75-3.00 standard deviations. The proposed two-stage intelligent monitoring scheme (2s-IMS) gave superior performance, namely, average run length, ARLl= 3.18-16.75 (for out-of-control process), ARLO = 335.01-543.93 (for in-control process) and recognition accuracy. RA = 89.5-98.5%. This scheme was validated in manufacturing of audio video device component. This research has provided a new perspective in realizing balanced monitoring and accurate diagnosis in BQC. 2014 Elsevier Ltd. All rights reserved. 1. Introduction In manufacturing industries, when quality feature of a product involves two correlated variables (bivariate), a n appropriate SPC charting scheme is necessary to monitor and diagnose these variables jointly. Specifically, process monitoring refers t o t h e iden- tification of process condition either in a statistically in-control or out-of-control, whereas process diagnosis refers t o t h e identifica- tion of t h e source variable(s) for out-of-control condition. In addressing this issue, the traditional SPC charting schemes for BQC such as x2 (Hotelling, 7947), multivariate cumulative sum (MCUSUM) (Crosier, 1988), and multivariate exponentially weighted moving average (MEWMA) (Lowly. Woodall, Champ, & Rigdon, 1992; Prabhu 8 Rungel-, 1997) are known t o be effective in monitoring aspect. Unfortunately, they a r e merely unable to provide diagnosis information, which is greatly useful for a quality practi- tioner in finding t h e root cause error and solution for corrective action. Since then, major researches have been focused on diagnosis * Corresponding author. Tel.: +607 4537700. E-mail addresses: ibraliim@~~Lli~n.edu.~ny (I. Masood), atlnan@fl~m.ul~ii.~ny (A. Hassan). URLs: l~~~p://www.~~~I~~n~edu~my (I. Masood), http://www.utm.my (A. Hassan). aspect. Shewhart-based control charts with Bonferroni-type control limits (Alt, 1985), principle component analysis (PCA) (Jacltson, 1991), multivariate profile charts (Fuchs & Ben,jamini, 1994), r2 decomposition (Mason, Tracy, & Young, 1995) and Minimax control chart (Sepulveda & Nachlas. 1997), among other, have been investi- gated for such purpose. Further discussions on this issue can be found in Lowry and Montgomery (1995), I 200). Nevertheless, the existing traditional SPC schemes were mainly designed by focusing on rapid detection of out-of- control condition (ARL, * I ) but it has limited capability in avoid- ing false alarm (ARL < 200). Fig. 1 illustrates the concepts of imbalanced monitoring vs. balanced monitoring as the central theme for this investigation. Based o n diagnosis viewpoint, an effective bivariate SPC scheme should be able to identify the source variable(s) of out-of-control condition a s accurate as possible. Nevertheless, it is difficult to cor- rectly recognize when dealing with small shifts ( ~ 1 . 0 standard deviation). Chih and Rollier (1994). Chih and Rollier (1995), Zorriassatine, Tannock, and O'Brien (2003), Chen and Wang (2004) and Yu and Xi (2009). for examples, have reported less than 80% accuracy for diagnosing mean shifts at 1.0 standard deviation. Among others, only Guh (2007) and Yu et al. (2009) reported the satisfied results ( >90% accuracy). The imbalanced monitoring and lack of diagnosis capability as mentioned above need further investigation. In order to minimize erroneous decision making in BQC, it is essential to enhance the overall performance towards achieving balanced monitoring (rap- idly detect process variationlmean shifts with small false alarm as shown in Fig. 1) and accurate diagnosis (accurately identify the sources ofvariation/mean shifts). Additionally, the BQCapplications are still relevant in today's manufacturing industries. In solving this issue, a two-stage intelligent monitoring scheme (2s-IMS) was designed to deal with dynamic correlated data streams of bivariate process. This paper is organized as follows. Section 2 describes a modeling of bivariate process data streams and patterns. Section 3 presents the frameworlc and procedures of the 2s-IMS. Section 4 discusses the performance of the proposed scheme in comparison to the traditional SPC. Section 5 finally outlines some conclusions. 2. Modeling of bivariate process data streams and patterns A large amount of bivariate samples is required for evaluating the performance of the 2s-IMS. Ideally, such samples should be tapped from real world. Unfortunately, they are not economically available or too limited. As such, there is a need for modeling of synthetic samples based on Lehman (1977) mathematical model. Further discussion on data generator can be found in Masood and Hassan (201 3). In bivariate process, two variables are being monitored jointly. Let Xl-i=(Xl.I,. . .,XI-24) and X2-i=(X2.1.. . . . X Z . ~ ~ ) represent 24 observation samples for process variable 1 and process variable 2 respectively. Observation window for both variables start with sam- ples i = (1,. . . -24). It is dynamically followed by (i + I ) , (i + 2) and so on. When a process is in the state of statistically in-control, samples from both variables can be assumed as identically and indepen- dently distributed (i.i.d.) with zero mean ( p o = 0) and unity standard deviation (go = 1). Depending on process situation, the bivariate samples can be in low correlation ( p = 0.1 -0.3), moderate correla- tion ( p = 0.4-0.6) or high correlation ( p = 0.7-0.9). Data correlation ( p ) shows a measure of degree of linear relationship between the twovariables. Generally, this data relationship is difficult to be iden- tified using Shewhart control chart as shown in Fig. 2. On the other hand, it can be clearly indicated using scatter diagram. Low corre- lated samples yield a circular pattern (circular distributed scatter plot), moderate correlated samples yield a perfect ellipse pattern, whereas high correlated samples yield a slim ellipse pattern. Disturbance from assignable causes on the component variables (variable-1 only, variable-2 only, or both variables) is a major source of process variation. This occurrence could be identified by various causable patterns such as mean shifts (sudden shifts), trends, cyclic, systematic or mixture. In this research, investigation was focused on sudden shifts patterns (upward and downshift shifts) with positive correlation ( p > 0). Seven possible categories of bivariate patterns were considered in representing the bivariate process variation in mean shifts as follows: N (0,O): both variables XI-i and X2-i remain in-control. US (1,O): shifted upwards, while X2.i remains in-control. US (0.1): X2.i shifted upwards, while remains in-control. US ( 1 , l ) : both variables Xl.i and X2.i shifted upwards. DS (1,O): shifted downwards, while X2.i remains in-control. DS ( 0 , l ) : X2.i shifted downwards, while remains in-control. DS ( 1 , l ) : both variables X1.i and X2.i shifted downwards. Reference bivariate shift patterns based on mean shifts f3.00 standard deviations are summarized in Fig. 3. Their structures are unique to indicate the changes in process mean shifts and data correlation. The degree of mean shifts can be identified when the center position shifted away from zero point (0,O). 3. Two-stage intelligent monitoring scheme As noted in Section I , an integrated MSPC-ANN was combined in a single-stage monitoring scheme (direct monitoring-diagnosis) as proposed in Chen and Wang(2004). Niaki and Abbasi (2005). and YLI et al. (2009). The other schemes based on fully ANN-based models as proposed in Zorriassatine, Tannocli, and O'Bricn (2003), C u h (2007). Yuand Xi (2009) and El-Midany et al. (2010) also can be classified as a single-stage monitoring scheme. In this research, two-stage mon- itoring scheme was investigated by integrating the powerful of MEWMA control chart and Synergistic-ANN model for improving the monitoring-diagnosis performance. Framework and pseudo- code (algorithm) for the proposed scheme are summarized in Figs. 4 and 5 respectively. It should be noted that an initial setting as fol- lows needs to be performed before it can be put into application: Load the trained raw data-ANN recognizer into the system. I. Masood, A. Hassan/Expert Systems with Applications 41 (2014) 7579-7595 Y I S e n s ~ t ~ v ~ t v in mean s h ~ R detection Capablhtv In false alann avo~dance I/ Shorter ARLl represents fastel Longer ARLO repiesenis smaller 1 detect~on of plocess mean sh~fts false alaim I t b U i 1 Current state I Imbalanced n~onitoring: able to detect process mean shifts rapidly (ARL, =1 1) but has limited capability to avoid false alarm (ARLO 5 200) I Desired state (for this research) 1 I / Balanced monltonn. [reasonable fol cu~rent ~ r a c t l c e i able to detect process mean 1 s h ~ f t s rapldly (ARL, Z, 1) and nlalnta~n small false alarm (ARLO >> 200) % I I I Ideal state i 1 Perfect balanced: able to detect process mean shifts as soon as possible (ARLI = 1) 1 I . . 1 w~tllout tnggenng any false alarm (ARLO = m) 1 L ~ ~ Fig. 1. Current state and desired state towards balanced monitoring. Set t h e values of means (p01,po2) and standard deviations (nol.ao2) of bivariate in-control process (for variables and X2.i). These parameters can be obtained based on historical or preliminary samples. Perform in-process quality control inspection until 24 observa- tion samples (individual or subgroup) to begin the system. Recognition window size is set to 24 observation samples (for variables Xl-i and X2_i) since it provided sufficient training results and statistically acceptable to represent normal distribution. Preli- minary experiments suggested that a smaller window size (<24) gave lower training result due to insufficient pattern properties, while a larger window size (>24) does not increase the training result but burden the ANN training. Rational to integrate the MEWMA control chart and the Synergistic-ANN model are based on preliminary experiments. Generally, the MEWMA control chart is ltnown to be effective for detecting bivariate process mean shifts more rapidly compared to the x2 control chart. Furthermore, it is very sensitive when deal- ing with small shifts (G1.00 standard deviations). Unfortunately, based on one point out-of-control detection technique, it gave lim- ited capability to avoid false alarm (ARb 6 200). This becomes more critical when the variables are highly correlated. In the related study, pattern recognition scheme using a Synergistic- ANN model gave better capability in avoiding false alarm (ARL,, > 200). As such, it can be concluded that process identifica- tion based on recognition of process data stream patterns (Synergistic-ANN model) is more effective compared to detection of one point out-of-control (MEWMA control chart). Nevertheless, different techniques should have their respective advantages in terms of pointlpattern discrimination properties. In order to fur- ther improve the monitoring performance (ARLl =. l , ARLO >> 200), it is useful to combine both discrimination properties (MEWMA control chart and Synergistic-ANN recognizer) by approaching I. Masood, R HassanlExpert Systems with Applications 41 (2014) 7579-7595 Shewhart Control Chart Scatter Diagram -- Low Correlation Low correlation 3 1 2 2 1 1 I 1 Y -2 I ' I' zo 41 m KI 100 1 '0 20 40 60 $0 Inn Nvrnbor orsamplcs -3 -2 -1 0 1 2 Number uisamplcs XI Moderate Correction Moderatc correlation 3 I 2 ..' ZO 40 60 SO 100 3~ 20 40 60 RO l W - 4 -3 -2 -1 0 1 2 3 Nlmbcl oisdnlplcs Numher urnampler X 1 High Correlation High correlation 3 3 I I I I 2 2 3~ 20 40 60 80 100 ' 0 20 10 60 80 IMI .4 -3 -2 -1 0 1 2 3 Nunlbel ofratnplcs Number ulsamples x1 Fig. 2. Shewhart control charts and its respective scatter diagrams. two-stage monitoring and diagnosis. In the first stage monitoring, the MEWMA control chart is used for triggering bivariate process mean shifts based on 'one point out-of-control' as per usual. Once the shift is triggered, the Synergistic-ANN recognizer will perform second stage monitoring and diagnosis through recognition of pro- cess data stream patterns that contain one of several out-of-control points. This approach is suited for 'recognition only when neces- sary' concept, that is, it is unnecessary to perform recognition while the process lies within a statistically in-control state. Besides, recognition is only necessary for identifying patterns sus- pected to a statistically out-of-control state. Besides producing smaller false alarm, this approach will also reduce computational efforts and time consumes for pattern recognition operation. 3.1. M E W M A control chart The MEWMA control chart developed by Lowry ct al. ( 1992) is a logical extension of the univariate EWMA control chart. In the bivariate case, the MEWMA statistics can be defined as follows: [U:(EWMAI, - 1 1 ) ' + u j ( E W M A z i - p 2 j 2 - 2 u : , ( E W M A l ; - p , ) ( E W M A z i - @ , j ] n M E W M A , = (u:.: - u:,, Covariance matrix of MEWMA: h M E W M A The standardized samples (Zli, Zzi) with cross correlation func- tion ( p ) were used. Thus, a1 = a2 = 1 ; 012 = p. Notations L and i rep- resent the constant parameter and the number of samples. The starting value of EWMA (EWMAo) was set as zero to represent the process target (/A,,). The MEWMA statistic samples will be out-of-control if it exceeded the control limit (H). In this research, three sets of design parameters (A, H: 0.05, 7.35; 0.10, 8.64; 0.20, 9.65) as reported in Prabhu and Runger (1997) were investigated. 3.2. Synergistic-ANN model pattern recognizer Synergistic-ANN model as shown in Fig. (5 was developed for ( 1 ) pattern recognizer. It is a parallel combination between two I. Masood, A. H a s s a n l E x p e r t S y s t e m s w i t h Applications 41 (2014) 7579-7595 . . . . . . .:. . -: - -5.0 n -5.0 -2.5 0.0 2.5 5.0 -5.0 -2.5 0.0 2.5 5.0 XI (Down-Shift) X 1 (Down-Shift) - Fig. 3. S u m m a r y o f bivariare s h i f t patterns for p = 0.1, 0.5 and 0.9 individual ANNs that are: ( i ) raw data-based ANN, and (ii) statisti- recognizers can be combined using simple summation: 01, = cal features-ANN as shown in Fig. 7. X(ORD-~,OF-~), where i = (1,. . . .7) are the number of outputs. Final Let OR^ = (ORD-i,. . . , ORD.7) and OF = (OF.l,. . . , OF.7) represent decision ( 0 ,,,,,,) was determined based on the maximum value seven outputs from raw data-based ANN and statistical features- from the c o m b i n e d ~ ~ u t ~ u t s : ANN recognizers respectively. Outputs from these individual Osynergy = max(O/l, . . . ,017) (5) 1. Masood, R HassanlExpert S y s t e m s w i t h Applications 41 (2014) 7579-7595 Bivariate shift patte~nsforrnoderate data correlation (p = 0.5) .- Partially developed shift Fully developed shift -5.0 -2.5 0.0 2.5 5.0 X l (Up-Shift) -5.0 -2.5 0.0 2.5 5.0 X I (Up-Shift) -5.0 -2.5 0.0 2.5 5.0 X I (Nonnal) h - 4 3 2.5 - w z . . b 0.0 3 .- . 5 - - -. 2 -25 5- -5.0 -5.0 -5.0 -2.5 0.0 2.5 5.0 -5.0 -2.5 0.0 2.5 5.0 X1 (Up-Shift) XI (Up-Shift) XI (Normal) u X I (Down-Shift) krFzFiT X I (Down-Shift) Fig. 3 ( c o n t i n u e d ) Multilayer perceptrons (MLP) model trained with back-propa- 48 neurons, while statistical features input representation gation (BPN) algorithm was applied for the individual ANNs. This requires only 14 neurons. The output layer contains seven neu- model comprises an input layer, one or more hidden layer(s) and rons, which was determined according to the number of pattern an output layer. The size of input representation determines the categories. Based on preliminary experiments, one hidden layer number of input neurons. Raw data input representation requires with 26 neurons and 22 neurons were selected for raw I. Mflsood, A. HossanfExpert S y s t e m s w i t h Applications 41 (2014) 7579-7595 X I (Down-Sh~ft) XI ( Down-Shift ) - --- - L --- -- ! Fig. 3 ( c o n t i n u e d ) data-based ANN and statistical features-ANN. The experiments did not improve the training results but provided poorer results. revealed that initially, the training results improved in-line with These excess neurons could burden the network computationally, the increment in the number of neurons. Once the neurons reduces the network generalization capability and increases the exceeded the required numbers, further increment of the neurons training time. I. Masood. R HassanlExpert Systems with Applications 41 (2014) 7579-7595 i......... ............................. ......................... ! First stage MEWMA , monitol-ing / control chart Next Yes (out-of-control) 11 Troubleshooting I and I i renew setting / / i ................................................ : I; t I YCS (out-of-control) I I ; Identify the sources of mean shift I ................................................................................................................................... Fig. 4. Frameworlc for the 2s-IMS. 3.3. lnput representation lnput representation is a technique to represent input signal into ANN for achieving effective recognition. There are various approaches could be used to represent input signal. Raw data (standardized samples) is the basic approach (Zorriassatine, Tannock, & O'Bricn. 2003). Besides raw data, feature-based approach that involves extracted features from raw data is one of the successful technique in image processing (Br~~nzcll & Eriltsson, 2000: I . ~ Raw data- own.. Kaw a a b a s e d Statistical F e a t u r e s Features- OF.S I I Fig. 6. Synergistic-ANN model. controlling or monitoring. Insufficient denoising will distort recognizer for improving pattern discrimination capability. Raw waveforms and introduce errors. Inversely, excessive denoising data input representation consists of 48 data, i.e., 24 'consecutive will over-smooth the sharp features of underlying signals by standardized samples of bivariate process (Z1.P1l Z1.p2,. . . .Z24.P1. recognizing them as noise or outliers. Z24.P2). Statistical features input representation consists of last In this research, raw data and improved set of statistical features value of exponentially weighted moving average (LEWMA]) with were applied separately into training of the Synergistic-ANN A = [0.25,0.20,0.15,0.10], mean ( p ) , multiplication of mean with I. Masood, A HassanlExpert Systems w i t h Applicatioiu 41 (2014) 7579-7595 Fig. 7. Individual ANN recognizer. W L g r e r - L w e r L W (48 01 02 0 3 0 4 4 0 6 0 7 Raw data-ANN standard deviation (MSD), and multiplication of mean with mean square value (MMSV). Each bivariate pattern was represented by 14 data as follows: LEWMA0.25-~~, LEWM&.20-P1, LEWM&.15.Pl, LEWMAO.IO-PI, , & I , MSDPI. MMSVpi, LEwMAo.25-~2, LEWMAo.2o-p2. LEWMAO.IS-P~. LEWMAO.IO-PZ~ PPZ, MSDPZ, MMSVp2. LEWMAn features were talten based on observation win- dow = 24. The EWMA-statistics as derived using Eq. (6) incorpo- rates historical data in a form of weighted average of all past and current observation samples (Lucas cF Saccucci. 1990): W L W - L w ~ L w (14 0 1 02 0 3 0 4 05 0 6 0 7 Statistical feature-ANN Xi represents the original samples. In this study, the standardized samples (Zi) were used instead of Xi so that Eq. (6) becomes: where 0 < A ,< 1 is a constant parameter and i = [ I , 2 , . . . ,241 are the number of samples. The starting value of EWMA (EWMb) was set as zero to represent the process target (PO). Four value of constant parameter (A = 0.25,0.20,0.15,0.10) were selected based on a range within [0.05,0.40] recommended by Lucas and Saccurci (1990). Besides resulting longer ARLO, these parameters could influence the performance of EWMA control chart in detecting process mean shifts. Preliminary experiments suggested that the EWMA with small constant parameter ( A = 0.05) were more sensitive in identify- ing small shifts ( ~ 0 . 7 5 standard deviations), while the EWMA with large constant parameter (A = 0.40) were more sensitive in identify- ing large shifts (>2.00 standard deviations). The MSD and MMSV features were used to magnify the magnitude of mean shifts ( P ~ ~ P Z ) : where (,u1,p2), ( a l , 0 2 ) (,uf,p:) are the means, standard deviations and mean square value respectively. The mathematical expressions of mean and standard deviation are widely available in textbook on SPC. The mean square value feature can be derived as in Hassail ct al. (2003). Further discussion on selection of statistical features can be found in Masood ancl Hassan (2013). 3.4. Recognizer training and testing Partially developed shift patterns and dynamic patterns were applied into the ANN training and testing respectively since these approaches have been proven effective to suit for on-line process situation (Gi~h, 2007). Detail parameters for the training patterns are summarized in Tablcs 1 ancl 2. In order to achieve the best training result for overall pattern categories, the amount of training patterns were set as follows: (i) bivariate normal patterns = [I500 x (total combination of data correlation)] and (ii) bivariate shift patterns = 1100 x (total combi- nation of mean shifts) x (total combinations of data correlation)]. In order to improve discrimination capability between normal and shift patterns, a huge amount of N (0,O) patterns was applied into ANN training. The US ( 1 , l ) and DS ( 1 , l ) pattern categories also require a huge amount of training patterns since it contain a more complex combination of mean shifts compared to the other bivar- iate shifts pattern categories. Guh (2007) reported that the utilization of partially developed shift patterns in ANN training could provide the shorter ARLl results. In order to achieve the best ARLl results for this scheme, different percentage of partially developed shift patterns were utilized for different range of mean shifts as shown in Table 2. The starting points of sudden shifts (SS) were determined empiri- cally. The actual value of data correlation is dependent to the var- iability in the bivariate samples. The simulated values ( p = 0.1,0.3, 0.5, 0.7. 0.9) as shown in Table 1 could only be achieved when the process data streams are in fully normal pattern or in fully devel- oped shift pattern. Input representations were normalized to a compact range between [-1,+1]. The maximum and the minimum values for normalization were talcen from the overall data of train- ing patterns. Based on BPN algorithm. 'gradient decent with momentum and adaptive learning rate' (traingdx) was used for training the MLP model. The other training parameters setting were learning rate (0.05) learning rate increment (1.05). maximum number of epochs (500) and error goal (0.001), whereas the network performance was based on mean square error (MSE). Hyperbolic tangent func- tion was used for hidden layer, while sigmoid function was used for an output layer. The training session was stopped either when the number of training epochs was met or the required MSE has been reached. 4. Performance results and discussion The monitoring and diagnosis performances of 2s-IMS were evaluated based on average run lengths (ARLo,ARLl) and recogni- tion accuracy (RA) as summarized in Table 4. The ARLs results were also compared to the traditional multivariate statistical process control (MSPC) charting schemes such as X 2 (Hotelling, 1947). MCUSUM (Pignatiello & Kunger, lf)90), and MEWMA (Lowry et dl., 1992), as reported in the literature. In order to achieve balanced monitoring and accurate diagnosis, the proposed 2s-IMS should be able to achieve the target perfor- mances as follows: I. Masood. A. Hossan/Expert Systems with Applications 41 (2014) 7579-7595 Table 1 . . Parameters for t h e training patterns. Pattern category M e a n shift ( i n standard deviations) Data correlation ( p ) Amount of training patterns N ( 0 , o ) X I : 0.00 0.1, 0.3, 0.5, 0.7, 0.9 1500 x 5 = 7500 X2: 0.00 US (1,o) X l : 1.00, 1.25,. . ..3.00 100 x 9 x 5 = 4 5 0 0 X2: 0.00, 0.00,. . .,o.oo US ( 0 , l ) X2: 0.00, O . O O , . . ..O.OO l O O x 9 x 5 = 4 5 0 0 Xl: 1.00, 1.25,. . ..3.00 us (1.1) Xl: 1.00, 1.00, 1.25, 1.25,. ..,3.00 100 x 25 x 5 = 12.500 X2: 1.00, 1.25, 1.00, 1.25. . . ..3.00 DS (1.0) X l : -1.00, -1.25.. . ., -3.00 l O O x 9 x 5 = 4 5 0 0 X2: 0.00, o.oo,.. ..o.oo DS ( 0 , 1 ) X2: 0.00, 0.00 ,..., 0.00 lOOx9x5=4500 X l : -1.00, -1.25,. . . , -3.00 0s ( 1 , 1 ) Xl: -1.00, -1.00, -1.25, -1.25,. . .,-3.00 100 x 25 x 5 = 12,500 X 2 : -1.00, -1.25, -1.00, -1.25,. . ..-3.00 Table 2 Parameters for the partially developed shift patterns. Range of mean shifts ( i n standard deviations) A m o u n t of partially developed shift patterns Starting point of sudden shift (SS) Sample 9 t h Sample 1 3 t h Sample 1 7 t h Table 3 Summary of monitoring-diagnosis capabilities. Traditional MSPC 2s-IMS Effective in monitoring (to identify out-of-control signal) Limited to avoid false alarm (ARLO r 200) Unable to identify the sources of variation (mean shifts) Comparable to the traditional MSPC in monitoring aspect Capable to maintain smaller false alarm (ARLO >> 200) High accuracy in identifying the sources of variation (mean shifts) (i) ARI, >> 200 to maintain small false alarm in monitoring bivariate in-control process. (ii) Short ARLl (average ARLl 6 7.5 for shifts range k0.75-3.00 standard deviations) to rapidly detect bivariate process mean shifts. (iii) High RA (average RA 2 95% for shifts range 20.75-3.00 stan- dard deviations) to accurately identify the sources of mean shifts. 4.1. Monitoring performance In monitoring aspect. the ARb represents the average number of natural observation samples of in-control process before the first out-of-control process signal exist as a false alarm. In other word, the ARLO measures how long a SPC scheme could maintain an in- control process running without any false alarm. On the other hand, the ARLl represents the average number of unnatural obser- vation samples before it is truly identified as out-of-control process signal. In other word, the ARLl measures how fast a SPC scheme could detect process mean shifts. Further discussion on this mea- sure can be referred to Montgoluery (2009). Ideally, a SPC scheme should provide ARLO as long as possible in order to minimize cost for investigating the discrepancy and trou- bleshooting while the process still within control. Meanwhile, it should provide ARLl as short as possible in order to minimize cost for reworlts or waste materials. Since t h e false alarm cannot be eliminated, the ARLO >> 200 is considered as the de facto level for balanced monitoring. In this research, the ARLs results of 2s-IMS were simulated based on correctly classified patterns. Generally, it can be observed that the smaller the mean shifts, the longer the ARLl values. This trend support the conclusion that process mean shifts with smaller magnitudes would be more difficult to detect. Specifically, the 2s-IMS indicated rapid detection capability for large shifts (shifts = 3 a . ARL1 = 3.18-3.19) and moderate shifts (shifts = 2 a , ARLl = 4.76-4.78) with short ranges of ARL1. It was also capable to deal with smaller shifts (shifts=Ilo.0.75a], ARL1= 110.33-10.60,15.69-16.751). In comparison to the X2 charting scheme, detection capability as shown by 2s-IMS was faster for small and moderate shifts (shifts = 0.75-20). In comparison to the MCUSUM and the MEWMA, it was slightly comparable in rapid detection for large shifts (shifts = 2.50, ARL,: 2s-IMS = 3.80-3.81, MCUSUM = 2.91, MEWMA=3.51) and moderate shifts (shifts=1.5o. ARL,: 2s-IMS = 6.41-6.52, MCUSUM = 5.23; MEWMA = 6.12). Similar trend can also be found when dealing with smaller shifts (shifts = la, ARLI: 2s-IMS = 10.33-10.60, MCUSUM = 9.28; MEWMA = 10.20). Meanwhile, based on the range of ARLO results ( p = 0.1,0.5, 0.9; ARLo=335.01, 543.93, 477.45), the 2s-IMS was observed to be more effective in maintaining smaller false alarm compared to the traditional MSPC (ARb r 200). It should be noted that the results for medium and high correlation processes have exceeded 370 as shown in the Shewhart control chart (Nels011, 1985; Shewhart, 1931). Overall, it can be concluded that the proposed scheme indicated balanced monitoring performance. 4.2. Diagnosis performance In diagnosis aspect, the RA measures how accurate is a SPC scheme could identify the sources of mean shifts towards diagnos- ing the root cause error and conducting troubleshooting. Generally, it can be observed that the smaller the mean shifts, the lower the RA results. This trend supports the conclusion that diagnosis infor- mation for small process mean shifts (61.0 standard deviations) - 7590 I. Masood, A. HassanjExpert Systems with Applications 41 (2014) 7579-7595 Table 4 . Performance comparison between t h e 2s-IMS and t h e traditional MSPC. Pattern category Mean shifts Average run lengths Recognition accuracv 2s-IMS X2 UCL = 10.6 MCUSUM k = 0.50 h = 4.75 MEWMA I = 0.10 H = 8.66 2s-IMS XI X2 ARLO for p = 0.1. 0.5. 0.9 ARLO for p = 0.0 RA for p = 0.1, 0.5. 0.9 N (0,o) 0.00 0.00 335.01. 543.93. 477.45 200 (0.005) 203 (0.0049) 200 (0.005) N A ARL, for p = 0.1, 0.5, 0.9 us (1,O) 0.75 0.00 17.60. 18.34, 20.00 92.7. 90.4. 89.5 US ( 0 , l ) 0.00 0.75 16.20, 15.99, 16.21 92.9. 89.3. 90.6 US (1.1) 0.75 0.75 13.64. 13.28, 14.17 82.4. 94.8, 99.9 DS (1,O) -0.75 0.00 16.31, 16.43, 17.35 92.3, 89.2. 89.4 Ds ( 0 , l ) 0.00 -0.75 16.94, 17.44, 18.75 92.3. 87.8. 88.5 DS (1.1) -0.75 -0.75 13.46. 13.37. 14.03 84.1. 96.1. 99.9 Average 15.69, 15.81, 16.75 89.5, 91.3, 93.0 us (1,o) 1 .OO 0.00 11.52. 11.57, 11.70 42-0.976 9.28-0.892 95.3, 93.1, 94.4 us ( 0 , l ) 0.00 1 .OO 10.50, 10.22, 10.20 95.8. 93.5, 94.4 Us ( 1 , l ) 1 .OO 1.00 9.1 6. 9.09. 9.66 90.0, 96.5, 100 Ds (1.0) -1 .OO 0.00 10.99, 10.86, 11.06 95.3. 93.2. 92.3 0 s ( 0 , l ) 0.00 -1.00 11.08, 11.12, 11.36 93.8, 92.1, 92.6 Ds (1,1) -1.00 -1.00 9.15. 9.12. 9.63 89.5. 98.0. 100 Average 10.40, 10.33, 10.60 93.3. 94.4. 95.6 us (1,O) 1.50 0.00 7.02, 7.07, 7.03 15.8-0.937 5.23-0.809 97.4, 96.5, 97.1 us (0,1) 0.00 1.50 6.54, 6.33, 6.40 97.1. 96.5. 96.2 U s ( 1 , l ) 1.50 1.50 5.82, 5.73. 5.94 91.7. 97.9. 100 DS (1,O) -1.50 0.00 6.81, 6.81, 6.92 97.4, 96.3, 95.5 0 s ( 0 , l ) 0.00 -1.50 6.82, 6.80, 6.85 96.2. 95.8. 95.6 0s (1,1) -1.50 -1.50 5.81. 5.69. 5.98 93.2. 99.0. 100 Average 6.47, 6.41, 6.52 95.5. 97.0. 97.4 us (1.0) 2.00 0.00 5.23, 5.15, 5.19 97.8, 97.1, 97.6 Us (0.1) 0.00 2.00 4.80. 4.72, 4.70 97.7. 97.8. 97.1 Us (1,1) 2.00 2.00 4.36, 4.32, 4.39 91.6, 98.4, 100 DS (1,o) -2.00 0.00 5.04, 5.04. 5.02 96.8. 96.7. 96.6 0 s (0,1) 0.00 -2.00 4.97, 5.03, 4.98 96.5. 96.5. 95.6 DS(1,l) -2.00 -2.00 4 29, 4.27 4 33 93.7. 98.9. 100 Average 4.78. 4.76. 4.77 95.7, 97.6. 97.8 Us (1,O) 2.50 0.00 4.10, 4.14, 4.12 98.0, 98.4, 98.0 US (0.1) 0.00 2.50 3.83, 3.81, 3.81 97.3. 97.4, 97.0 US (1.1) 2.50 2.50 3.54, 3.49, 3.53 93.2, 98.4, 100 DS (1,O) -2.50 0.00 3.99. 3.96. 3.95 97.3. 97.3, 97.0 0 s ( 0 , l ) 0.00 -2.50 3.97. 4.02. 3.98 96.5. 96.6. 97.0 D S ( 1 , l ) -2.50 -2.5O 3.41.3.40. 3.46 94.9. 98.8. 100 Average 3.81, 3.80, 3.81 96.2. 97.8. 98.2 Us (1.0) 3.00 0.00 3.47, 3.46, 3.46 98.6, 98.3. 98.2 U s (0.1) 0.00 3.00 3.20, 3.20. 3.21 97.8. 97.8, 98.0 US ( 1 , l ) 3.00 3.00 2.98, 2.93, 2.98 93.8. 98.4, 100 DS (1,o) -3.00 0.00 3.31, 3.30. 3.27 98.0, 97.1. 97.6 0 s ( 0 , l ) 0.00 -3.00 3.33.3.32, 3.32 96.7, 97.1, 97.1 DS(1,l) -3.00 -3.0° 2.84. 2.85. 2.90 94.6. 99.1. 100 Average 3.19, 3.18. 3.19 96.6. 98.0, 98.5 Grand average +(0.75-3.00) 7.39, 7.38, 7.61 94.5. 96.0, 96.8 Note: Design parameters for MEWMA control c h a r t in 2s-IMS (1= 0.1, H = 8.64). would be more difficult to identify. Specifically, the 2s-IMS indicated accurate diagnosis capability for large shifts (shifts = 30, RA = 96.6-98.5%) and moderate shifts (shifts = 20, RA = 95.7- 97.8%) with high ranges of RA. Although the results were slightly Flange Internal diametc degraded, it is still effective to deal with smaller shifts (shifts = [lo,0.75o], RA= [93.3-95.6%,89.5-93.0%]). It should be 1' noted that the RA results for medium and highly correlated pro- /' cesses were higher compared to low correlation process, which is i effective for practical case. Since the traditional MSPC charting i 1, n schemes were unable to provide diagnosis information, diagnosis \, 1" capability as shown by 2s-IMS was absolutely capable in solving \~ such issue. Overall, it can be concluded that the proposed scheme '> 1. indicated accurate diagnosis performance. Table 3 summarizes the comparison of monitoring-diagnosis capabilities between the Groove & flange view Roller head 2s-IMS and the traditional MSPC. Fig. 8. Functional features of roller head. I. Masood, A. HassanjExpert Systems with Applications 41 (2014) 7579-7595 Extnision round bar Tuining to rough size Turning to size Honing inner diameters N~ckel electroplating Beal-ings assembly Fig. 9. Process plan for the manufacture of roller head. 5. Industrial case study . Broadly, the need for BQC could be found in manufacturing industries involved in the production of mating, rotational or mov- ing parts. Investigation for this study was focused on the manufac- turing o f audio video device (AVD) component, namely, roller head. This investigation was based on the author's working experience in manufacturing industry in Johor, Malaysia. In an AVD, the roller head functions to guide and control the movement path of a film tape. Inner diameters of roller head (ID1 and ID2) as shown in Fig. 8 a r e two dependent quality characteristics (bivariate) that need for joint monitoring-diagnosis. In current practice, such func- tional features are still widely monitored independently using Shewhart control charts. It is unsure why the MSPC was not imple- mented. Based on the author's point of view, it could be due to lack of motivation, ltnowledge and sltills to adapt new technology. The process plan for the manufacture of roller head can be illus- trated in Fig. 9. Initially, an aluminium extrusion round bar was turned t o rough size (rough cut machining). Then, it was turned to size (finish cut machining) to form functional features such as inner diameters, and groove and flange, among others. The machining of inner diameters was then extended into honing pro- cess to achieve tight tolerance for bearing assembly. Hard coated surface was also necessary. As such, the machined work-piece was electroplated using nickel alloy before assembly. rTool rRoller head <. ..-.---.-. ,./' Tool bluntness Loading error (Decrement in ID) (Increment in ID) Fig. 10. Process variation occurred in turning-to-size operation. automatically loaded into pneumatic chuck using a robotic system. Bluntness in the cutting tool will cause gradual decrement in both inner diameters (ID1 ,ID2) with positive cross correlation ( p > 0). In another situation. such inner diameters could be suddenly increased simultaneously and yields positive cross correlation ( p > 0) due to loading error. Based on two examples of bivariate process variation, industrial process samples were simulated into the 2S-IMS for validating its applicability in real world. The first case study involves tool bluntness. The mean ( p ) and standard deviation (o) of bivariate Bivariate process variation can be found in turning to size oper- in-control process were determined based on the first 24 samples ation d u e to tool bluntness and loading error as illustrated in (observations lst-24th). Tool bluntness begins between observa- Fig. 10. These disturbances will cause unnatural changes in t h e tion samples 41st-50th. Validation results are summarized in process data streams as shown in '1-able 5. The work piece is Table (5, whereby the determination of process status (monitoring) Table 5 Sources of variation in machining inner diameters. Stable process Process noise N (0,O) Tool bluntness DS ( 1 , l ) Loading error US (1.1) XI., ([Dl) 1~:: f l w w t . I-"_- -- N o r m a l _ N o r m a l " k M L-.. Down-Trend I -" Down-Trend h L Up-Shift Scatter diagram i.01-1 u XI ( Down-Trend ) I. Masood R HassanjExpert Systems with Applications 41 (2014) 7579-7595 - - 7592 Table 6 . - . Inspection results based on tool bluntness case. i Original samples Standardized samples Window range Monitoring-diagnosis decision Xi.1 (ID11 Xi-2 (ID2) 4 1 ((Dl) Zi-1 (ID2) 1 7.9420 7.9428 0.3393 1.0790 2 7.941 2 7.9420 -1.1414 -0.591 7 3 7.9412 7.941 6 -1.1414 -1.4271 4 7.9420 7.9428 0.3393 1.0790 5 7.9412 7.9420 -1.1414 -0.5917 6 7.9412 7.941 6 -1.1414 -1.4271 7 7.9420 7.9428 0.3393 1.0790 8 7.9424 7.9420 1.0797 -0.5917 9 7.941 6 7.9420 -0.401 0 -0.591 7 1 0 7.941 2 7.941 6 -1.1414 -1.4271 11 7.941 6 7.9424 -0.401 0 0.2437 1 2 7.9428 7.9432 1.8201 1.9144 1 3 7.9420 7.9424 0.3393 0.2437 1 4 7.941 6 7.9424 -0.4010 0.2437 1 5 7.9424 7.9428 1.0797 1.0790 16 7.9412 7.942 -1.1414 -0.5917 1 7 7.941 2 7.9416 -1.1414 -1.4271 1 8 7.9420 7.9424 0.3393 0.2437 1 9 7.9428 7.9428 1.8201 1.0790 20 7.9420 7.9424 0.3393 0.2437 2 1 7.941 2 7.9416 -1.1414 -1.4271 22 7.9424 7.9428 1.0797 1.0790 23 7.9424 7.9424 1.0797 0.2437 2 4 7.9420 7.9424 0.3393 0.2437 1-24 N 25 7.941 2 7.941 6 -1.1414 -1.4271 2-25 N 26 7.9424 7.9420 1.0797 -0.591 7 3-26 N 27 7.9424 7.9428 1.0797 1.0790 4-27 N 28 7.941 2 7.9420 -1.1414 -0.591 7 5-28 N 29 7.9420 7.9428 0.3393 1.0790 6-29 N 3 0 7.9420 7.9424 0.3393 0.2437 7-30 N 3 1 7.9412 7.9420 -1.1414 -0.5917 8-31 N 3 2 7.9420 7.9428 0.3393 1.0790 9-32 N 3 3 7.9428 7.9424 1.8201 0.2437 10-33 N 3 4 7.941 6 7.9424 -0.4010 0.2437 11 -34 N 3 5 7.9424 7.9432 1.0797 1.9144 12-35 N 3 6 7.9428 7.9424 1.8201 0.2437 13-36 N 3 7 7.941 6 7.9420 -0.401 0 -0.591 7 14-37 N 3 8 7.9420 7.9424 0.3393 0.2437 15-38 N 3 9 7.9424 7.9420 1.0797 -0.5917 16-39 N 40 7.941 6 7.9420 -0.4010 -0.5917 17-40 N 41 7.9408 7.9412 -1.8818 -2.2625 18-41 N 42 7.9408 7.9408 -1.8818 -3.0978 19-42 N 43 7.9404 7.9408 -2.6222 -3.0978 20-43 N 44 7.9404 7.9408 -2.6222 -3.0978 21-44 DS ( I 1 ) 4 5 7.9404 7.9404 -2.6222 -3.9332 22-45 DS ( I I ) 46 7.9400 7.9404 -3.3626 -3.9332 23-46 D5 (1 1 ) 4 7 7.9400 7.9400 -3.3626 -4.7686 24-47 DS (1 1 ) 48 7.9396 7.9400 -4.1029 -4.7686 25-48 DS (1 1) 4 9 7.9396 7.9396 -4.1 029 -5.6040 26-49 DS ( I I ) 50 7.9396 7.9396 -4.1 029 -5.6040 27-50 DS ( I 1 ) ( p l , p z ) = (7.9417,7.9422): (g,,02) = (4.6687 x 10-~,4.2495 x Note: Observation samples highlighted in bold (41st-50th) represent out-of-control process. Table 7 Outputs of the scheme for tool bluntness case. Reco~nition window (RW) 1-24 2-25 3-26 4-27 5-28 6-29 7-30 8-31 9-32 P Decision based on MEWMA control chart RW P Decision based on MEWMA control chart RW P N US (I 0 ) US ( 0 1 ) us (1 1 ) DS (1 0) DS ( 0 1 ) DS (1 1 ) Note: Bold value represents the maximum output of ANN that determines pattern category. I. Mosood. A. HonanfExpert Systems with Applications 41 (2014) 7579-7595 7593 and sources of variation (diagnosis) are based on output of the scheme as shown in Table 7. In t h e first 40 samples, this scheme was able to correctly recog- nize t h e bivariate process data streams as in-control patterns (N). In this case, it was effective to identify bivariate in-control process without triggering any false alarm. Bluntness of the cutting tool begins a t sample 41st, whereby this scheme was able to correctly recognize bivariate process data streams as Down-Shift patterns (DS (1 , I ) ) starting from sample 44th (at window range 21st-44th). In overall diagnosis aspect, this scheme was observed to be effective to identify the sources of variation in mean shifts without mistalte. The second case study involves loading error. Similar as in the first case study, the mean ( p ) and the standard deviation (o) of bivariate in-control process were computed based on the first 24 observation samples. Loading error exist between samples 40th- 50th. Validation results and related output of the scheme are sum- marized in Tables S a n d 9 respectively. Based on t h e first 39 samples, this scheme is effective to cor- rectly recognize the bivariate process data streams as in-control patterns (N). In this situation, the process was running smoothly without false alarm. Improper condition of pneumatic chuck and robotic arm causes loading error between samples 40th-50th. In this situation, this scheme was able to correctly recognize the bivariate process data streams as Up-Shift patterns (US (1 , l ) ) start- ing from sample 40th (at window range: 17th-40th). In overall diagnosis aspect, this scheme is capable to correctly identify the sources of variation in mean shifts without mistalte. Table 8 Inspection results based on loading error case. i Original sa~nples Standardized samples Window range Monitoring-diagnosis decision XILI ([Dl 1 x,.z (ID21 Z i ~ l (ID1 ) zi-i (ID21 1 7.941 6 7.9420 -0.2856 -0.5099 2 7.9412 7.9428 -1.1424 1.3727 3 7.9420 7.9424 0.5712 0.43 1 4 4 7.941 2 7.941 6 -1.1424 -1.4512 5 7.9420 7.9428 0.571 2 1.3727 6 7.941 2 7.9420 -1.1424 -0.5099 7 7.941 2 7.941 6 -1.1424 -1.4512 8 7.941 6 7.9424 -0.2856 0.4314 9 7.9424 7.9420 1.4279 -0.5099 10 7.941 6 7.9420 -0.2856 -0.5099 11 7.941 2 7.941 6 -1.1424 -1.4512 1 2 7.9424 7.9428 1.4279 1.3727 13 7.9420 7.9424 0.571 2 0.4314 14 7.941 6 7.9424 0 . 2 8 5 6 0.4314 15 7.9412 7.9416 -1.1424 -1.4512 16 7.9424 7.9428 1.4279 1.3727 17 7.941 6 7.9420 -0.2856 -0.5099 18 7.9420 7.9424 0.5712 0.4314 19 7.9412 7.9420 -1.1424 -0.5099 20 7.9424 7.9424 1.4279 0.4314 21 7.9420 7.9424 0.5712 0.4314 22 7.9420 7.9424 0.5712 0.43 1 4 23 7.941 2 7.9416 1 . 1 424 -1.4512 24 7.9424 7.9428 1.4279 1.3727 1-24 N 25 7.9420 7.9424 0.571 2 0.4314 2-25 N 26 7.9412 7.9416 -1.1424 -1.4512 3-26 N 27 7.9424 7.9420 1.4279 -0.5099 4-27 N 28 7.9424 7.9428 1.4279 1.3727 5-28 N 29 7.9412 7.9420 -1.1424 -0.5099 6-29 N 30 7.9420 7.9428 0.5712 1.3727 7-30 N 3 1 7.9428 7.9424 2.2847 0.4314 8-31 N 32 7.9420 7.9424 0.5712 0.43 1 4 9-32 N 33 7.941 2 7.9420 -1.1424 -0.5099 10-33 N 34 7.9420 7.9428 0.5712 1.3727 11-34 N 35 7.9428 7.9424 2.2847 0.43 1 4 12-35 N 36 7.941 6 7.9424 -0.2856 0.4314 13-36 N 37 7.9424 7.9428 1.4279 1.3727 14-37 N 3 8 7.9416 7.9420 -0.2856 -0.5099 15-38 N 39 7.9428 7.9424 2.2847 0.43 1 4 16-39 N 40 7.9428 7.9432 2.2847 2.3140 17-40 US (1 1 ) 41 7.9432 7.9428 3.1415 1.3727 18-41 US (1 1 ) 42 7.9436 7.9432 3.9982 2.3140 19-42 US (1 1 ) 43 7.9428 7.9432 2.2847 2.3140 20-43 US (1 1 ) 44 7.9432 7.9428 3.1415 1.3727 21 -44 US (1 1 ) 45 7.9436 7.9432 3.9982 2.3140 22-45 US (1 1 ) 46 7.9428 7.9432 2.2847 2.3140 23-46 US (1 1 ) 47 7.9432 7.9428 3.1415 1.3727 24-47 US (1 1 ) 48 7.9428 7.9436 2.2847 3.2553 25-48 US(1 I ) 49 7.9428 7.9432 2.2847 2.3140 26-49 US (1 1 ) 50 7.9436 7.9432 3.9982 2.3140 27-50 US(1 1 ) (/L,,/L~) = (7.9417,7.9422); ( u 1 . u 2 ) = (4.6687 x 10-4,4.2495 x Note: Observation samples highlighted in bold (40th-50th) represent out-of-control process. I. Masood, A. HassanlExpert Systems with Applications 41 (2014) 7579-7595 Table 9 . - Outputs of t h e scheme for loading error case. - - W i n d o w range (RW) 1-24 2-25 3-26 4-27 5-28 6-29 7-30 8-31 9-32 P 0.6896 0.6910 0.8333 0.7723 0.7733 0.7822 0.7733 0.7234 0.7407 Decision based o n MEWMA control c h a r t N N N N N N N N N RW 10-33 11-34 12-35 13-36 14-37 15-38 16-39 17-40 18-41 P 0.7924 0.7753 0.7202 0.6941 0.7075 0.7254 0.6693 0.6973 0.7040 N N N N N N N N 0.8220 0.4510 US (1 0) 0.5067 0.6528 US (01) 0.1 665 0.1370 US (1 I ) 0.9479 1.1855 DS ( 1 0) 0.0985 0.0719 DS (01) 0.1533 0.1453 DS (1 1) 0.1257 0.1129 RW 19-42 20-43 21-44 22-45 23-46 24-47 25-48 26-49 27-50 P 0.7546 0.7504 0.7561 0.781 5 0.7806 0.7486 0.7258 0.7228 0.6886 N 0.1775 0.1609 0.1053 0.0434 0.0403 0.0376 0.0313 0.0318 0.0200 US (1 0) 0.7116 0.4926 0.5886 0.6606 0.4464 0.5521 0.3923 0.2824 0.3099 US (01) 0.1184 0.1124 0.1317 0.1724 0.1631 0.1540 0.1762 0.1711 0.2213 US (1 1 ) 1.4012 1.6147 1.5717 1.5235 1.6681 1.5983 1.7006 1.7666 1.7163 DS (1 0) 0.0632 0.0852 0.0817 0.0766 0.0802 0.0970 0.1002 0.1142 0.1221 DS (0 1) 0.1537 0.1508 0.1350 0.1830 0.2023 0.1735 0.1754 0.2060 0.2468 DS (1 1) 0.1304 0.1144 0.0875 0.0636 0.0597 0.0384 0.0434 0.0493 0.0283 Note: Bold value represents t h e m a x i m u m o u t p u t of ANN t h a t determines p a t t e r n category. G . Conclusions This paper proposed two-stage monitoring approach in moni- toring and diagnosis of bivariate process variation in mean shifts. Based on the frameworlc of 2s-IMS that integrates the powerful of MEWMA control chart and Synergistic-ANN recognizer, it has resulted in a smaller false alarm ( A R b = 335.01-543.93), rapid shifts detection (ARL1 = 3.18-16.75), and accurate diagnosis capa- bility (RA = 89.5-98.5%) compared to the traditional SPC charting schemes for BQC. Since the monitoring and diagnosis performances were evaluated using modeling data, real industrial data were used for the purpose of validation. The case studies involved tool bluntness and loading error in machining operations, whereby the proposed scheme has shown a n effective monitoring capability in identifying the bivariate in-control process without any false alarm. The scheme also effective in diagnosis aspect, that is, in cor- rectly identifying the sources of mean shifts when process becomes out-of-control. Based on the promising results, the 2s-IMS could be a reference in realizing balanced monitoring and accurate diagnosis of bivariate process variation. In the future work, further investigation will be extended to other causable patterns such as trends and cyclic. Aclcnowledgements The authors would like to thank Universiti Tun Hussein Onn Malaysia (UTHM). Universiti Teltnologi Malaysia (UTM), and Ministry of Higher Education (MOHE) of Malaysia who sponsoring this work. References Al-Assdf, Y. (2004). Recognition of control c h a r t p a t t e r n s using ~nulti-resolution wavelets analysis and ncul-dl networlts. Conlputers. a r ~ d I~~dustrialEngir~ecr~~~g, 47, 17-29. AIL, F. B. (1985). Multivariate quality contl-ol. In N. L. Johnson 8 S. IKotz (Etls.). Encvclopedia of.stotistica1 mcierices (vol. 6). New Yol-It: Wiley. Assaleh. I<.. 4 AI-Assal, Y. (2005). Features extraction a n d analysls for cldss~fying cdusable patterns in c o n t ~ - o l rhal-ts, CoinputeIs aitd Industrial Engirzrering, 49, 168..181. Bersimis, S., Psaraliis, 5 . . 8 I'anaretos, 1. (2007). Multivariate ~ t a t i s l ~ c a l pi-ocess control c h a m : An ovel-view. @ruliLy and Rellabilrty Engineeririg I ~ ~ r e r r ~ a t i t r n a i , 23, 5 17-543. BI-iinzell, I-]., 8 Eriltssnn. J. (2000). Feature recluction for cla.ss~fication of r n u l t i d ~ m e n s ~ o n a l data. Porter-ri Recognition, 33(121, 1741 - - 1748. Clien, 2.. Lu. S., 8 Lam, S. (2007). A hybrid system Ibr SI'C concurrenl patlel-11 recognilion. /ldvclnced Enginc.eriilg Inloi-~iialics. 21, 303-3 10. Clien. L. H.. & Wang, T. Y. (2004). AI-Lificial neural networlts lo classify mcan shifts from niultivariate %%hart s ~ g n a l s . Computers olld 111dush.iol Engineerillg, 47, 195-~205. Clieng, C. S., & Cheng, 1-1. 1'. (2008). I d e n t i f y ~ n g the source of variance shifts in the rnultivai-iafe pl-occss using neural netwol-Its and s u p p o r l vector machines. Expert 5,y.stern.s wit11 Applico1ror1.s. 3.5, 198-206. Chih. W. H.. 4 Kollier, D. A. (1994). U ~ a g n o s i s charactcristics Sor hivariale p a t t e r n recogliltion s c h c m e in SPC. Iruernalionul Jouri~o! o j r2uolity u r ~ d Reliability Molitr.yrnelit, l I ( 1 ) . 53-66. Chih. W. I<., 8 Rollier. I). A. (1995). A ~ ~ i e t h o d o i o g y of pattern recogmtion s c h e m e s for t w o variables in SPC. Il~terr~ationol jou1-11a1 of Q~ialin, and I(eliability Mnnngenrent, 12(3), 86-107. Crosiel-, R. B. (1988). M u l t ~ v a r ~ a t e g e n e ~ a l ~ z a t i o n s of c u ~ n u l a t i v e s u m q w l ~ t y c.ontrol schemes. 7'ech11ometrics. 30(3), 291-303. El-Midany. I . T., El-Baz, M. A,. & Abtl-Elwahecl. M. S. (2010). A proposccl Irarneworlt for conlrol c h a r t p a t t c r n recognition in multivariale process using &I-tificial neural networlts. Expert Systems with Aljplicutio~is, 37, 1035- 1042. Fuchs. C.. 4 Benjamini, Y. (1994). M u l t ~ v a r i a t e profile chart? for s t a t ~ s t i c a l ~ l r o c e s s control. Technom~trics, 36(2), 182- 195. Gauri, S. I<., 8 Chaltraborty, 5 . (2006). Feature-l~ased recogiiltlon of control c h a r t patterns. Conlputen und lndlrstriul Erigir~eerin,y, 51. 726-742. Gauri. S. I<.. 4 Chakraborly, S. (2008). Improved recognition of control c h a r t p a l l c r n s using artificial neural netwol-Its. I ~ ~ t e n ~ n t i o r ~ c i l J o u n ~ u l ofAdvnnted Manlrfuucturirlg Tei:hnolom,, 36, 1191--1201. Guh, R. S. (2007). On-line identification and q u a n t i l i c a t ~ o n of mean shifts In b i v a r ~ a t e processes u s ~ n g a NN-bared approach. Quality and Reliabilrtj~ Engineering I ~ ~ t e r n o t i o n a l . 23, 367 -- 385. Cull, R. S . (2010). Simultaneous process mean and variance monitol-ing using artiiicial neul-al nctworlts. Cornptrters rind Ir~dlistrifll Er~girirering, 58, 739-753. Guh. K, S., 8 Shiue, Y. K. (2005). On-line idenlification of contl-ul c h a r t p a t t e r n s using selt:o~-ganizing approaches, lnternatioiiu! lour1101 of Production Kesenrch, 43(6). 1225--- 1254. H a c h ~ c h a , W., 4 Ghorbel. A. (2012). A survey of contl-ol-chart pattern-~'ecognitio~i I ~ t e r a t u r e (1991-2010) based oli a new conceptual clasrificat~on s c h e ~ n e . C(~nlputers o r ~ d Incll~striol Eli~@l~ee~-in,y, 63, 204-222. Hassan, A., Nabi Baksli, M. S., S h a h a r o u n , M. A,, K. Jamaludin, H. (2001). Impl-ovecl SPC c h a r t paltern recognition using statislical features. 111Li.i-norionoljouii~al o j Producrion Rescorch, 41(7). 1587-1 603. 1.lotelling. 1-3. H. (1947). Multival-late quality contl'ol. In C. Eisenhdrt, M. W . Hastay, 8 W. A. W a l l ~ s (Eds.), Teclli~iques ofstub.sticul ailnlysi.~. New York: McGraw-I-1111. Jacltson. J. E. (1391). A user's guide to principle compollerrt. NL'W Jel-spy: Wiley. IKlosgen, W.. 8 Zytltow, J. M. (2002). HaildbooI< of dotti minii~g arid knowledge discovery. Lontlon: OxFord University I'ress. I