ICOTS 2, 1986: Tadakazu Okuno

STATISTICAL QUALITY CONTROL I N JAPANESE INDUSTRY AND EDUCATION PROGRAMS FOR ENGINEERS Tadakazu Okuno Dean, Faculty of Engineering Science University of Tokyo Japan

1 Introduction
Development of quality control Modern o r Statistical Quality Control (SQC) took the f i r s t form i n 1930s i n the U.S. when W. A. Shewhart of Bell Telephone Laboratory applied t h e Control Chart invented b y him t o American industries. The modern statistical theories originated i n the U.K. and the application of them t o SQC was attempted as early as 1935. I n Japan, the study of the modern statistics was engaged before the War b y a limited number of researchers. The U.S. Armed Forces stationed i n Japan after the War were concerned with the low and erratic quality characteristics of Japanese communication i n struments and facilities. From around 1950, SQC utilizing the Control Chart and the Sampling I n spection came t o be used i n many Japanese plants after the line of D r . W. E. Deming. QC was f i r s t engaged only b y engineers and factory workers. The original concept of QC was such t h a t the conformity of quality could be assured through inspection, which is referred t o as the Inspectionoriented QC.

If goods of non-conformity result continually from the production process, a r i g i d inspection process would become uneconomic. It is t o be preferred t o reduce the number of goods of non-conformity b y means of eliminating negative factors i n the production process. This stage of QC is called t h e Process-control-oriented QC.
Problems concerning reliability, safety and economy of products as well as of failing design and u n f i t materials cannot be solved only b y the QC e f forts of the production department. These problems are t o be dealt with b y a comprehensive'QC containing the phases of development, plan and design of new products. This stage is called the Design-oriented QC. Features of Japanese quality control The present Japanese QC are represented i n the following six items. Recently, some enterprises i n other countries follow the similar lines of QC. 1) Company-Wide Quality Control (CWQC), which indicates the QC with all the departments and employees of a corporation participated;

2) QC with emphasis on education and training f o r all the employees from t o p managers t o workers;

ICOTS 2, 1986: Tadakazu Okuno

3) Q u a l i t y Control w i t h QC c i r c l e activities; Q u a l i t y C o n t r o l w i t h QC audits, by t h e p r e s i d e n t o f t h e concerned c o r poration a n d by t h e Deming P r i z e Committee, whose members a r e composed o f academic a n d research specialists f r o m n o n - p r o f i t o r g a n i z a tions;

5) QC w i t h e f f e c t i v e u t i l i z a t i o n o f s t a t i s t i c a l methods;
6) Promotion o f QC a c t i v i t i e s by t h e w a y o f nation-wide movements.
In t h i s paper we shall deal w i t h Items 2) a n d 5) a n d i l l u s t r a t e several cases i n automotive i n d u s t r i e s .

2. Q u a l i t y C o n t r o l Education Scheme
V e r y f e w educational programmes have e v e r been devoted t o r e a r i n g QC e x p e r t s i n Japan. Presently, consultation specialists i n t h e QC have h a d t h e educational b a c k g r o u n d s o f chemical, mechanical, electric, e l e c t r o n i c o r i n d u s t r i a l engineering. O n l y a small number o f them have majored i n mathematics o r mathematical statistics. T a b l e 1. Main Education Courses for 1986 of JUSE
Courses [ C in general1 P 1 Cowrehensive Basic Course . 2. Elenentan Basic Course 3. PC TOPWsmgemnt Course 4. P Executives Course C 5. PC Middle Waaageaent C w n e 6. Basic Course for Forenen 7. "KC Instructor Course 8. PC-Circle Instructor Course 9. PC-Circle Leader Course Cjwcific Topies1 1 Desiw of Experimnt (DE) . 2. Eleaentarr Course in DE 3. Elenentan Multivariate Analysis 4. Mvancd Hultinviata h a l n i s 5. Rslinbilib 6. Elcnentlw Rsliability 7. lhma.llant Course in Reliability 8. FWA-FfA 9. Desiw Reviev 10. Owrations Research 11. Wsrketing Research 12. Senson Inswction Tem(tf kl./t days h w n Fm./m

30 8

189

4
3 15

54 24
18 103

4 4
2 3 30 17 1 1

28 24
13 18 190 110

67

100 140 1W 90 100 100 100 130 100 50 90 60

2 7
3 1

3 7 3 11 8 1 1 1

A number o f QC education courses a r e h e l d e v e r y y e a r for s t a f f members o f all r a n k s o f e n t e r p r i s e s by i n s t i t u t i o n s l i k e JUSE, JSA etc. A s an example, t h e education courses f o r 1986 o f JUSE a r e shown i n T a b l e 1.

ICOTS 2, 1986: Tadakazu Okuno

Table 2. Curriculum f o r Design o f Experiment Course

Subfecc
0 . Preli#&ury

No. of Units* Couraa (2)

2

~ubjecc

Ib. of Unit#

S c a c i s c i u l Hachods 1 krcroduccion (4) . b n i c Cancapc Factor cnd Lnnrl

02 fn Production

5. Techniquaa f o r Plandug d k s i g r t i n g Produccs ( l l f

D f o r designing E 1 1 1 quality Ihplopenc syac.n

WFrA

3 3

3
2 2

Proc~a Use of D E
2. Design m d AMlysis of Experinmcs (22)

ku1y.t. of Process Cilplbfllcy 6. Specific Topics (9)

Signal-co-Noise

&Kin

Yulci-way L y o u c s Split-ploc Iksignr Use of Orchogotul Arrays Caapo.tce Oasigns 2

j
'
;

W c i v a r i a c e ANOVA

1 1 1 1
2

9
2

Encoaplece Block lkaiqn 2 MElySi. of Qcagorical Lbca 3. U8e of Suppltmencary Variablas (3) Regression Analysis
4. Theory (3)

/
I
I

;
!

us. of Kfcro-compucu
k u l y ~ Of ~ SeMOv i Impeccion b c a k p l o r a c o r y OIra Aarlyaia
Practical Exnnplea

Spacinl h c c u r e Exarcisas and Others (11) OEGurva (Group Caapecicioo) Exarcisas Croup D b c w ~ a i o n s Concluding U v r k a

j 7.
3

2.5

Linear S c a c i s c i d Enfarmce

5 3
0.5

r A unit dmoces 3-hour Laccure andlor praccica.

Table 3. Curriculum f o r Elementary and Advanced Multivariate Analysis (MA) Courses

Elementary MA course Subject What is MA Simple regression Multiple regression (MR) ( p * = 2 ) Principal Component Analysis (PCA) Selection of variables in MR Computer Program Package Quantification Method I Exploratory Data Analysis Discussion No. of Units 1 1 1 1 1 0.5 1 1 0.5 Advanced MA course Subject Multiple regression and PCA Discriminant Analysis Cluster Analysis Multi-dimensional scaling and correspondense analysis Outlier detection and regression diagnostics Use of Micro-computer and Large-scale computer (SAS) Discussion 0.5 1 No. of Units

1
1

0.5

*p

denotes number of explanatory variables.

Specific curricula f o r the courses, "Design of Experiment" and "Multivariate Analysis", are presented i n Tables 2 and 3. These educational p r o grammes tend t o illustrate various statistical techniques, which could be applied t o respective occasions of industrial activities. It is, however, t o be favored t h a t those techniques are incorporated into a whole set of analyses of the existing procedures f o r locating ill-functioning processes. A following step is t o modify o r t o relieve the failed processes, i n which t h e statistical techniques could be an important tool f o r implementation.

ICOTS 2, 1986: Tadakazu Okuno

Fig. 1 Problem-solving paths

effect

effect

Approach

effect

good^
effect
I

bad effect
I
1

mod
effect
I

effect
t
I

effect
I

* _ _ _ _ ______, *
1

: : Cause-effect d i m n : I_~di-~.t_~.!!!_"l -=_rpp kst~ - ; t !_ _? - i J

Exclusion
?P;?P;?P;?P;_?P;?P;

,---_--__-__A_______-----.

) b l a r e the

very

cause

:~-~tatistica~_a_"a~_y_~~i~~t

Fig. 1 will help t o explain the problem-solving paths i n the QC Education Scheme. I n Approach 1 t h e goods of non-conformity (bad effect) resulting from a failed process (ill cause) are excluded o r adjusted i n an additional process t o make satisfactory goods. The approach does not deal with t h e v e r y cause so t h a t t h e first-hand production of non-conformity goods continues. The cost of production cannot be reduced. I n Approach 2 t h e cause-effect diagram is introduced t o l i s t plausible causes. Every measure p u t f o r t h by all QC-circle members is undertaken for improving products. This approach works normally t o the desired effect, b u t it provides no i n formation on t h e whereabouts of failed processes. A series of excessive measures is v e r y often taken u p consequently, which leads t o an added cost of products. I n Approach 3 e v e r y failed process can be located through an effective use of the multiple regression analysis on existing data and the design of e x periment. This approach can indicate the processes which cause nonconformity goods. It will be able t o lead t o a complete elimination of failed processes and also t o reduce non-conformity goods b y a great degree.

3. Utilization of Multivariate Analysis Techniques

The author and f i v e mechanical engineers have recently published a book entitled "Application of Multivariate Techniques i n Automotive I n d u s t r y " . Those coauthors w o r k f o r Toyota Motors and related companies. Eight cases from the production processes are dealt with i n the publication. Ref e r r i n g t o t h e book, the author illustrates examples of the utilization of multiple regression analysis and related techniques.

ICOTS 2, 1986: Tadakazu Okuno

Case 1. Estimating optimum p a i n t i n g conditions o f t h e b o d y o f a motor-car

Fig. 2. Cause-Effect Diagram for Painting Efficiency

properties o f Paints Ambient Conditions

' 7 Specltlc viecoeity

Height

I.)+
tlu~uldl ty D i l u t i o n Rate
( ~ 1 0 )

d Speed Painting

Speed of Movement of Spray Gun tiuntber o f P a i n t i n g e

Pressure o f Air Spray

(xs)
Spread of Sprayed P a i n t i ( ~ o Spray nethods

T h e p a i n t i n g e f f i c i e n c y y ( = amount o f p a i n t s coated/amount o f p a i n t s s p r a y e d ) i s influenced by t h e q u a l i t y a n d q u a n t i t y o f p a i n t s sprayed, s p r a y methods a n d ambient conditions as shown i n Fig. 2. We have f o u n d a model f o r estimating y by f o u r variables - temperature o f p a i n t s x8, d i s tance o f s p r a y g u n xq, spread o f s p r a y e d p a i n t s X6 a n d s p r a y e d selected f r o m 10 variables by u s i n g n = 38 observations. amount x 7 T a b l e 4 shows several statistics f o r t h e b e s t seven subsets w i t h each specif i e d number o f variables. T h e "best" means t h e smallest values o f Residual Sum o f Squares (RSS) a n d t h e seven subsets a r e a r r a n g e d in t h e i n creasing o r d e r o f RSS values. Prediction Sum o f Squares (PSS) a n d RSS a r e d e f i n e d by

w h e r e fi a n d Pi* (i=1 ,2,. . ,n ) denote, respectively, t h e r e g r e s s i o n estimates d e r i v e d f r o m n observations a n d those f r o m n-1 observations e x c l u d i n g t h e i-th observation. T h e a u t h o r normally recommends a d o p t i n g t h e subset w i t h t h e minimum value o f PSS.

ICOTS 2, 1986: Tadakazu Okuno

Table 4. Statistics for t h e "bestn Seven Subsets with Each Number of Variables
Nunbar

of

Varlablna

P-vsluas for tsytlng,nlgnlfleanoa of

The l e f t half of Table 4 is f o r PSS, RSS and squared multiple correlation R2, i t s adjusted value b y degrees o f freedom Rf2 and i t s doubly-adjusted value R"' derived from the expected value of PSS. The minimum value of PSS is shown 1924.93 i n the subset with four variables mentioned above. I n the r i g h t half of Table 4 are indicated the F-values f o r testing significance of the partial regression coefficients. 10- denotes t h a t the F-value is larger than 10. The model becomes explicitly,

qi reveals an outlier,

The scatter diagram of 38 residuals yi - qi and regression estimates which is excluded i n the calculations t o follow.

The similar analyses were also conducted on the other two criterion v a r i ables related t o thickness of coating. It is found t h a t the models f o r these two variables can be expressed b y the same four explanatory variables as t h e painting efficiency. Case 2. Analysis of gear noises of transmission Sensory tests on gear noises of transmission b y workers were conducted f o r discrimi nati ng two groups "Conformity and non-conformity" This sort of tests is laborious and the reproducibility is open t o question. I n t h i s study, measurements of sound pressure at 28 frequencies were used f o r calculating stepwise discriminant functions and 5 frequencies a t 40Hz, 160Hz, 630Hz, 2.5KHz and 10.OKHz were chosen as the most appropriate frequencies.

ICOTS 2, 1986: Tadakazu Okuno

Fig. 3. Group Means of Sound Pressure at 28 Frequencies and Their Coefficients in t h e Discriminant Function
Frequency Variable
o
0,.
1,

(Hz)
31.5
10

Group Means
I*r.-lt.
-t.-.

a , , - (

Change in Sound P r e s s u r e . with F r e q ~ e n c i e s ~ , ~ ~ , - -!. ., I ~ I 1

5
11 1

6
3211 533
SJI~

51)

61 IN1 n ltll 11 125 01. IWI I . 21~1 gI , 2511 .XU 315 1,. .I1111 .r,, 5INl 01,. 61UI .rsr # # I x,. 1 . n. t.Sk XI. I.W)k rt. 2l~lk O.,, 'LYlk rrt 3.15k m 4.lll k ran 5.lXtk 12. G.3llk m H.ILIk r. r.
1

xn
xu

I'Lllk I6.lXlk

16117 19117 26.21 3230 3.15 211.ai 1.188 1.15.1 I . IOJNI 7.74 I~IAll 2.2 1 2. 21 I!L25 .1 X V 1748 2l.4ti IRHX 111112 H.4l 51 .7 235 11.23 IIS'Z lCUl

-~.GI 591 -1511 81 15 . -11112 -1 2 1. 3 163 -0.70 19.77 112IUi 5.5 53 ILZ5 3211 -il.Gl Zi.76 l8lil 067 143 11.3 0.76 13.711 I -1.1.1 1.17 11.17 -234 lUli -2.47 1087 -11.17 ZLSB ! I ! -11.7l !.! -1.07 4.4 -271 211.~1 -207 2353 1.77 211.65 1.73 11.75 Ifil 111115 13!1 GXG J-'= -11-3'' ZG7 p , -1l.13 D31i I 1 1 1192 -11.l5 01 ;5

C 3

Is

25

J I

O -Z N I.I

fllXXi8
- O X

0.UlI.l 01XXlR

11a1 1x3 o.tiz fl.18 11.1~l

OU117 fllN68 0.lYXII -011171

(122 1.19 11117 I-%

0.IXXiI 01XXl 01X115 OJXXFJ

! I

IUI 0.1 1 11.23

L I. Jl )i

1 !

~I J IZ X ? OJXlli! U.IXX1)( fl.lX~11~ -1J:li l OJlf l i - OJXIBI

W . I 11 15 ; 1.5 11 1.5 11 11.72 2!liW !!.!Xi Il.'B)

'

1 ) Mark o designates the v a r i a b l e s s e l e c t e d by stepwise discriminant a n a l y s i s

2 ) C o e f f i c i e n t of each variable i n the discriminant function

using 28 v a r i a b l e s .

3) $-value is used f o r t e s t i n g t h e s i g n i f i c a n c e of the c o e f f i c i e n t ,

t [71:0.[15) =?.OO. f ( 7 1 :[).Of) =Z.tiG

Fig. 3 shows the means of the two groups and their difference f o r each of 28 variables as well as the coefficients of the 28 variables i n the discriminant function with corresponding t-values f o r testing their significance. It is interesting t o note t h a t a t the above five frequencies both the differences of two-group means and the t-values are distinctly larger than those at other frequencies. This result has led t o the design of an instrument f o r discrimination as shown i n Fig. 4.

Fig. 4. An Instrument for Discirninating Gear Noises of Transmission

r l l r u c tny..pllflar
c.lcul.~l,op

ICOTS 2, 1986: Tadakazu Okuno

Case 3. Analysis of deficit i n diecasting work

Table 5. Data with Category Numbers of 6 ltems (factors)

Noce: The bracketed obsecvacions have che same comblnacion of cacepories. The diEEerences of these data are used i n escinacing che variance of pure error.

The fractional-factorial design with 6 factors each at 3 levels was originally planned with the resulted specific weight data (coded) representing t h e degrees of deficit as shown i n Table 5 b u t a few combinations of levels were proved t o be d i f f i c u l t t o realize. The resulting 53 combinations do not guarantee the orthogonality of t h e six factors. The comprehensive analysis of such complicated and incomplete data is conducted on computers b y using "Quantification theory of t h e f i r s t type" of Hayashi, which can be considered as the "Multiple regression analysis for categorical data", o r "Analysis of multi-way classification data with unequal numbers of observations i n cells".

ICOTS 2, 1986: Tadakazu Okuno

Table 6. An Illustration o f Computer Output by Haga (Quantification Method o f t h e First Type)

Item(factor) No. of Category 1 (A) corresponding t-value Partial

T h e computer o u t p u t is i l l u s t r a t e d i n Table 6. T h e score bi. represents t h e e f f e c t o f t h e j - t h category (level) o f t h e i - t h item (factorj, subject t o t h e condition: f o r each i, where n i j denotes t h e number o f observations i n t h e concerned categ o r y . I n t h e middle t h r e e columns f o r t h e differences in t h e scores and t h e i r t-values, t h e lower off-diagonal elements f o r item A indicate t h e differences i n scores l i k e b l l - b12 = 3.185 - (-2.956) = 6.141. T h e corresponding t-values are g i v e n i n t h e u p p e r off-diagonal elements.
Fig. 5. Change of PSS and RN2with Pdumber of Items

Number of Itens

Selection o f item i n place o f selection o f each dummy variable is performed t h r o u g h t h e algorithm newly-developed by T. Haga o f Science U n i v e r s i t y o f Tokyo. T h e progressive results i n t h e intermediate steps are summarized i n Fig. 5. T h e model accommodating o n l y t w o items, temperature o f

ICOTS 2, 1986: Tadakazu Okuno

melted aluminum (factor B) and speed o f moulding (factor A), give the best subset b y the criteria RI2, R W 2 and PSS.

is proved t o

The above illustrations display clearly t h e characteristic points of o u r data analysis, which are summarized as follows:

1) Unbalanced data often derive from failed experimental designs. T h e y can, however, be analysed b y so-called "quantification methods", which are multiple regression and discriminant analyses f o r categorical data. 2) The model (or variable) selection i n multiple regression analysis and detection of outliers are not connected normally with unique solutions. This is the characteristic o f "Exploratory Data Analysis (EDA)".
3) The application of EDA will promote t h e development of SQC. The cases illustrated could serve as examples of methods t o be adopted i n such studies.

4. Concludina Remarks
The application of statistical methods should not be left exclusively f o r QC experts. Extension of the use of statistical techniques t o those who concern t h e quality is required since data are acquired i n their normal line of work. However, their understanding on the significance of data acquisition seems t h e least developed. Collecting data i n the daily work could be decisive i n reducing non-conformity goods. It i s generally recommended t h a t the collectors of data be supplied with 5W1H on their work: Who, When, Where, What, Why and How. The educational programmes should include simple methods of graphical display and data analysis on the personal computer f o r general engineers. The advanced statistical analysis b y use of program packages such as SAS on large-scale computers might be reserved i n the educational programme f o r a f a i r l y limited number of engineers.