24/05/2021

Quality Data Analysis

Multivariate control chart
Bianca Maria Colosimo – biancamaria.colosimo@polimi.it

Reference:
Quality Data Analysis- BM Colosimo
QualityMontgomery –
Data Analysis Chapter 11
1

Multivariate SPC: control chart for the mean

Assume to be interested in monitoring more than a single variable. We want to
simultaneously control all the quality characteristics. Let’s apply the known methods
for monitoring the variable means.

x1

x2

2
Quality Data Analysis

1
 24/05/2021

Multivariate SPC
Control limit computation:
P Allarme 1 In Controllo 1  
1

P Allarme 2 In Controllo 2   2

Assume variables to be independent and:
1   2  

P  A llarm e proc . in controllo   1  1   1   

Generally speaking, for p tests on independent variables:
   1  1    p   1  1    1 p
 p  2    0.0975 ARL  10
 We expected
A few examples:   0.05 p  3    0.143 ARL  7
 p  4    0.185 ARL=20
3  ARL  6
Quality Data Analysis

Multivariate SPC
The problem gets more complicated if variables are dependent.

In this case one has to use the Bonferroni’s inequality or to determine the joint 

probability density function.

 ' i
i  1,.., p Bonferroni’s inequality
“familywise” 

If  test i we choose i= /p

 ’  p /p=  being ’ the Type I error for the whole set of p tests

4
Quality Data Analysis

2
 24/05/2021

Multivariate random variables

Consider a p‐component vector, i.e. a vector of p random variables x '  [ x1 , x 2 ,..., x p ]

Expected value μ '  E ( x )  [ E ( x1 ), E ( x 2 ),..., E ( x p )]  [  1 ,  2 ,...,  p ]

 2  12 ... 1p 
Variance-Covariance Matrix  1 
  22 ... 2p
V ( x )  E [( x  E ( x ))( x  E ( x ))' ]  Σ   12 
 ... ... ... ... 
pxp px1 1xp  2p ...  2p 
 ij  E [( x i   i )( x j   j )]  cov( x i , x j )  
1p

Product between a constant vector and a random vector:

Scalar (1xp)(px1)V ( a ' x )  E [( a ' x  E ( a ' x ))( a ' x  E ( a ' x )) '] 
 E [( a ' x  a ' E ( x ))( a ' x  a ' E ( x )) ']  a ' E [( x  E ( x ))( a '( x  E ( x ))) '] 
 a ' E [( x  E ( x ))( x  E ( x )) ']a  a ' Σa
Quality Data Analysis 5

correlation  1  12 ... 1 p 
cov( x i x j )  1 ...  2 p 
P  
12
 ij 
V ( x i )V ( x j )  ... ... ... ... 
 
  1 p 2 p ... 1 
Correlation matrix
Bivariate nomal distribution:
2
1  x 
1   
2  
f ( x)  e -  x < + 
2  2

2
x 
 
1
   x    
2
x     x - μ  '  Σ  1  x - μ 
  
1
1   x - μ '  Σ -1  x - μ 
f (x)  e 2 -  x j < +  j = 1,2,..., p
 2 
p/2
| Σ |1 / 2
6
Quality Data Analysis

3
 24/05/2021

Multivariate normal distribution

Bivariate case: p=2

X Y
Y

X X

Y Y

Y Y
X X X X

Quality Data Analysis

7

Some relevant results

Remind:
X  μ
2

X ~ N ( μ, ) 2
 Z 2 ~ χ 2 (1)
 2

X  μ
2
 2
X ~ N (μ , ) n  Z 2 ~ χ 2 (1)
n  2

Multivariate case:
Has been
X i  N p (μ , Σ ) proven that (X  μ )'Σ  1 (X  μ )  Z 1 2 + Z 2 2 + ...+ Z p 2 ~  2 ( p )
( i  1, 2 , ..., n )
Has been proven that
X  N p ( μ , (1 / n ) Σ ) n (X  μ )'Σ  1 (X  μ )  Z 1 2 + Z 2 2 + ...+ Z p 2 ~  2 ( p )

1xp pxp px1

Quality Data Analysis 8

4
 24/05/2021

Case p=2
x '   x1 x2    12  12  1 1  2
2
 12 
Σ     Σ   
μ '  1 2    1 2  22  | Σ |    1 2  12 

By solving the matrix multiplications, n ( x  μ )'Σ  1 ( x  μ ) ~  2 ( p )

We got:
1   22   1 2   x1  μ 1 
n  x1  μ 1 x2  μ2  2 2  2    ~ 2 ( p)
 1  2   12 2   12  1   x2  μ2 
Thus, the statistical quantity:
n
 02    2 2 ( x1   1 ) 2  2 1 2 ( x1   1 )( x 2   2 )   1 2 ( x 2   2 ) 2 
 1  2   12
2 2 2

Follows a chi squared distribution with 2 degrees of freedom:

 02 ~  2 2 

Quality Data Analysis 9

Being known that:

n
 02    2 ( x   1 ) 2  2 12 ( x1   1 )( x 2   2 )   1 2 ( x 2   2 ) 2    2 ( 2)
 1  2   12 2  2 1
2 2

A joint control region is applicable:

Alarm if :  02  2 2 

2 2
 x1   1   x2   2   1 2  x1   1   x 2   2   1 2 2 2   1 2 2  2 ( 2 )
      2   
 1   2   1 2   1    2   1 2 2 2 n

10
Quality Data Analysis

10

5
 24/05/2021

Ellipse control region - equation

2 2
 x1   1   x2   2   1 2  x1   1   x 2   2   1 2 2 2   12 2  2 (2 )
      2   
 1   2   1 2   1    2   1 2 2 2 n
2 2
 12  0    0  x1   1   x2   2   2 (2)
When                            it becomes:      
 1   2  n
2 2
 x1   1   x2   2 
   
 1     2   1
 2 (2)  2 (2)
n n
1
Semi‐axes:  2 ( 2 ) su x1
n
2
 2 ( 2 ) su x 2
n


1   2  
If                     the ellipse is a circle with radius  2 ( 2 )
n

Quality Data Analysis

11

11

Ellipse control region - equation

2 2
 x1   1   x2   2   1 2  x1   1   x 2   2   1 2 2 2   1 2 2  2 ( 2 )
    2   
 1   2   1 2   1    2   1 2 2 2 n

 12  0    0  12  0    0

12
Quality Data Analysis

12

6
 24/05/2021

x1 Region C
UCL1
Region B

Region A

LCL1
x2
LCL2 UCL2
Region A: out-of-control not signaled by traditional charts
Region B: false out-of-control signaled
Region C: ‘twice’ false out-of-control (because of x1 and x2) signaled
Quality Data Analysis 13

13

2 control chart
A statistical control scheme can be implemented directly into the plane x1 - x2 by
using the control ellipse:
• But the information about the temporal sequence would be lost;
• It would be difficult to depict the control region for 3 variables, and even
impossible for larger numbers of variables

A control chart can be designed to monitor the quantity  0 by using the control
2

limit    p  . In the most general case:

2

scalare
1xp pxp px1

Quality Data Analysis 15

15

7
 24/05/2021

T2 control chart
Assume we don’t know      and      and we m samples of size n to be used in design 
phase

Estimators 
(considering m samples)

For each sample k (k=1,…,m):

 02k  n ( x k  μ )' Σ -1 ( x k  μ ) T k2  n ( x k  x )' S -1 ( x k  x )
1xp pxp px1

The T2 statistic is referred to as Hotelling’s statistic and it follows the F distribution (not the 

squared chi distribution), corrected by a constant

Quality Data Analysis

19

19

In particular:
- Phase 1 (design phase)
T k2  n ( x k  x )' S -1 ( x k  x ) ~ c 1 ( m , n , p ) F ( p , m ( n  1)  ( p  1))

p ( n  1)( m  1)
c1 ( m , n , p )   UCL  c 1 ( m , n , p ) F ( p , m ( n  1)  ( p  1))
m ( n  1 )  ( p  1)
LCL  0
‐ Phase 2 (future observations): 
Under the assumption that m* samples during the design phase

T k2  n ( x k  x )' S -1 ( x k  x ) ~ c 2 ( m *, n , p ) F ( p , m * ( n  1)  ( p  1))

p ( n  1)( m *  1)
c 2 ( m *, n , p )   UCL  c 2 ( m *, n , p ) F ( p , m * ( n  1)  ( p  1))
m * ( n  1)  ( p  1)
LCL  0

20
Quality Data Analysis Quality Engineering
20

8
 24/05/2021

It is possible to prove that, for large values of m:

c 1 ( m , n , p ) F ( p , m ( n  1)  ( p  1))   2 ( p )

Remind that:  2 ( 1 )
lim F(  1 ,  2 ) 
2 1

Quality Data Analysis 21

21

Example: ultimate 
tensile strenght and 
diameter of a textile 
fiber
n=10; m=20; p=2.

 02 
n
 12 2 2   12 2
 2
2
( x1  1 ) 2  2 12 ( x1  1 )( x2   2 )   1 ( x2   2 ) 2
2

n  S 2 ( x  x ) 2  2 S ( x  x )( x  x )  S 2 ( x  x ) 2 
p  2T2 
S1 S 2  S12 2  
2 2 2 1 1 12 1 1 2 2 1 2 2

22
Quality Data Analysis

22

9
 24/05/2021

No out‐of‐control data.

By using the Phase II limit advocated by Montgomery we have UCL=15.16.

If we used the 2 approximation we would get UCL=13.816, close to the Phase I limit, but 
different from the Phase II limit.

Quality Data Analysis 23

23

Problem:
When a control limit violation occurs, how to search for an assignable cause?

We could exploit univariate control charts (one for each variable), with Bonferroni’s control 

limits, i.e., by replacing Z  / 2 with Z  /( 2 p )

A different approach consists of decomposing the T2 statistic into components that reflect the 
contribution of each individual variable.
If T2 is the current value of the statistic, and  T2(i) is the value of the statistic for all process
variables except the ith one, then:
d i  T 2  T (2i ) is an indicator of the relative contribution of the ith variable.

Thus, when an out‐of‐control signal is generated, we can compute the values of  di for each

process variable and to search for assignable causes associated to the variables for which
are relatively large.

Quality Data Analysis 24

24

10
 24/05/2021

T2 control chart
Example:
Standardized vars:
x ij   j
y ij 
( m  1) 2j

Out‐of‐control

Cut‐off for di
Quality Data Analysis 25

25

Individual observations
In some SPC applications (e.g., chemical industry) the size of the sample is n=1. In 
this case, the Hotelling’s statistic is defined as follows:

The Phase II control limits are:

When the number of samples

(individuals) is very large (m>100) or
the following Phase II control
limits may be used:
Some authors suggest the use of Beta 
distribution for Phase I control limit
computation.
Quality Data Analysis 26

26

11
 24/05/2021

Individual observations
A relevant issue in the presence of individual observations is the variance‐covariance 
matrix        estimation

The ‘usual’ estimator (long period) is:

This estimator is particularly sensitive to outliers or out‐of‐control data in the original 
sample of m observations

Alternative:

A different estimator (short period) is:

27
Quality Data Analysis

27

Example (Sullivan e  Individual observations

Woodall – 1996)
Composition: 

L=percentage classified as large

M=percentage classified as
medium
S=percentage classified as small
Control chart just on the first 
two components
(sum=100%) 

x   5 . 682 ,88 . 22 

Quality Data Analysis 28

28

12
 24/05/2021

Individual observations
Example (Sullivan e Woodall – 1996)

x   5 . 682 ,88 . 22 

Quality Data Analysis 29

29

Latent Structure methods (briefly)

Problem: The control‐charting schemes previously presented 
become less and less effective as the number of monitoring 
variables (p) increases
(satisfactory performances for p<=10).

Example: MEWMA control chart
The ARL value when the process is out‐of‐control increases as p
increases

We are interested in different control‐charting schemes, that are 
suitable to monitor processes where the variability is not equally 
influenced by all the p variables (the sistem dynamics is driven by 
a subset of process variables)
30
Quality Data Analysis

30

13
 24/05/2021

Principal Components – PCA

Principal Components Analysis
Given a set of p process variables, the goal is to perform a reference system
transformation:

The key idea consists of finding a new orthonormal

reference system that allows maximizing the 
variability of original data and, simultaneously, 
reducing the number of variables that are necessary to describe the process

z1: largest percentage

of data variability

Quality Data Analysis 31

31

Principal Components
Example: (chemical process)        95% confidence region
Scatter plot matrix

(actually, it is the sample correlation matrix)

32
Quality Data Analysis

32

14
 24/05/2021

Latent Structure methods

Example: 
“Principal Components Analysis” command (Minitab)      

Principal components scores for first 20

observations:

Quality Data Analysis 33

33

Latent Structure methods

Scatter plot z1- z2 (with 95% confidence
region: used as control region)

34
Quality Data Analysis

34

15
 24/05/2021

Out‐of‐control

Quality Data Analysis 35

35

Latent Structure methods

Drawbacks: 
1. A very large number of variables z could be computationally infeasible (the 
scatter plots matrix is too huge)
2. Difficult interpretation (out‐of‐controls on transformed variables that are 
linear combinations of original ones)
3. Promising approach: generate the r‐variate vector of retained principal 
components and apply the MEWMA on it
4. Shifts on directions that are different from the ones associated with the 
retained principal components can be difficult to detect (especially if 
orthogonal directions are concerned): in those cases one could use a MEWMA 
control chart on the remaining principal components zr+1,…,zp.

If r=2, it is not possible to detect shifts

orthogonal to this plane

36
Quality Data Analysis

36

16