Multivariate Control
Charts
(Ref. Ch. 11, SQC, D.C. Montgomery)
Prof. Sayak Roychowdhury
Joint Control Region
The joint control
region is NOT the shaded
region. The joint control
Region is distorted here,
𝑃 𝑇𝑦𝑝𝑒 𝐼 𝐸𝑟𝑟𝑜𝑟 = 1 − 1 − 𝛼 𝑝 ,
Where p is the number of variables
Multivariate Normal Distribution
𝒙 = [𝑥1 , … . 𝑥𝑝 ] is a random vector with multivariate normal
distribution.
Let 𝝁 = [𝜇1 , … , 𝜇𝑝 ] denote the mean
Let Σ denote 𝑝 × 𝑝 dimensional variance-covariance matrix
The multivariate normal probability density function
1
1 − 𝒙−𝝁 𝑇 Σ−1 (𝒙−𝝁)
𝑓 𝑥1 , … 𝑥𝑝 = 𝑝 1
exp 2
2𝜋 2 Σ 2
𝐷 = 𝒙 − 𝝁 𝑇 Σ −1 𝒙 − 𝝁 is also known as Mahalanobis
Distance, distance from point 𝒙 from distribution
𝐌𝐕𝐍(𝝁, 𝚺)
Joint Control Region
Joint control region Joint control region
When 𝑥1 and 𝑥2 are When 𝑥1 and 𝑥2 are
independent not independent
Hotelling’s 2
𝑇 Control Chart
• Two quality characteristics 𝑋1 and 𝑋2 jointly distributed as
bivariate normal distribution
• Target mean values are 𝑋ത1 and 𝑋ത2
• Sample means 𝑋ത1 and 𝑋ത2
• Sample variances 𝑠12 and 𝑠22
• Test statistic
𝑛 2 2
2
𝑇 = 𝑠22 𝑋ത1 − 𝑋ത1 + 𝑠12 𝑋ത2 − 𝑋ത2 − 2𝑠12 𝑋ത1 − 𝑋ത1 𝑋ത2 − 𝑋ത2
𝑠12 𝑠22 −𝑠12
2
2 2
• Compare 𝑇 2 with 𝑇𝛼,2,(𝑛−1) , if 𝑇 2 > 𝑇𝛼,2,(𝑛−1) at least one
of the characteristics is out of control
𝑇 2 Hotelling Control Chart
• The Hotelling 𝑇 2 Control Chart uses the following test
statistic
𝑻 −𝟏
𝑇2 ഥ
ഥ − 𝒙 𝑺 (ഥ
=𝑛 𝒙 ഥ)
𝒙−𝒙
Source: Introduction to statistical quality control,
D.C.Montgomery
Control Limits
Individual Contribution of Characteristics
• 𝑇 2 chart does not identify specific variables responsible
for shifts
• 𝑇 2 decomposition method is used to determine the
individual contributions of each of the p variables to the
statistic
2
• 𝐷𝑖 = 𝑇 2 − 𝑇1,2,…𝑖,𝑖+1,..𝑝 for 𝑖 = 1,2, . . 𝑝
2
• 𝑇1,2,…𝑖,𝑖+1,..𝑝 denotes the 𝑇 2 statistic when 𝑖 𝑡ℎ variable is
left out
Issues with 𝑇 2 Charts
• Difficult to identify which variable has shifted
• Difficult to interpret out of control signals
• It is directionally invariant chart, it can only reveal the
shift from mean vector in terms of magnitude, not in
direction
MEWMA Charts
• Multivariate EWMA charts is the logical extension of
single variable EWMA
Here 0 ≤ 𝜆 ≤ 1 and 𝒁𝟎 = 𝟎
The quantity plotted on control chart is given by
Control Limits MEWMA Charts
• The control limits are derived based on shift size 𝛿, and
ARL performance. The control limit H is given for In-
control 𝐴𝑅𝐿0 = 200 (Table 11.3), and
𝐴𝑅𝐿0 = 500, 1000 (Table 11.4)
Other Control Charts
• Regression Adjusted Control Charts
• PCA Based control charts
