Abstract
Low-emittance (low-E) glass manufacturing has become an important sector of the glass industry for energy efficiency of such glasses. However, the quality control scheme in the current processes is rather primitive and advanced statistical quality control methods need to be developed. As the first attempt for this purpose, this article considers monitoring of optical profiles, which are typical quality measurements in low-E glass manufacturing. A Bayesian hierarchical approach is proposed for modeling the optical profiles, which conducts model selection and estimation in an integrated framework. The effectiveness of the proposed approach is validated in a numerical study, and its use in Phase I analysis of optical profiles is demonstrated in a case study. The proposed approach will lay a foundation for quality control and variation reduction in low-E glass manufacturing.
Appendices
Appendix A:. The Gibbs sampling procedure in Section 3
Starting step: Specify starting values
A simple method to find starting values is fitting a regular polynomial model to each profile. The coefficient estimates will be used as starting values of β, the sample variances of those estimates will be used as starting values of {σ2p, …, σ02}, and the estimate of random error variance will be used as the starting value of σ2ϵ. The random effects can simply be set to zero.
Generate a sample {α(g)i, i = 1, …, m} from the conditional posterior in EquationEquation (9)
(9) .
Generate a sample {σ2(g)p, …, σ02(g)} from the conditional posterior in EquationEquation (10)
(10) .
Generate a sample
from the conditional posterior in EquationEquation (11)
(11) .
Then go back to Step 1 and repeat this process for g = 1, … , G.
Appendix B: Proof of EquationEquation (9)
(9)
We will first prove a proposition and then prove EquationEquation (9)(9) based on that.
Proposition A1. Let U1, U2, … , Un be n independently normally distributed random variables with a common mean and scaled variances; that is,
where c1, … , cn are known non-zero constants, σ2 is known, and θ is the unknown mean that needs to be estimated. Let u1, u2, … , un be the corresponding observations of these variables. Under a normal prior for θ; that is,
the posterior is
where
Proof. Under the given prior, the posterior of θ is
which is a normal distribution with mean and variance
The result in the proposition can be obtained by simple manipulations of the above formulas.▪
Coming back to the HLME model, by EquationEquation (3)(3) :
which can also be written as
(A1) Let zi = [zi1, … , zin]′, xk = [xk1, …, xnk]′, and ϵi = [ϵi1, … , ϵin]′, by Equations (A1) and (1),
or, equivalently,
We can treat zij/xkj, j = 1, … , n, as n independently normally distributed variables with a common mean αi,k and scaled variances σ2ϵ/(xjk)2, which falls into the exact situation assumed in Proposition A1. Also, by EquationEquation (2)
(2) , the prior of the common mean αi,k is a normal distribution
According to Proposition A1, the corresponding posterior is
By expressing the above formula in matrix form, EquationEquation (9)
(9) will be obtained.▪
Appendix C: The model selection procedure in Section 4
For each given model Mi:
Generate posterior samples
by the procedure in Appendix A.
Estimate the log marginal likelihood
by EquationEquation (15)
(15)
2.1. Calculate the mean, θ*, of the posterior sample of θ.
2.2. Calculate the log-prior
by EquationEquation (17)
(17) .
2.3. Calculate the log-prior
by EquationEquation (18)
(18) and treat it as
.
2.4. Calculate the log posterior ordinate
by EquationEquation (21)
(21) .
2.5. Calculate
by EquationEquation (15)
(15) using the results in Sections 4.1 and 4.2.
Calculate the log BF of Mi vs. Mi-1 by EquationEquation (14)
(14) .
If the log BF is larger than three, conclude that Mi is better than Mj.
Appendix D: Proof of EquationEquation (20)
(20)
By the independence of parameters (β,σ2ϵ) and {σ2p, …, σ02},
(A2) So to find the joint conditional posterior at the left-hand side, we just need to find the three marginal conditional posteriors at the right. Under the uniform prior in EquationEquation (6)
(6) :
under the uniform prior in EquationEquation (3)
(3) :
and under the weakly informative prior in EquationEquation (7)
(7) :
Plugging in these distributions into EquationEquation (A2)
(A2) will lead to EquationEquation (20)
(20) .▪
Additional information
Notes on contributors
Li Zeng
Li Zeng is an Assistant Professor in the Department of Industrial and Manufacturing Systems Engineering at the University of Texas at Arlington. She received her B.S. degree in Precision Instruments and M.S. degree in Optical Engineering from Tsinghua University and Ph.D. in Industrial Engineering and M.S. degree in Statistics from the University of Wisconsin–Madison. Her research interest is process monitoring and control in complex manufacturing and healthcare delivery systems. She is a member of INFORMS and IIE.
Nan Chen
Nan Chen is an Assistant Professor in the Department of Industrial and Systems Engineering at the National University of Singapore. He obtained his B.S. degree in Automation from Tsinghua University and his M.S. degree in Computer Science, M.S. degree in Statistics, and Ph.D. degree in Industrial Engineering from the University of Wisconsin–Madison. His research interests include data analytics in manufacturing systems, prognostics and health management, data-driven modeling, and control of service systems. He is a member of INFORMS, IIE, and IEEE.