Gaussian Process Regression and Bayesian Inference Based

May 3, 2018 - As seen from eq 12, the predictive distribution is also Gaussian, and the mean vector and covariance matrix are σ–2x new TA–1XTy an...
0 downloads 3 Views 570KB Size
Subscriber access provided by the Library Service | Stellenbosch University

Process Systems Engineering

Gaussian Process Regression and Bayesian Inference Based Operating Performance Assessment for Multiphase Batch Processes Yan Liu, Xiaojun Wang, Fuli Wang, and Furong Gao Ind. Eng. Chem. Res., Just Accepted Manuscript • DOI: 10.1021/acs.iecr.8b00234 • Publication Date (Web): 03 May 2018 Downloaded from http://pubs.acs.org on May 5, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



Gaussian Process Regression and Bayesian Inference Based



Operating Performance Assessment for Multiphase Batch Processes



Yan Liu a, b, c*, Xiaojun Wang d, Fuli Wang a, b,1Furong Gao c



(a. College of Information Science & Engineering, Northeastern University, Shenyang, Liaoning, 110819, China;



b. State Key Laboratory of Synthetical Automation for Process Industries (Northeastern University), Shenyang,



Liaoning, 110819, China;



c. Department of Chemical and Biological Engineering, The Hong Kong University of Science and Technology,



Clear Water Bay, Kowloon, Hong Kong;



d. Key Laboratory of Advanced Design and Intelligent Computing, Ministry of Education, Dalian University,

10 

Dalian, 116622, China)

11 

Abstract: Batch processes have been playing a significant role in modern industrial

12 

processes. However, even if the operating conditions are normal, the process

13 

operating performance may still deteriorate away from optimal level, and this may

14 

reduce the benefits of production, so it is crucial to develop an effective operating

15 

performance assessment method for batch processes. In this study, a novel operating

16 

performance assessment method of batch processes is proposed based on both

17 

Gaussian process regression (GPR) and Bayesian inference. It is committed to solving

18 

the challenges of multiphase, process dynamics and batch-to-batch uncertainty that

19 

contain in most of batch processes. To characterize different dynamic relationships

20 

within each individual phase, multiple localized GPR-based assessment models are

21 

built firstly. Furthermore, the phase attribution of each new sample is determined, and

22 

two different identification results are obtained, i.e., a certain interval and a fuzzy

23 

interval between two adjacent phases. Then different online assessment strategies are

24 

designed correspondingly. When the operating performance is nonoptimal, cause                                                               *

 Corresponding author at: College of Information Science & Engineering, Northeastern University,

3 Lane 11, Wenhua Road, Heping District, Shenyang, Liaoning, 110819, China. Tel.: +86 024 83680351; Fax: +86 024 23890912. Email:[email protected]  1   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60



variables are identified by variable contributions. Finally, the effectiveness of the



proposed method is demonstrated by the fed-batch penicillin fermentation process.



Keywords: Batch processes, operating performance assessment, Gaussian process



regression, Bayesian inference, nonoptimal cause identification



1. Introduction



Batch processes are commonly used in chemical, materials, pharmaceutical,



biotechnology and semiconductor industries for the production of low-volume while



high-value-added commodities, and they play a more and more important role in



modern industrial production.1,

2

However, as time goes on, process operating

10 

performance may deteriorate away from optimal state due to process disturbances,

11 

noise, and other uncertainties, and this may cancel the benefits of preliminary designs

12 

for process optimization and result in a degraded operating behavior. Therefore, it is

13 

very necessary to develop an effective operating performance assessment method for

14 

batch processes.

15 

In the past several decades, many process and quality monitoring approaches have

16 

been proposed for batch processes.3-7 Among them, multiway principal component

17 

analysis (MPCA) and multiway partial least squares (MPLS) are most widely used. 6-9

18 

They extend the applications of principal component analysis (PCA) and partial least

19 

squares (PLS) techniques from continuous processes to batch processes, and allow

20 

process variable trajectory information to be projected into low-dimensional latent

21 

variable spaces. Therefore, batch process performance and product quality can be

22 

easily analyzed and monitored in the reduced space. As the study moving on,

23 

extensions to taking into account various factors are available, such as dynamic

24 

characteristics, 10, 11 nonlinearity, 12, 13 non-Gaussian distributions, 14, 15 and multiscale

25 

16

26 

MPCA/MPLS model cannot efficiently reveal the multiphase data behavior, even

27 

though it is very common in many batch processes. The phases defined by Undey and

28 

Cinar

29 

caused by operational or phenomenological (chemical reactions, microbial activities,

of batch processes. However, it has recently been explored that the traditional

17

is that ‘‘Steps occurring in a single processing unit as succession of events

2   

ACS Paragon Plus Environment

Page 2 of 37

Page 3 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

18



etc.) regimes’’. Furthermore, Lu et al.

defined phases from the perspective of



process monitoring, where phases are determined according to the changes in the



underlying process correlations. In the present work, we apply the phase definition



provided by Lu et al.18 Being different from stages that lay their focuses on describing



the physical operation units of a batch process, phases put the emphasis on the



expression of process characteristics, such as variable relationships, data distributions,



dynamic trajectories, amplitudes of measurements and so on. Compared with the



stages, although the phases are more abstract, it helps learn more about the intrinsic



process characteristics and enhance process understanding. Therefore, a batch process

10 

is desired to be divided into several phases to improve the modeling performance. In

11 

recent years, different phase division methods have been proposed, 19-21 and different

12 

modeling methods have been developed that take the phase effects into consideration.

13 

11, 22, 23

14 

work.

The pioneer work has provided abundant theoretical bases for our following

15 

In actual processes, the main task of process monitoring is to maintain the

16 

production process under normal operating conditions, but it can’t satisfy the quest by

17 

enterprises for profits any longer. For most of plants, the production goal is to profit,

18 

and an effective way is to ensure that the process operates on optimal level throughout

19 

the batch production. By this point, the operating performance assessment of

20 

industrial processes came into being. 24 The purpose of process operating performance

21 

assessment is to get a measure on how far the current operating condition is from the

22 

optimum (or how optimal the current operating state is), and the potential assumption

23 

is that the operating conditions are normal. According to the process characteristics

24 

and plant personnel’s attitudes of the operating performance, the performance levels

25 

can be divided into several grades, such as optimal, suboptimal, general, and poor.

26 

Through operating performance assessment, operators and managers can make a

27 

deeper understanding and mastering with the process operating performance, and

28 

propose reference suggestions on the operating adjustment and performance

29 

improvement. Despite a rich body of literatures in process monitoring, 25-27 studies in

30 

operating performance assessment of industrial processes are still in its infancy. In our 3   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60



previous work, some methods in respect to continuous and multimode continuous



processes have been developed on this issue. 28-30



This paper is devoted to the operating performance assessment for multiphase batch



processes with the considerations of process dynamics and batch-to-batch uncertainty.



It is generally believed that the process operating performance has a close relationship



with the comprehensive economic index, such as cost, profit, total revenue, product



quality or the weighted integration of several production indices. If the comprehensive



economic index approaches to or reaches the history optimal level, the process



operating performance is usually considered as optimality. Thus, the comprehensive

10 

economic index is applicable to evaluate the process operating performance. However,

11 

it is hard to get online and usually obtained at the end of a batch production, which

12 

seriously affects the timeliness of online assessment. As an alternative way, the

13 

predicted value of the comprehensive economic index can be used in operating

14 

performance assessment. Many quality prediction methods of batch processes have

15 

been proposed so far.

16 

processes, the quality prediction methods depend on the phase division strategies to

17 

some extent. For instance, the multivariate statistics based quality predictions are

18 

often applied to the batch processes whose phases are divided according to the

19 

variable relationships or covariance structures.

20 

characterized by non-stationary, batch-to-batch uncertainty and process variables

21 

frequently comprise deterministic trends, the process data of the whole batch often

22 

present non-Gaussian distribution. When the process variables have different change

23 

rates or amplitudes in two adjacent phases and the change rates are not very large in

24 

each individual phase, the local process data will show different distributions in a

25 

cycle of the batch production. In this sense, the batch processes can be divided into

26 

multiple phases based on the distribution characteristics, and each individual phase is

27 

actually represented by an approximate multivariate Gaussian distribution of the

28 

unfolded process data. This is quite consistent with the methodology of Gaussian

29 

mixture model (GMM), and some offline phase division algorithms based on GMM

30 

have been established, such as MPCA-GMM, 34 multiway Gaussian mixture model

31, 32

Considering the multiphase characteristics of batch

33

Since batch processes are usually

4   

ACS Paragon Plus Environment

Page 4 of 37

Page 5 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

35

and GMM-based phase successive division (GMM-PSD). 36 Under the



(MGMM)



framework of probability, Gaussian process regression (GPR)



appropriate quality prediction method matching GMM. Compared with the



deterministic modeling methods, GPR is more suitable for characterizing the complex



relationships between the process and quality variables caused by process stochastic



feature and system uncertainty.

37

is regarded as an



In the present study, with the consideration of process dynamics and batch-to-batch



uncertainty, a new operating performance assessment method based on GPR and



Bayesian inference is proposed for multiphase batch processes. Considering the

10 

advantages of GMM-PSD in dealing with the uneven-length batch processes, it is

11 

used in offline phase division firstly. Then multiple local GPR-based assessment

12 

models are developed to characterize different dynamic relationships of the identified

13 

multiple phases, and the predictive distribution of the comprehensive economic index

14 

is estimated by averaging and weighting all the regression model parameter values

15 

with their posterior probabilities. This kind of assessment models can not only

16 

effectively handle the stochastic feature caused by process dynamics and

17 

batch-to-batch uncertainty but also ensure the accuracy of the assessment result. In

18 

online assessment, based on the result of offline phase division, the phase type can be

19 

identified as a certain interval or a fuzzy interval between two adjacent phases. In a

20 

certain interval, the phase type of the new sample can be uniquely identified, and only

21 

the corresponding assessment model is invoked for online assessment. While, if the

22 

new sample falls into a fuzzy internal between two adjacent phases, the assessment

23 

result is the weighted sum of those from two adjacent phases with the Bayesian

24 

inference-based posterior probabilities as the adaptive weights. It effectively reduces

25 

the error rate caused by misclassification and is very challenging for traditional

26 

assessment methods. When the process operating performance degenerates, the

27 

possible cause variables can be identified based on variable contributions, which helps

28 

managers and operators take appropriate operating adjustment strategy on production

29 

improvement. The contributions of the proposed method are summarized as follows: (i) to ensure

30 

5   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 37



the accuracy of the assessment result, the GPR-based assessment models are



developed to deal with process dynamics and batch-to-batch uncertainty; (ii) for the



new sample whose phase type is unknown, its posterior probabilities with respect to



the assessment models of two adjacent phases are set as the adaptive weights, and



then integrate the corresponding local assessment results for online operating



performance



misclassification; (iii) the cause variables responsible for the nonoptimal operating



performance can be determined by variable contributions.

assessment,

which

avoids

the

error

evaluation

caused

by



The remainder of this article is organized as follows. Section 2 briefly reviews the

10 

methodologies of GMM and GPR. Then the proposed operating performance

11 

assessment method is introduced in Section 3. Section 4 demonstrates the

12 

effectiveness of the proposed assessment approach by the fed-batch penicillin

13 

fermentation process. Finally, some conclusions are drawn in Section 5.

14 

2. Preliminaries

15 

2.1 Gaussian mixture model

16 

GMM is usually used to model a collection of random variables arising from a

17 

number of latent classes or states.38 Consider a J dimensional random variable

18 

x  R J 1 , the GMM with M components is written as:

19 

G ( x Θ)   ωm g ( x θm ),

M

(1)

m 1

20 

where ωm , m  1, 2,..., M is the prior probabilities of the m th Gaussian component Cm

21 

and satisfies the conditions of 0  ωm  1 and

M

ω m 1

m

 1 ; θm   μm , Σ m  and

22 

Θ = ω1 , ω2 ,..., ωM , θ1 , θ2 ,..., θM  separately represent the local and global Gaussian

23 

model parameters; μm and Σ m are the mean vector and covariance matrix of the m th

24 

Gaussian component, respectively; g ( x θm ) is the corresponding multivariate

25 

Gaussian density function and expressed as follows:

6   

ACS Paragon Plus Environment

Page 7 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

g ( x θm ) 



1 (2π )

J 2

Σm

12

 1  exp   ( x - μm )T Σ m1 ( x - μm )  , m  1, 2,..., M .  2 

(2)



In order to establish the GMM, the unknown model parameters Θ need to be



estimated. Some learning methods, such as maximum likelihood estimation (MLE),



expectation maximization (EM), and Figueiredo–Jain (F–J) algorithm, are usually



used in parameter estimation of mixture model.



algorithm



and estimating their statistical distribution parameters, it is adopted for model



parameters estimation in this study. With the initialized model parameters



Θ(0) = ω1(0) , ω2(0) ,..., ωM(0) , θ1(0) , θ2(0) ,..., θM(0)  , the s th two-step iteration involved in the

10 

40

39, 40

In view of the ability of F-J

in automatically optimizing the number of Gaussian model components

F-J algorithm is as follows: E-step:

11 

P ( s ) (Cm xn ) 

12 

ωm( s ) g ( xn μm( s ) , Σ (ms ) ) M

ω m 1

(s) m

(s) m

(3)

.

(s) m

g ( xn μ , Σ )

M-step:

13 

N

( s 1) m

μ

14 



P n 1 N

(s)

P

(Cm xn ) xn ,

(s)

n 1

N

Σ (ms 1) 

15 

P n 1

(s)

ωm( s 1)

(Cm xn )

(Cm xn )( xn  μm( s 1) )( xn  μm( s 1) )T N

P n 1

16 

(4)

, (s)

(5)

(Cm xn )

 N  max 0,  P ( s ) (Cm xn )  V 2   n 1  ,  M N   max 0,  P ( s ) (Cm xn )  V 2   m 1  n 1 

(6)

17 

where xn is the n th sample of the modeling data X =  x1 , x2 ,.., x N   R N  J ;

18 

V  J 2 2  3J 2 is the number of scalar parameters specifying each Gaussian

19 

component; μm( s 1) , Σ(ms 1) , and ωm( s 1) are the mean vector, covariance matrix, and prior

T

7   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 37



probability of the m th Gaussian component at the ( s  1) th iteration, respectively.



The F-J algorithm is implemented in an iterative way until all parameters converge to



the optimal solution. 40 For an arbitrary new sample xnew , the posterior probability corresponding to the m

4  5 

th Gaussian component can be represented as follows:

P(Cm xnew ) 



ωm g ( xnew μm , Σ m ) M

ω m 1

m

.

(7)

g ( xnew μm , Σ m )



For a class of multiphase batch processes, where process variables have different



change rates or amplitudes in two adjacent phases and the change rates are not very



large in each individual phase, the phases can be characterized by their distribution

10 

functions. Although some variables show particular trends throughout the batch

11 

production, these trends can be significantly weakened in each individual phase. Thus,

12 

it can be considered that the unfolded process data from each individual phase

13 

approximately follow a multivariate Gaussian distribution, and GMM is applicable to

14 

cluster the batch process data from different phases.

15 

2.2 Gaussian process regression

16 

In many industrial processes, process variables often show stochastic feature owing

17 

to process dynamics and batch-to-batch uncertainty. Thus, the deterministic modeling

18 

strategies become ill-suited for characterizing the complex relationships between

19 

process variables and comprehensive economic index. To deal with this issue, GPR 37

20 

is proposed. GPR focus its interest in two aspects: (i) making inferences about the

21 

relationship between process inputs and outputs and (ii) obtaining the conditional

22 

distribution of the outputs given the inputs, rather than the multivariate distribution of

23 

process inputs. Denote the comprehensive economic index data corresponding to X

24 

as

25 

y =  y1 , y2 ,.., yN   R N 1 . Assuming that the relationships between process variables

26 

and comprehensive economic index are approximately linear, and then the regression

27 

model between X and y can be described as below:

T

8   

ACS Paragon Plus Environment

Page 9 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



yn  f ( xn )    xnT α   ,



where α  R J 1 is the regression parameter vector and follows a Gaussian prior



distribution with zero mean and covariance  ;  is the Gaussian noise with zero



mean and standard deviation  .37 Moreover, given the process inputs and regression model parameter, the conditional

5  6 

(8)

probability density function of the output is represented as follows 37: N

P ( y X ,  )   g ( yn x n ,  ) n 1

 ( yn  xnT  ) 2  1  exp    2 2 2 n 1   1 2  1  exp   2 y  X   , 1 2  2  (2 ) N 2  2 I N



(9)

  ( X , 2 I ) 8 

where I  R N  N is a unit matrix.



According to the posterior distribution over the regression model parameter, the

10 

posterior probability density function can be estimated through Bayesian inference

11 

strategy as follows:

P ( X , y ) 

12 

13 

P( y X ,  ) P( ) P( y X ,  ) P( ) ,  P( y X )    ( y X , ) ( ) P P d 

(10)

and the corresponding distribution can be expressed as P( X , y )   ( , A1 ),

14  15 

where A   2 X T X  Σ α1 ;    2 A1 X T y and

16 

covariance matrix, respectively.

(11)

A1 are the mean vector and

17 

In deterministic modeling strategies, a single parameter is typically chosen by some

18 

criteria to predict the output of the new sample xnew . However, under the consideration

19 

of process stochastic feature, all possible regression model parameters are weighted

20 

with their posterior probabilities. Thus, the predictive distribution of output

21 

yˆnew  f ( xnew ) is deduced as follows: 9   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

P ( yˆ new xnew , X , y )   P( yˆ new xnew , α )P(α X , y )dα



T T   ( 2 xnew A1 X T y, xnew A1 xnew ).

(12)

As seen from Eq. (12), the predictive distribution is also Gaussian, and the mean

2  3 

Page 10 of 37

T T A1 X T y and xnew A1 xnew , respectively. vector and covariance matrix are  2 xnew

Naturally, the predicted output of xnew is given by

4  5 

T yˆ new   2 xnew A1 X T y.



3. Gaussian process regression and Bayesian inference based



operating performance assessment of multiphase batch processes

(13)



Due to the dynamic characteristics and random uncertainty, 19, 20, 22, 41 the stochastic



assessment modeling method is more suitable for multiphase batch processes. In this

10 

study, a novel operating performance assessment method for multiphase batch

11 

processes is developed by integrating GPR with probabilistic inference strategy. The

12 

GMM-PSD algorithm is first used to divide different phases of batch processes.

13 

Furthermore, the GPR-based assessment models are established to characterize

14 

different dynamic relationships of the identified multiple phases, meanwhile, reveal

15 

the influences of different underlying process information on the process operating

16 

performance. During online assessment, the phase type of the new sample is identified

17 

firstly. If the new sample falls into the certain interval of a phase, only the assessment

18 

model of the corresponding phase is invoked for online assessment. While, if it

19 

belongs to a fuzzy internal between two adjacent phases, the assessment result can be

20 

calculated by incorporating those from two adjacent phases, where the Bayesian

21 

inference-based posterior probabilities are used as the adaptive weights. In this way,

22 

the error evaluation owing to misclassification is avoided effectively. When the

23 

process operating performance degenerates, the possible cause variables responsible

24 

for the nonoptimality can be identify based on variable contributions.

25 

3.1 Establishment of assessment models

26 

For a multiphase batch process, I uneven-length training batches are collected

27 

from the historical data. The process data of the i th batch are expressed in the form 10   

ACS Paragon Plus Environment

Page 11 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

T



of X i   xi ,1 , xi ,2 ,..., xi , Ki  , where Ki is the number of samples of the i th batch,



i  1, 2,..., I . The corresponding comprehensive economic index measured at the end



of the batch is denoted as yi . In order to remove the process noise and dynamics to



some extent, all training batches are unfolded into a two-dimensional block matrix



X  [ X (Tk 1) , X (Tk  2) , , X (Tk  Kmin ) , , X (Tk  Kmax ) ]T  R K × J via variable-wise unfolding as



shown in Fig. 1, where K min  min( Ki ) and K max  max( Ki ) are the minimum and



maximum numbers of samples of the training batches, K   Ki . Then the

1i  I

1i  I

I

i 1



normalization approach proposed in Refs. 42 and 43 is applied to X , i.e., xk ,i , j 



xk ,i , j  x j sj

(14)

,

10 

where xk ,i , j is the j th variable of the k th sample in the i th batch, x j and s j are mean

11 

and standard deviation of the j th variable of X .

X ( k=1) X ( k=2)

X Î R K ×J

X ( k = Kmin )

X ( k = Kmax )

12  13 

Fig.1. Illustration of the variable-wise unfolding of uneven-length training batches

Since the sample numbers of the comprehensive economic index and process

14 

11   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 37



variables must be the same in modeling, the comprehensive economic index of each



batch is extended to a vector by stacking it Ki times repeatedly, and then yi is



extended as yi   yi , yi ,..., yi   RKi 1 . Furthermore, the comprehensive economic index



data of all training batches are unfolded as y  [ y(Tk 1) , y(Tk 2) ,, y(Tk Kmin ) ,, y(Tk Kmax ) ]T  RK×1 .



For simplicity, the normalized forms are still denoted as X and y , respectively.

T



In multiphase batch processes, if process variables have different change rates or



amplitudes in two adjacent phases and the change rates are not very large in each



individual phase, the unfolded process data approximately follow a multivariate



Gaussian distribution in each individual phase. Hence, GMM-PSD strategy

36

is

10 

adopted for offline phase division. Firstly, with the unfolded data X , GMM is

11 

estimated through the F-J algorithm as formulated in Eqs. (3)~(6). Then the means of

12 

the posterior probabilities of the first h samples of each training batch are calculated

13 

with respect to each Gaussian component as follows: P (Cm xi , h ) 

14 

ωm g ( xi , h μm , Σ m ) M

ω m 1

m

,

g ( xi ,h μm , Σ m )

(15)

i  1, 2,..., I , m  1, 2,..., M , h  1, 2,..., h,

Pm 

15 

1 I h  P(Cm xi,h ), m  1, 2,..., M . Ih i 1 h1

(16)

16 

The target Gaussian component of phase 1 is determined by searching for the

17 

Gaussian component with the largest posterior probability, i.e., m  arg max( Pm ) .

18 

Thereafter, from the h  1 th sample of each training batch, only the posterior

19 

probability with respect to the target Gaussian component of phase 1, Cm , is

20 

calculated and used to determine whether the sample is still belonging to phase 1.

21 

Based on Bayesian inference, the posterior probabilities of the samples of phase 1

22 

should be greater than those with respect to other Gaussian components. Hence, if the

23 

posterior probability is greater than a given threshold, it can be sure that the sample

24 

belongs to phase 1. In this way, the end moments of phase 1 of each batch are

m

12   

ACS Paragon Plus Environment

Page 13 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



determined finally. Because the phase may be uneven-length across different training



batches, the end moments of phase 1 are different. As phase 1 is identified, the



process data belonging to phase 1 are removed from each batch, and then the



aforementioned procedures is applied to identify the second phase. This process is



conducted iteratively until all phases are divided. Correspondingly, y is divided into



different data sets along with X . The detailed steps of offline phase division by



GMM-PSD method can be referred in Ref. 36. Assuming that C phases are identified from offline phase division, the process

8  9 

data and comprehensive economic index data of phase c T

are denoted as

T

10 

X c   x1c , x2c ,..., x Nc c   R Nc  J

y c   y1c , y2c ,..., yNc c   R Nc 1 , c  1, 2,..., C

11 

respectively. Nc is the number of samples of phase c . Then the GPR-based

12 

assessment model of phase c is established and formulated as follows:

and

,

Nc

P ( y c X c ,  c )   g ( ync xnc ,  c ) n 1 Nc

 n 1

13 



 ( ync  xncT  c ) 2  1 exp    2( c ) 2 2 ( c ) 2   2  1 y c  X c c  exp   c 2  2( ) 

1 (2 ) Nc

2

( c ) 2 I

(17)

12

  ( X c c , ( c ) 2 I ), 14 

where  c is the regression parameter vector and follows a Gaussian prior distribution

15 

with zero mean and covariance Σcα ;  c is the standard deviation of the Gaussian noise. For a new sample xnew that belongs to phase c , the predictive distribution of

16 

c yˆ new  f ( xnew ) can be formulated as follows:

17 

c c P ( yˆ new xnew , X c , y c )   P ( yˆ new xnew , α c )P (α c X c , y c )dα c

18 

19 

  (( ) x c 2

T new

c 1

cT

c

(A ) X y , x

T new

c 1

( A ) xnew ),

(18)

where Ac  ( c ) 2 X cT X c  ( Σαc ) 1 .

20 

In modeling, some computations require inversions of covariance matrices.

21 

However, it doesn’t necessarily exist in real world scenarios for the collinearity of 13   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 37



process variables. In this case, the Moore–Penrose pseudo inverse 44 is calculated to



solve the inverse of a singular matrix.



3.2 Online operating performance assessment



In online assessment, the basic premise is to choose the correct active assessment



model, and this involves the issue of online phase identification for the new sample.



As all phases have been identified during the offline phase division, the start and end



moments of phase c of the i th batch are known and denoted as kic,in and kic,out ,



c ] respectively. Then the maximum interval of phase c is written as [kinc , kout



c accordingly, where kinc  min(kic,in ) and kout  max(kic,out ) are the earliest start and latest

10 

end moments of phase c . Since the uneven-length duration also presents in the single

11 

c 1 phase across different batches, kinc may be less than kout , c  2, 3,..., C  1 , and this

12 

c 1  . further results in an overlapping interval between two adjacent phases, i.e.,  kinc , kout

13 

c 1 , kinc 1 ) , the phase type of the new sample can In the nonoverlapping interval, i.e., (kout

14 

be identified uniformly, thus it is called as certain interval; while the phase type is

15 

unknown if a sample falls into the overlapping interval of two phases, and this kind of

16 

interval is described as fuzzy interval. According to the intervals of each phase, the

17 

phase type of the new sample can be determined. Table 1 gives a demonstration of the

18 

intervals and the corresponding phase identification results.

1 i  I

19 

Table 1 Intervals and online phase identification result Interval

Phase type

Phase identification result

[1, k ) [k , k ]

certain interval fuzzy interval

1 1 or 2





2 in

2 in

1 out



c 1 ( k out , kinc 1 )

fuzzy interval certain interval

c 1or c

 

 

 

C 1 C ( k out , k out ]

certain interval

C

c 1 [ kinc , k out ]

c

Supposing that the new sample belongs to phase c certainly, the assessment model

20  21 

1 i  I

of phase c is thus invoked, and the predicted output of xnew is given by 14   

ACS Paragon Plus Environment

Page 15 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

c T yˆ new  ( c ) 2 xnew ( Ac )1 X cT y c .



(19)



In order to simplify the assessment process, the predicted comprehensive economic



index should be normalized, and the normalized form is defined as the assessment



index and used for online operating performance assessment. Without loss of



generality, assuming that the operating performance is optimal when comprehensive



economic index is high, and the assessment index, γc , is calculated as follows:



c c  if yˆ new  ymax 1,  c c  yˆ  ymin c c c γc   new  yˆ new  ymax , if ymin , c c y y  min  max c c   ymin 0, if yˆ new 



c c where ymax  max( ync ) and ymin  min( ync ) , n  1, 2,..., Nc , are the maximum and



minimum values of the comprehensive economic index of phase c , respectively. As

10 

seen from Eq.(20), γc is between 0 and 1. When γc is close to 1, it means that the

11 

predicted value is close to the optimal value appeared in historical data, and the

12 

process operating performance is usually optimal; otherwise, if γc is nearly 0, the

13 

process operating performance is likely to be nonoptimal. In order to distinguish

14 

optimality from nonoptimality strictly, an assessment index threshold, (0.5    1) ,

15 

is introduced. If γc   , it means that the process operating performance is optimal. On

16 

the contrary, we can say that the process is operating on the nonoptimal performance

17 

grade. The value of  can be determined through historical data and expert

18 

experience. According to the historical data, the maximum and minimum values of the

19 

comprehensive economic index can be counted, and then experts who are familiar

20 

with the production process can give a dividing value between optimality and

21 

nonoptimality. Furthermore,  is calculated as in Eq. (20) by replacing the predicted

22 

value of the comprehensive economic index with the dividing value.

n

(20)

n

23 

c 1 ] , both the Particularly, if the new sample falls into the fuzzy interval [kinc , kout

24 

assessment indices γc 1 and γc are calculated, and Bayesian inference-based posterior 15   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 37



probabilities of xnew with respect to phase c  1 and c are set as the adaptive weights to



integrate the corresponding local assessment results. The posterior probabilities are calculated as follows:



P



c 1

ω c 1 g ( xnew μ c 1 , Σ c 1 )  c 1 , ω g ( xnew μ c 1 , Σ c 1 )  ω c g ( xnew μ c , Σ c )

(21)

ω c g ( xnew μ c , Σ c ) , μ c 1 , Σ c 1 )  ω c g ( xnew μ c , Σ c )

(22)

Pc 



ω c 1 g ( xnew



where ω c 1  ωc1 (ωc1  ωc ) and ω c  ωc (ωc 1  ωc ) represent the prior probabilities



with respect to phase c  1 and c , respectively. Furthermore, the assessment indices can be integrated as follows:



γ  P c 1γc 1  P c γc .



(23)

Then the process operating performance can be evaluated by comparing the

10  11 

assessment index γ with the threshold .

12 

3.3 Nonoptimal cause identification

13 

Although the nonoptimal operating performance is not expected, it may occur due

14 

to manual operating error or process characteristics drift in production. Therefore,

15 

when the process operating performance is nonoptimal, it is much important to

16 

identify the possible cause for actual production processes. Being similar to the

17 

contribution plots-based fault diagnosis methods,

18 

identification method is proposed based on variable contributions in this section. By

19 

approximately decomposing the assessment index into the weighted sum of different

20 

process variables, the variable contributions are constructed firstly. Then compare the

21 

contribution values with their statistical scopes obtained from the optimal level, and

22 

identify the responsible variables according to the given criterion. Since the

23 

contribution is a function of the process variable, the nonoptimal cause can be

24 

identified as long as the nonoptimality can be reflected through the measurable

25 

process variables. Therefore, the proposed method is applicable to nonoptimal

26 

conditions caused by sensors and process abnormalities.

45, 46

16   

ACS Paragon Plus Environment

the nonoptimal cause

Page 17 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



As known from Eq.(20), the variable contribution to the assessment index γc is



c c equivalent to the contribution to yˆ new , so yˆ new is further decomposed into the



following form: J

c T yˆ new  xnew z c   xnew, j z cj ,



(24)

j 1



where z c = ( c )2 ( Ac )1 X cT y c ; z cj is the j th element of z c , and xnew, j is the j th



variable of xnew . Accordingly, the raw contribution of the j th process variable with



respect to the assessment index γc is denoted as below: c c contrraw , j  xnew, j z j , j  1, 2,..., J .



(25)



c ˆc Intuitively, small contrraw , j causes small ynew , and further leads to small γc . So it

10 

looks like that the process variables with smaller raw contributions should be

11 

considered as the responsible variables for nonoptimality. However, even though

12 

under the optimal operating performance, the raw contributions of different process

13 

variables are not the same, and some of them may be very small or negative.

14 

Therefore, it is more reasonable to compare each raw variable contribution with its

15 

average level under optimal operating performance, rather than only considering its

16 

absolute value. In view of this, I batches under the optimal operating performance

17 

are selected from I training batches and used to calculate the mean of the raw

18 

contribution of each process variable. The reference data of each phase are denoted as

19 

( X 1 , y 1 ),( X 2 , y 2 ),...,( X C , y C ) , where X c  [ x1c , x 2c ,..., x Nc ]T and y c  [ y1c , y 2c ,... y Nc ]T ,

20 

c  1, 2,..., C .

21 

related parameters. The mean of the raw contribution of the j th process variable of

22 

phase c is calculated as follows:

c

 c ) 1 and A c  ( c ) 2 X cT X c  ( Σ α

z c = ( c ) 2 ( A c )1 X cT y c

c

are the

N c

c  c  c N c , contrmean , j   xn , j z j

23 

n 1

24 

where z cj is the j th element of z c , and xnc, j is the j th variable of x nc , n  1, 2,..., N c . 17   

ACS Paragon Plus Environment

(26)

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 37

Then we define the variable contribution as below:



c c contrjc  contrraw , j -contrmean , j , j  1, 2,..., J .



(27)



In theory, if contrjc  0 , it means that the j th variable is the responsible variable for



the nonoptimal operating performance. Nevertheless, it is difficult to achieve an



c c absolutely equal between contrraw , j and contrmean , j in actual processes. Hence, the



upper and lower confidence limits of the variable contributions of the reference data



are counted by kernel density estimation 47 for different phases, and they are denoted



c as CU cj and CLcj , respectively. If CLcj  contrjc  CU cj , it means that contrraw , j and



c contrmean , j are approximately equal statistically; otherwise, it can be concluded that

10 

they are obviously unequal. Furthermore, the process variables with contributions that

11 

are significantly less than CLcj are considered as the cause variables responsible for

12 

the nonoptimal operating performance. For the process variables whose contributions

13 

are greater than CU cj , they may play a role in improving the operating performance.

14 

c 1 ] , both contrjc 1 and contrjc should be If xnew belongs to the fuzzy interval [kinc , kout

15 

calculated, and the posterior probabilities are used as the adaptive weights to integrate

16 

them as the final variable contribution, i.e.,

17 

contrj  P c 1contrjc 1  P c contrjc , j  1, 2,..., J .

18 

Correspondingly, the upper and lower confidence limits are constructed as follows: CU j  P c 1CU cj 1  P cCU cj ,

19 

CL j  P c 1CLcj1  P c CLcj , j  1, 2,..., J ,

(28)

(29)

20 

The step-by-step procedure of GPR and Bayesian inference based operating

21 

performance assessment approach is summarized as below and the flow diagram is

22 

shown in Fig. 2.

23 

(1) Collect the training batches for the establishment of assessment models.

24 

(2) Unfold the three-dimension data into two-dimension matrix via variable-wise

25 

unfolding and perform the normalization on the unfolded matrix. (3) Estimate GMM and identify all different phases based on GMM-PSD strategy.

26 

18   

ACS Paragon Plus Environment

Page 19 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



(4) Develop GPR-based assessment models for different phases.



(5) Identify the phase type for new sample and invoke the corresponding model.



(6) Predict the comprehensive economic index of the new sample and construct the



assessment index with it.



(7) Online assessment using the constructed assessment index.



(8) Identify the cause variables responsible for the nonoptimal operating



performance based on variable contributions.

8  9 

Fig.2. Diagram of GPR and Bayesian inference based operating performance assessment approach

19   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60



4. Case study



4.1 Fed-batch penicillin fermentation process description

The fed-batch penicillin fermentation process



48, 49

is a typical multiphase batch



process and has received wide attentions for its academic and industrial importance.



There are two physical stages in this process: preculture and fed-batch. The preculture



stage starts with small amounts of biomass and substrate. Most of the initially added



substrate is consumed by the microorganisms after about 45 h, and the process is



switched from preculture stage to fed-batch stage which usually continues about 355 h.



In fed-batch stage, the penicillin increased exponentially until the stationary

10 

procedure. Because the biomass growth rate must be kept constant for the aim of

11 

optimizing penicillin production, the substrate is supplied continuously into the

12 

fermentor, rather than being added all at once in the beginning. Meanwhile, in order to

13 

maintain the constant temperature and pH values, two proportional-integral-derivative

14 

(PID) control loops are implemented in the fermentor for manipulating the acid/base

15 

and hot/cold water flow ratios, respectively. In 2002, a modular simulator (PenSim

16 

v2.0) for fed-batch fermentation was developed by Ali Çinar et al. from the

17 

Monitoring and Control Group of the Illinois Institute of Technology. It can simulate

18 

the concentrations of biomass, CO2, hydrogen ion, penicillin, carbon source, oxygen,

19 

and heat generation under various operating conditions. The diagram of the penicillin

20 

fermentation process is shown in Fig. 3.

21 

Known from the mechanism of penicillin fermentation process, the process

22 

operating performance depends on the final penicillin concentration. Under the

23 

normal operating conditions, the higher penicillin concentration corresponds to the

24 

better operating performance, while the poorer operating performance usually leads to

25 

lower penicillin concentration. Thus, the final penicillin concentration is used as the

26 

comprehensive economic index in this study. In addition, the process variables related

27 

to the penicillin concentration are selected for operating performance assessment and

28 

listed in Table 2. Table 3 gives the initial operating conditions of the production.

20   

ACS Paragon Plus Environment

Page 20 of 37

Page 21 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research

pH

FC Substrate Acid

Fermentor T

Base FC Cold Water Air

1  2 

Hot Water



Table 2 Process variables used for operating performance assessment

Fig. 3 Flow diagram of the penicillin fermentation process

No. 1 2 3 4 5 6 7 8 9 10 11 4 

Process variables

Normal operating ranges

Aeration rate (L/h) Agitator power (W) Substrate feed rate (L/h) Substrate feed temperature (K) Substrate concentration (g/L) Dissolved oxygen concentration (%) Biomass concentration (g/L) Culture volume (L) Carbon dioxide concentration (mole/L) pH Generated heat (kcal/h)

3-10 20-50 0.035-0.045 296-298 5-50 1.16 0-0.2 100-150 0.5-1 4-6 0

Table 3 Initial operating conditions of penicillin fermentation process Process variables

Initial operating conditions

Substrate concentration(g/L) Dissolved oxygen concentration (%) Biomass concentration (g/L) Penicillin concentration (g/L) Culture volume (L) Carbon dioxide concentration (mole/L) pH Fermentor temperature (K) Generated heat (kcal/h)

15 1.16 0.1 0 100 0.5 5 298 0

To establish the assessment models, a total of 50 training batches are produced



21   

ACS Paragon Plus Environment

Industrial & Engineering Chemistry Research 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60



under the normal operating conditions, and the durations are between 390h and 420h



with the sampling interval of 1h. By using GMM-PSD method, five phases are



identified from offline phase division. Table 4 shows the result of offline phase



division, and the preculture stage and fed-batch stage are further divided into two and



three phases, respectively. This can be attributed to the fact that a stage may contain



multiple data distributions caused by the intrinsic biochemical reactions or the



changes in external operating conditions. In preculture stage, most of the initially



added substrates are consumed quickly by the microorganisms. To capture the



regularity of data distribution, meanwhile, ensure that the data of each phase

10 

approximately follow a multivariate Gaussian distribution, the preculture stage is

11 

further divided into two phases, although its duration is only about 45 h. The

12 

fed-batch stage, whose duration is about 355 h and almost eight times length of

13 

preculture stage, is just divided into three phases. It is because that the variation trends

14 

of the variable trajectories noticeably slows down in this stage, and three phases are

15 

enough to describe the diversity of data distributions. In addition, the batch-to-batch

16 

variations cause the inconsistent phase durations across different batches, thus form

17 

the fuzzy intervals between phases.

18 

Table 4

Result of offline phase division

Phase no.

Earliest start moment kinc

Latest end c moment kout

1

1

34

2

30

47

3

43

125

4

106

223

5

195

420

Phase division result Certain interval of phase 1: [1,30) ; Fuzzy interval between phase 1 and 2: [30,34] ; Certain interval of phase 2: (34, 43) ; Fuzzy interval between phase 2 and 3: [43, 47] ; Certain interval of phase 3: (47,106) ; Fuzzy interval between phase 3 and 4: [106,125] ; Certain interval of phase 4: (125,195) ; Fuzzy interval between phase 4 and 5: [195,223] ; Certain interval of phase 5: (223, 420] .

19 

To test whether the process data of each phase follow Gaussian distributions, the

20 

multivariate normality test algorithm proposed in Ref. 50, i.e., F-Straight Method

21 

Based on Mahalanobis Distance (FSMD), is implemented before modeling. The basic

22 

idea of FSMD is that, when the process data approximately follow a multivariate 22   

ACS Paragon Plus Environment

Page 22 of 37

Page 23 of 37 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Industrial & Engineering Chemistry Research



Gaussian distribution, the distribution function of the squared Mahalanobis distances



of samples, D( n ) , is a specific F distribution, and it can be replaced by its empirical



distribution function. Denote the quantile of F distribution as Fn , and the above



argument is equivalent to test whether D( n )  Fn . Let the linear regression equation



between D( n ) and Fn be Fn  a  bD( n ) , the significance test is performed on it based on



the regression standard deviation, s , and the mean of the quantile, F . If



s F  0.15 , the significance test is valid and the linear regression equation can reflect



the real relationships between D( n ) and Fn ; otherwise, the significance test is invalid.



Furthermore, if both of the conditions a   and b  1   are satisfied, where  and

10 

 are usually set as F  5% and 0.2, it is usually considered that the fitting line,

11 

Fn  a  bD( n ) , passes through the original point with the slope of 1. Thus, the

12 

hypothesis test that the data follows a Gaussian distribution should not be rejected.

13 

Otherwise, reject the null hypothesis. In summary, when the conditions s F  0.15 ,

14 

a   and b  1   are satisfied at the same time, the data follow a Gaussian

15 

distribution. The detailed procedures of FSMD are given in Appendix A. As shown in

16 

Table 5 and Fig. 4, the test results illustrate that, in each individual phase, the unfold

17 

process data approximately follow a multivariate Gaussian distribution.

18 

Table 5 Results of multivariate normality test of each phase Phase no.

s / F (