A Modified Dynamic PLS for Quality Related Monitoring of Fractionation Processes

Proceedings, 10th IFAC International Symposium on Shenyang, Liaoning, July 25-27, 2018 Advanced Control of China, Chemical Processes Proceedings, 10th IFAC International Symposium on Shenyang, Liaoning, China, July 25-27, 2018 online Available at www.sciencedirect.com Advanced Control Chemical Processes Proceedings, 10th of IFAC International Symposium on Shenyang, Liaoning, China, July 25-27, 2018 Advanced Control of Chemical Processes Shenyang, Liaoning, China, July 25-27, 2018

A Modified Dynamic PLS for Quality Related Monitoring of Fractionation ScienceDirect A Modified Dynamic PLS Related A Modified Dynamic PLS for for Quality Quality Related Monitoring Monitoring of of Fractionation Fractionation Processes IFAC PapersOnLineProcesses 51-18 (2018) 315–320 A Modified Dynamic PLS for Quality Related Monitoring of Fractionation Processes Xue Xu Qiang Liu Jinliang Ding A Modified Dynamic PLS for Related Monitoring of Fractionation XueQuality XuProcesses Qiang  Liu Jinliang Ding XueQuality Xu Qiang Ding A Modified Dynamic PLS for Related Monitoring of Fractionation  Liu Jinliang Processes  Liu Jinliang Ding Xue Xu Qiang State Key of Laboratory of Synthetical Automation for Process Industries, Northeastern University, Shenyang Processes 

State Key of Laboratory of Synthetical Automation for Process Industries, Northeastern University, Shenyang 110819,Liu China Xue XuAutomation Qiang State Key of Laboratory of Synthetical forJinliang Process Ding Industries, Northeastern University, Shenyang 110819, China  Xue Xu Qiang Liu Jinliang 110819, China State Key of Laboratory of Synthetical Automation for Process Ding Industries, Northeastern University, Shenyang  110819, China [email protected]; [email protected]; State Key(e-mail: of Laboratory of Synthetical Automation for Process [email protected]). Industries, Northeastern University, Shenyang [email protected]; [email protected]; [email protected]). State Key(e-mail: of Laboratory of Synthetical Automation for Process Industries, Northeastern University, Shenyang 110819, China (e-mail: [email protected]; [email protected]; [email protected]). 110819, China (e-mail: [email protected]; [email protected]; Abstract: The fractionation process is a typical dynamic process,[email protected]). and practitioners highly pay attention Abstract: The fractionation process is a typical dynamic process, and highlydynamic pay attention to the quality-related abnormal in the real refining processes. In this practitioners paper, a modified PLS (e-mail: [email protected]; [email protected]; [email protected]). Abstract: The fractionation process is a typical dynamic process, and practitioners highly pay attention to the quality-related abnormal in the real refining processes. In this paper, a modified dynamic PLS (MDPLS) modeling method and the corresponding process monitoring strategy are proposed. The main (e-mail: [email protected]; [email protected]; [email protected]). to the quality-related abnormal in theis real refining processes. In and this practitioners paper, a modified PLS Abstract: The fractionation process a typical dynamic highlydynamic pay attention (MDPLS) modeling the corresponding processprocess, monitoring strategy are proposed. main contributions of the method proposedand method are in the following. First, a clear dynamic relation isThe captured (MDPLS) modeling method and thethe corresponding process monitoring strategy are proposed. The main to the quality-related abnormal in real refining processes. In this paper, a modified dynamic PLS contributions offractionation the method in thedynamic following. First,and a clear dynamic relation is attention captured Abstract: The process is are a typical process, practitioners highly pay between process dataproposed and quality indices. Moreover, the process quality space are comprehensively contributions of the proposed method are in the following. First, and a clear dynamic relation isThe captured (MDPLS) modeling method and the corresponding process monitoring strategy are proposed. main between process data and quality Moreover, the process space comprehensively Abstract: The fractionation process a typical dynamic process, and practitioners highly pay to the quality-related abnormal inindices. theis real refining processes. Inand thisquality paper, a modified PLS divided into dynamic quality-related subspace, static quality-unrelated subspace asare well asdynamic theattention residual between process dataproposed and quality indices. Moreover, the process and quality space are comprehensively contributions of the method are inrefining thestatic following. First, a clear dynamic relation isThe captured divided into dynamic quality-related subspace, quality-unrelated subspace as proposed. well asdynamic the residual to the quality-related abnormal in the real processes. In this paper, a modified PLS (MDPLS) modeling method and the corresponding process monitoring strategy are main space for improving the performance of monitoring. Finally, the effectiveness of the proposed algorithm divided into dynamic quality-related subspace, static the quality-unrelated subspace as well as the residual between datamethod and quality indices. Moreover, process and quality space comprehensively space forprocess improving the performance ofare monitoring. Finally, the effectiveness ofare theare proposed algorithm (MDPLS) modeling and the corresponding process monitoring strategy proposed. The main contributions of the proposed method in the following. First, a clear dynamic relation is captured is demonstrated with the data from a real fractionation process. space forinto improving the performance subspace, of monitoring. Finally, the effectiveness of the proposed algorithm divided dynamic quality-related quality-unrelated subspace asare well as the residual is demonstrated the data from a realare fractionation process. contributions of with the method in thestatic following. First,and a clear dynamic relation is captured between process dataproposed and quality indices. Moreover, the process quality space comprehensively is demonstrated with the data from a real fractionation process. space for improving the performance of monitoring. Finally, the effectiveness of the proposed algorithm © 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: Fractionation process, Dynamic modeling, Quality related monitoring between process data and quality indices. Moreover, process and quality space comprehensively divided into dynamic quality-related subspace, static the quality-unrelated subspace asare well as the residual Keywords: Fractionation process, modeling, Quality related monitoring is demonstrated with the the data fromDynamic a real fractionation process. divided dynamic quality-related subspace, static quality-unrelated subspace as well as thealgorithm residual space forinto improving performance of monitoring. Finally, the effectiveness of the proposed Keywords: Fractionation process, Dynamic modeling, Quality related monitoring  space for improving of monitoring. the effectiveness of the activity proposedinalgorithm is demonstrated with the the performance data fromDynamic a real fractionation process.  Finally, Keywords: Fractionation process, modeling, Quality related monitoring property, decay of catalyst reactor, operation  1. INTRODUCTION is demonstrated with the data from a real fractionation process. property, decay of catalyst activity in reactor, operation transform, Alghazzawi A et al. [17] proposed recursive INTRODUCTION Keywords:1.Fractionation process, Dynamic modeling, Quality property,related decaymonitoring of catalyst activity in reactor,a operation  transform, Alghazzawi A et al.strategy [17] proposed a recursive 1.cracking INTRODUCTION multi-block PCA monitoring which can quickly Fluid catalysis and unit (FCCU) is an important Keywords: Fractionation process, Dynamic modeling, Quality related monitoring transform, Alghazzawi A et activity al.strategy [17] in proposed a operation recursive property, decay catalyst reactor, multi-block PCAofmonitoring which can quickly Fluid catalysis and 1.cracking unit (FCCU) is more an important  identify and isolate abnormality in the crude distillation unit. process for converting heavy oil fractions into valuable INTRODUCTION PCA monitoring strategy which can quickly Fluid catalysis and cracking unit (FCCU) is an important multi-block transform, Alghazzawi A et al. [17] proposed a recursive identify and isolate abnormality in the crude distillation unit. process for converting heavy oil fractions into more valuable property, decay of catalyst activity in reactor, operation  In the real industry, auto-correlation and lag cross-correlation light production in petroleum processing. Fractionation and isolate abnormalitystrategy in the crude distillation unit. process for converting heavy oil fractions intois more valuable identify INTRODUCTION multi-block PCA monitoring which can quickly Fluid catalysis andin1.cracking unit (FCCU) an important In the real industry, auto-correlation and lag cross-correlation light production petroleum processing. Fractionation property, decay of catalyst activity in reactor,a variables. operation transform, Alghazzawi A et large-scale al. [17] proposed recursive or the dynamics exist among process columnproduction is an indispensable unit that produces Fractionation gasoline and In real auto-correlation andcrude lag cross-correlation light in1. INTRODUCTION petroleum processing. andindustry, isolate abnormality in[17] the distillation unit. process for converting heavyunit oil fractions intois more valuable or dynamics exist among process column is inan indispensable that produces gasoline and identify transform, Alghazzawi A proposed et large-scale al.strategy proposed a variables. recursive multi-block PCA monitoring which can quickly Fluid catalysis and cracking unit (FCCU) an important Therefore, Gao et al. [18] indiscernibility dynamic diesel oil fluid catalysis and cracking processes. The two dynamics exist among large-scale process variables. column is an indispensable unit that produces Fractionation gasoline and or In the real industry, auto-correlation and lag cross-correlation light production in petroleum processing. Therefore, Gao et al. [18] proposed indiscernibility diesel oil oil inconverting fluid catalysis and cracking processes. The twoa multi-block PCA monitoring strategy which can dynamic quickly identify andGao isolate abnormality in the crude distillation unit. Fluid cracking unit (FCCU) an important process heavy oil fractions intoisEndpoint more valuable kernel PCA, where two indiscernibility and kinds catalysis of areand major products of FCCU. is et al. [18]parameters--the proposed indiscernibility dynamic diesel oilfor inan fluid catalysis and cracking processes. The two or dynamics exist among large-scale process variables. column isoil indispensable unit that produces gasoline anda Therefore, kernel PCA, where two parameters--the indiscernibility and kinds of are major products of FCCU. Endpoint is identify and isolate abnormality in the crude distillation unit. In the real industry, auto-correlation and lag cross-correlation process for converting heavy oil fractions into more valuable light production in petroleum processing. Fractionation the degree of cross are defined for reducing data dimension. critical quality index of gasoline and diesel. Therefore, it is kernel PCA, where two parameters--the indiscernibility and kinds of oil are major products of FCCU. Endpoint a the degree Gao et al.are [18] proposed indiscernibility dynamic diesel oil inanfluid catalysis and cracking processes. The is of cross defined for reducing data dimension. critical quality index ofpetroleum gasoline and diesel. Therefore, ittwo is Therefore, In the real industry, auto-correlation and lag cross-correlation or dynamics exist among large-scale process variables. light production in processing. Fractionation column is indispensable unit that produces gasoline and The algorithm that can extract the dynamics is utilized to playing a dominant role in monitoring endpoint-related faults degree of where cross are defined for reducing data dimension. critical quality index of gasoline and diesel. Therefore, it isa the kernel PCA, two parameters--the indiscernibility and kinds of oil are major of produces FCCU. Endpoint is The algorithm that can extract the indiscernibility dynamics is utilized to playing ais dominant role inproducts monitoring endpoint-related faults or dynamics exist among large-scale process variables. Therefore, Gao et al. [18] proposed dynamic column an indispensable unit that gasoline and diesel oil in fluid catalysis and cracking processes. The two monitor distillation column with lower missed alarm rate than with the fractionation column to help improve product quality algorithm that column candefined extract thereducing dynamics is dimension. utilized to playing a dominant role monitoring endpoint-related faults the degree of where cross data critical quality index of in gasoline and diesel. Therefore, ittwo isa The monitor distillation with for lower missed alarm rate than with the fractionation column to cracking help improve product quality Therefore, Gao etmethod. al.are [18]parameters--the proposed indiscernibility dynamic kernel PCA, two indiscernibility and diesel oil in fluid catalysis and processes. The kinds of oil are major products of FCCU. Endpoint is the conventional Process monitoring based on PCA and operation safety. monitor distillation column with lower missed alarm rate than with the afractionation column to help improve product quality The algorithm that candefined extract the dynamics is dimension. utilized to playing dominant role monitoring endpoint-related faults the conventional method. Process monitoring based on PCA and operation safety. kernel PCA, two parameters--the indiscernibility and degree of where cross are for reducing data kinds ofquality oil are major products of FCCU. Endpoint is critical index of in gasoline diesel. Therefore, it isa the algorithm monitors only the process variables in unique The operation fractionation processes with and multiple process variables conventional method. Process monitoring based on layer, PCA and safety. monitor distillation column with lower missed alarm rate than with the fractionation column to help improve product quality algorithm monitors only the process variables in unique layer, The fractionation processes with multiple process variables the degree of cross are for data dimension. The algorithm that candefined extract the dynamics isunique utilized to critical quality index of in gasoline it to is algorithm playing a dominant role monitoring endpoint-related faults while abnormal variability may notreducing affect product quality, are highly dynamics and correlated thatdiesel. make itTherefore, be difficult monitors only the process variables in layer, The fractionation processes with and multiple process variables conventional method. Process monitoring based on PCA and operation safety. while abnormal variability maylower not affect product quality, are highly dynamics and correlated that make itproduct be difficult to the The algorithm that can extract the dynamics is utilized to monitor distillation column with missed alarm rate than playing a dominant role in monitoring endpoint-related faults with the fractionation column to help improve quality due to compensation from feedback control in the build first principles model accurately [1-3]. As a result, abnormal variability may notvariables affect control product quality, are highly dynamics and correlated that make it be difficult to while algorithm monitors only the process in unique layer, The fractionation processes with multiple process variables due to compensation from feedback in the build first principles model accurately [1-3]. As a result, monitor distillation column with lower missed alarm rate than the conventional method. Process monitoring based on PCA with the fractionation column to help improve product quality and operation safety. fractionation processes. Unlike PCA, the PLS method a model-based process model monitoring and[1-3]. faultAs diagnosis toabnormal compensation from feedback control in is build first dynamics principles accurately a result, while variability may not affect product quality, are highly and correlated thatand make it be difficult to due fractionation processes. Unlike PCA, the PLS method isthea model-based process monitoring fault diagnosis the conventional method. Process monitoring based on PCA algorithm monitors onlyestablishing the process variables in unique layer, and operation safety. The fractionation processes with multiple process variables powerful approach for relationships of data from approaches are not applicable. Fortunately, due to the fractionation processes. Unlike PCA, the PLS method is a model-based process model monitoring and[1-3]. faultAs diagnosis due toabnormal compensation from feedback control in layer, the build first dynamics principles accurately a to result, powerful approach for establishing relationships data from approaches are not applicable. Fortunately, due the algorithm monitors only the process variables in of unique while variability may not affect product quality, The fractionation processes with multiple process variables are highly and correlated that make it bedue difficult to powerful two layers [19, 20]. PLS-based process monitoring extensive application of distributed control system in to recent approach for establishing relationships of data from approaches areprocess not applicable. Fortunately, the fractionation processes. Unlike PCA, the PLS method is model-based monitoring and fault diagnosis two layers [19, 20]. factors PLS-based process monitoring extensive application of distributed control system in recent while abnormal variability may feedback not affect product quality, due to compensation from control in thea are highly dynamics and correlated that make it be difficult to two build first principles model accurately [1-3]. Asproduction a result, approaches extract latent according to themonitoring maximum years, a large number of valuable data with layers [19, for 20]. PLS-based process extensive application of distributed control system in recent powerful approach establishing relationships of data approaches are not applicable. Fortunately, due to the approaches extract latent factors according to the maximum years, a large number of valuable data with production due to compensation from feedback control in from fractionation processes. Unlike PCA, the PLS method isthea build principles model accurately [1-3]. a on result, model-based process monitoring and faultAs diagnosis co-variance criterion via nonlinear iterative partial least processfirst been collected. The researchers rely the approaches extract latent factors according to the maximum years, a have large number of valuable data with production two layers [19, 20]. PLS-based process monitoring extensive application of distributed control system in recent co-variance criterion via nonlinear iterative partial least process have been collected. The researchers rely on the fractionation processes. Unlike PCA, the[22] PLSdeveloped method isana powerful approach for establishing relationships of data from model-based process monitoring and fault diagnosis approaches are not applicable. Fortunately, due to squares algorithm [21]. The work of available data to achieve data-based operation optimization, criterion viafactors nonlinear iterative partial least process been collected. The researchers rely on the co-variance approaches extract latent according to the maximum years, a have large ofdata-based valuable data system withoptimization, production squares algorithm [21]. The work of [22] developed an available data tonumber achieve operation powerful approach for establishing relationships ofmonitoring data from two layers [19, 20]. PLS-based process approaches are not applicable. Fortunately, due to the extensive application of distributed control in recent improved multiscale PLS method, which combined PLS process monitoring as well as fault diagnosis [4-10]. squares algorithm [21]. The work of [22]combined developed an available data been to achieve data-based operation optimization, co-variance criterion viafactors nonlinear iterative partial least process have collected. The researchers rely on the improved multiscale PLS method, which PLS process monitoring as well as fault diagnosis [4-10]. two layers [19, 20]. PLS-based process monitoring approaches extract latent according to the maximum extensive application of control system in [4-10]. recent improved years, a monitoring large number ofwell valuable data with monitoring production algorithm and wavelet PLS analysis to buildwhich the model, with using Especially, multivariate statistical process multiscale method, combined PLS process as distributed as fault diagnosis squares algorithm [21]. The work developed an available data tonumber achieve data-based operation optimization, algorithm and wavelet analysis to buildof the[22] model, with using Especially, multivariate statistical process monitoring approaches extract latent factors according tointhe maximum co-variance criterion via nonlinear iterative partial least years, a large of valuable data with production process have been collected. The researchers rely on the generalized likelihood ration (GLR) testing the residual (MSPM) is the most popularly utilized [11-15]. The principal algorithm and wavelet analysis to build the model, with using Especially, multivariate statistical process monitoring improved multiscale PLS method, which combined PLS process monitoring as well as fault diagnosis [4-10]. generalized likelihood ration (GLR) testing in theapproach residual (MSPM) is the most popularly utilized [11-15]. The principal co-variance criterion via nonlinear iterative partial least squares algorithm [21]. The work of [22] developed an process have been collected. The researchers relyprincipal on the generalized available data to achieve data-based operation optimization, space. The monitoring performance of in this component analysis (PCA), partial least squares (PLS) likelihood ration (GLR) testing the residual (MSPM) is the most popularly utilized [11-15]. The and wavelet analysis to buildof the[22] with using Especially, multivariate statistical process monitoring space. The monitoring performance ofmodel, this approach component analysis (PCA), partial least squares (PLS) algorithm squares algorithm [21]. The work developed an improved multiscale PLS method, which combined PLS available data to achieve data-based operation optimization, process monitoring as well as fault diagnosis [4-10]. outperforms than PLS-based algorithm. method project high-dimensional and correlated process The likelihood monitoring performance of in thistheapproach component analysis (PCA), utilized partial [11-15]. least squares (PLS) space. generalized ration (GLR) testing residual (MSPM) is the most popularly The principal outperforms than PLS-based algorithm. method project high-dimensional andprocess correlated process improved multiscale PLS method, which combined PLS algorithm and wavelet analysis to build theconsider model, with using process monitoring as well as fault diagnosis [4-10]. Especially, multivariate statistical monitoring However, the above some methods only the static variables into lower-dimensional latent space in chemical than PLS-based algorithm. method project high-dimensional andleast correlated process space. The monitoring performance thisthe approach component analysis (PCA), partial squares (PLS) outperforms However, thelikelihood above some methods only consider the static variables into lower-dimensional latent space inmonitoring chemical algorithm and wavelet analysis to build theof model, with using generalized ration (GLR) testing in residual Especially, multivariate statistical process (MSPM) the most popularly utilized [11-15]. The principal latent variables, whichsome cannot capture the dynamics. Then it industry. isinto However, the above methods only consider the static variables lower-dimensional latent space in chemical outperforms than PLS-based algorithm. method project high-dimensional and correlated process latent variables, which cannot capture the dynamics. Then it industry. generalized likelihood ration (GLR) intesting in the residual space. The monitoring performance of this approach (MSPM) is the most popularly utilized [11-15]. The principal component analysis (PCA), partial least squares (PLS) will result in high missed alarm process monitoring. As for the fractionation processes, Thomas C et al. [16] latent variables, whichsome cannot captureonly the dynamics. Then it industry. However, the above methods consider the static variables into lower-dimensional latent space in chemical will result in high missed alarm in process monitoring. As for the fractionation processes, Thomas C et al. [16] space. The monitoring performance of the thisdynamics approach outperforms than PLS-based algorithm. component analysis (PCA), partialand least squares (PLS) method project high-dimensional correlated process Although a few methods took into account in employed PCA method for extracting latent variables, with result in high missed alarm inthe process monitoring. As for the fractionation processes, Thomas C et al. [16] will latent variables, which cannot capture dynamics. Then it industry. Although athe few methods took into account the dynamics in employed PCA method for extracting latent variables, with outperforms than PLS-based algorithm. However, above some methods only consider the static method project high-dimensional and correlated process variables into lower-dimensional latent space in chemical terms of augmented data, which cannot make accurate two models based on hourly and minute-by-minute for Although a few methods took into account the dynamics in employed PCA method for extractingThomas latent variables, with will result in high missed alarm in process monitoring. As for the fractionation processes, C et al. [16] terms of the augmented data, which cannot make accurate two models based on hourlyof and minute-by-minute for However, some methods only consider static latent variables, which cannot capture the dynamics. Then it variables into lower-dimensional latent space in industry. description of above the dynamic and static structure in the the data. monitoring distillation FCCU. Taking intochemical account of aaugmented data, which cannot make accurate two models based oncolumn hourly minute-by-minute for terms Although few methods took into account the dynamics in employed PCA method for extracting latent variables, with description of the dynamic and static structure in the data. monitoring distillation column of and FCCU. Taking into account latent variables, which cannot capture the dynamics. Then it will result in high missed alarm in process monitoring. industry. As for the fractionation processes, Thomas C et al. [16] Therefore, in this paper, a new dynamic PLS method for the time-varying behavior, such as changes of crude oil description of this the dynamic static structure in the data. monitoring distillation of FCCU. Taking into account of aaugmented data, which cannot make accurate two models based oncolumn hourly minute-by-minute for Therefore, in paper, a and new dynamic PLS method for the for time-varying behavior, suchand as Thomas changes oil terms will result in high missed alarm in process monitoring. Although few methods took into account the dynamics in As thePCA fractionation processes, Cofetcrude al. with [16] employed method for extracting latent variables, in this dynamic paper, a and new static dynamic PLS in method for the time-varying behavior, such as changes ofinto crude oil Therefore, description of structure the data. monitoring distillation of and FCCU. Taking account fewthemethods tookwhich into account the dynamics in terms of aaugmented data, cannot make accurate employed PCA method for extracting latent variables, with two models based oncolumn hourly minute-by-minute for Although in paper, a and new dynamic PLS method for the time-varying behavior, such as changes crude oil 309 Therefore, terms of augmented data, which cannot make accurate Copyright © 2018 IFAC on description of this the dynamic static structure in the data. two models based hourly minute-by-minute for monitoring distillation column of and FCCU. Takingofinto account Copyright © 2018 IFAC description of the dynamic and static structure in the data. monitoring distillation columnsuch of FCCU. Takingofinto account the time-varying behavior, as changes crude oil 309 Copyright © 2018 IFAC 309 Therefore, in this paper, a new dynamic PLS method for the time-varying behavior, such as changes of crude oil Therefore, in this paper, a new dynamic PLS method for Copyright © 2018, 2018 IFAC 309Hosting by Elsevier Ltd. All rights reserved. 2405-8963 © IFAC (International Federation of Automatic Control) Peer review under responsibility of International Federation of Automatic Copyright © 2018 IFAC 309Control. 10.1016/j.ifacol.2018.09.319 Copyright © 2018 IFAC 309

2018 IFAC ADCHEM 316 Shenyang, Liaoning, China, July 25-27, 2018 Xue Xu et al. / IFAC PapersOnLine 51-18 (2018) 315–320

quality-related monitoring is proposed and applied into fractionation processes. The contributions of the work are in the following. First, the process data space is decomposed into dynamic quality-related and static quality-unrelated subspace, respectively. For quality-unrelated space, it can contain a few large variations that is further divided by performing PCA decomposition to improve the monitoring performance. The remainder of this paper is organized as follows. In section 2, fractionation process of FCCU is described. In section 3, the PLS method is reviewed briefly and a modified dynamic PLS approach is presented, and the corresponding monitoring strategy is given in later. The proposed algorithm is applied in the real fractionation processes in section 4. Finally, the conclusions are concluded and the following work is discussed in the section 5.

Wet gas Cold reflux

1 2 4 Top Pumparound

7

Sewage

9

Gasoline

13 14 17 Mid Pumparound

21

Stripper

27 Steam

2. FRACTIONATION PROCESS DESCRIPTION

29

Fractionating column is an important unit in catalytic cracking processes, whose process structure diagram is shown in Figure 1. The high-temperature oil-gas mixture enters the fractionation column from column bottom, and then are separated several fractions according to different boiling temperatures. . Among them, overhead product is gasoline and side-lines product is diesel. Gasoline and diesel are the main products of FCCU. The endpoint is an important quality index of gasoline and diesel. The major process variables relevant endpoint of gasoline are top tower temperature, top pressure, flow of top reflux, the 1st pump-around reflux flow, that of diesel are top pressure, column bottom temperature, the 1st pump-around reflux temperature, respectively. According to gasoline, diesel oil as the major fuel, endpoint of gasoline is extremely high, which will make it difficult to evaporate and burn the gasoline completely. The endpoint of diesel will affect the mobility under different climate conditions. Further, the endpoint also has an effect on product distribution in the production processes. Therefore, a great deal of researchers focus on anomalies monitoring of fractionation process. In the early researches, researchers look forward to establish first principle model for solving the abnormal monitoring, whereas several assumptions are made to simplify the model because of difficulties in the solution of quite complex mechanism model. However, the above strategy will lead to a great difference between the built model and the real process [15]. As a result, considering the above drawbacks, multivariate statistical process monitoring method provides a data-driven framework to monitor the fractionation process. The fractionation process is a typical dynamic process, with dynamic characteristics derived from the interactions among complex recycle reflux. Therefore, a quality related modified dynamic PLS monitoring method is proposed for monitoring the fractionation processes.

Bottom Pumparound

Diesel

30

Gas-oil Feed

Heat Exchangers

Oil slurry

Fig.1 The flowchart around a typical fractionation column Collect normal process and quality data and construct inputs X 

n m

outputs

containing

Y 

n p

n

samples with

containing

n

m

variables, and

samples with

p

variables.

Then, the scaled data is projected to low-dimensional space as the following PLS model [23]: X     Y 

where

T



l



l

nl

i

i

T

T

E

T

T

F

t ip i  E = T P t ip i  F = T Q

(1)

is the score matrix of inputs X ,

P 

m l

is

pl

the loading matrix of inputs X , Q   is the loading matrix of outputs Y , l is the number of latent factors, E and F are residuals for inputs and outputs, respectively. The detailed PLS method is described in literatures [24, 25]. The matrix W is weight matrix of deflation matrix T X i  X i  t i  1 p i  1 . In order to represent t i in terms of original data X , the following equation is utilized. T  XR

where

T

R  W (P W )

1

from [26].

3.2 A modified dynamic PLS (MDPLS) modelling algorithm Generally, both auto-correlations and cross-correlations exist inside real industrial data simultaneously. The above static PLS model cannot capture dynamics in variables. As a result, missed alarm rate may increase when static model is applied to monitor dynamic processes. In this paper, a new dynamic latent variable monitoring algorithm is proposed and applied to quality-related fault detection.

3. A MODIFIED DYNAMIC PLS MODELING FOR PROCESS MONITORING 3.1 Reviews of PLS

310

2018 IFAC ADCHEM Shenyang, Liaoning, China, July 25-27, 2018 Xue Xu et al. / IFAC PapersOnLine 51-18 (2018) 315–320

The explicit dynamic and static relations between variables is modeled by the dynamic inner PLS (DiPLS) algorithm[27]:

The score vector statistic:

317

can be monitored via Hotelling’s

tx

T

2

T

 X  T P  E  1 1 T T  Y  G 1 ( z ) t 1 q 1  ...  G l ( z ) t l q l  F

2

where

where G i ( z  1 )   iT z , z  [1, z  1 , ... z  s  1 ] describes the dynamic relationship between the quality data and the dynamic scores t i , z  1 denotes delay operator. l is the number of latent variables, which is obtained by crossvalidation method. However, the above approach focuses on building a regression model, rather than supply a statistical model for process monitoring. Further, the procedures of DiPLS algorithm is similar to that of PLS, thus scores T include variations orthogonal to Y. In order to overcome the aforementioned shortcomings, a modified monitoring strategy is proposed. According to DiPLS method, the predicted quality variable can be shown [27]: T T T (3) Yˆ = Tˆ s Q = T c B Q  Z s  1 R B Q where

B =   s  1  s  2 ...  1 

T

,

T c   T1 T 2 ... T s  1 

Yˆ = L d C d H

T

Yˆ

R

† d

of

T

 (R d R

d

)

1

Span R

T d

,

X

s

is performed

 X  K dR

lx

T

1

vk

2

Tv

 

g

,

 k



 gn T      i  td  i    i  g 1 

k

 1  , ..., t d

(12)

2

where



v

  t d

k

 g  

T

.

. T

Tv  v k 

is covariance of

1 v

(13)

vk

V

,

V  [ v 1 , ..., v n ]

T

.

Table 2 Monitoring statistics and control limits Statistic

Calculation

Control limit 2

l x ( n  1) 2

Tx

,where

T

tx

1 x

n (n  lx )

tx

Fl

x

, n  l x ;

2

X

s

l d ( n  1) Tv

with

2

1 v

n ( n  ld )

vk

Fl

d

, n  l d ;

2

(S x / 2  x )  (2  x / S x )

ex

In this paper, the monitoring statistics and corresponding control limits are listed in the Table 2. Among them, where l d , l x are the number of the latent variables, n is the number

(6)

is the principal component score matrix that is

useless to predict outputs, E x represents residuals, respectively. According to the proposed MDPLS method, the outer model of input and output spaces are decomposed into the following form:

of training samples, ld , n  ld 2

d

represents

of freedom, where

(7)

Fl

, n  l d ;

indicates

degrees of freedom under

 (2  x / S x )

†

  X  L d R d  Tx Px  E x  ˆT T   Y  Ts Q  F

T

vk 

2

Qx X s  Tx Px  E x

Tx

(10)

can be time independent, it can be monitored by

components

where

ex

As

. Due to the large variation in the residual

subspace, PCA decomposition is further realized on

2

ˆ    1 , ... 

statistic

d

.

where

 LdQ d

R d 

tx

statistic:

Q

 gn T  ˆ      i    i    i  g 1 

where R d  R Q T H d C d 1 For the remaining static qualityirrelevant process variables via projecting onto the orthogonal

(9)

tx

where  i is the parameters of auto-regressive model, g is model order, it can be obtained by Bayesian Information Criterion. The parameters are solved by the least squares algorithm as following:

, with

†

x

The t d is not time independent series, which is not suitable to monitor. As a consequence, it can be described as a stationary time series model with auto-regressive mode as following: t d ( k )   1 t d ( k  1)  ...   p t d ( k  g )  v  k  (11)

T



1

is the covariance of

x

Qx =

(4) where Q d  H d C d contains l d nonzero singular values in descending order and the corresponding right singular vectors. 1 T 1 (5) L d  Yˆ H d C d  Z s  1 R B Q H d C d  Z s  1 R d d



can be monitored via

ex

Ti  X i R

. The dynamic predictable quality variable singular value decomposition as following:

T

Tx = t x 

(2)

Sx



2



F

-distribution with

confidence coefficient.

-distribution with 2  x , S x degrees

means the variance of the training

samples, and  x represents mean of training samples, respectively [28, 29].

3.3 MDPLS-Based Monitoring Given a new sample

x new

4. THE SCENARIO STUDIES IN QUALITY RELATED MONITORING OF FRACTIONATION PROCESSES

, it can be calculated as follows:

T

t d  R d x new T

t x  Px x s

e x  (I  Px Px )x s

(8)

311

This section applies the proposed MDPLS method to endpoint relevant monitoring in fractionation processes. The collected data contain 28 process variables and 2 quality variables (endpoint of gasoline and endpoint of diesel). In this paper, a number of 600 samples were collected, of which

2018 IFAC ADCHEM 318 Shenyang, Liaoning, China, July 25-27, 2018 Xue Xu et al. / IFAC PapersOnLine 51-18 (2018) 315–320

flowrate of PA-1 (℃ )

300 samples under fault-free condition, and additional 300 samples under fault condition. The prediction performance of two quality variables with PLS and dynamic PLS can be seen in the Fig. 3 and Fig. 4, respectively. The dominant process variables are shown in the Fig. 5. It is obvious that the dynamic PLS have better prediction accuracy than PLS.

top temperature (℃ )

The real value The predicted value

200

195 0

50

100

150

200

250

300

temperature of PA-1 (℃ )

Gasoline endpoint

205

The real value The predicted value

360

340 330

310 300

0

50

100

150

200

250

300

Fig.3. Quality prediction of PLS The real value The predicted value

Gasoline endpoint

205

211 210 209 208

0

50

100

150

200

250

300

50

100

150

200

250

300

0

50

100

150

200

250

300

0

50

100

150

200

250

300

124 122 120 118

0

164 162 160 158

118 116 114 112 110

Fig.5 The original dominant process variables To demonstrate the effectiveness of the algorithm, quality related faults 1 and faults 2 are modeled and monitored, comparing with conventional dynamic PLS (DPLS) methods. In the fractionation processes, 1st pump-around flowrate and the top tower pressure with too low are common faults. The former (fault 1) can lead to a higher temperature with the top and middle of the fractionating column, even further lead to the phenomenon of punching column, which will make the endpoint of gasoline and diesel rise, and also affect the heat source supply for the downstream devices. The latter (fault 2) will also have impact on the product quality. The monitoring performance with traditional DPLS and the MDPLS under fault scenario 1 is described in Fig.6. It is seen from T v2 statistic with Fig.6 (b) that the endpoint relevant fault is detected precisely from the 183th to 267th sample and have a trending of returning to the normal condition after the 267th samples because of the closed-loop controller to decrease the impact of the fault on process. Nevertheless, the 2 T statistic from Fig.6 (a) cannot detect the influence of feedback control on quality indexes, which still above the control limit after the 267th sample considering as the false alarms. Even if T x2 statistic in Fig 6 (b) above the control limit, which is a process-specific variation considering as

200

195 0

50

100

150

200

250

300

The real value The predicted value

360 350

Diesel endpoint

212

320

top pressure (MPa)

Diesel endpoint

350

213

340 330 320 310 300

0

50

100

150

200

250

300

Fig.4 Quality prediction of dynamic PLS

312

2018 IFAC ADCHEM Shenyang, Liaoning, China, July 25-27, 2018 Xue Xu et al. / IFAC PapersOnLine 51-18 (2018) 315–320

319

30

Q statistic

normal situation when concentrating on the relevant endpoint fault. Therefore, the effectiveness of proposed monitoring method is demonstrated via the above experimental results. When the fractionation process under fault 2 situation, the monitoring results with DPLS and the MDPLS algorithms are displayed in Fig 7(a) and Fig (b), respectively. The T v2 statistic with Fig 7 (b) tends to exceed the control limit from the 145th sample, which denotes the quality-related faults can be dete cted successfully. However, the T statistic with DPLS method from Fig.7 (a) have missed alarms among 165-185, 231-240,267-280 samples.

x

20 10 0

0

50

100

150

200

250

300

(b) MDPLS

Fig. 6 Fractionation process monitoring results (faulty 1)

Fault 1: Fault 2:

60

60

T2 statistic

T2 statistic

40 20 0

0

50

100

150

200

250

40 20

300

0

5

0

50

100

150

200

250

0

200

250

300

50

100

150

200

250

300

0

(a) DPLS

60

T statistic

60 40

v

T statistic

150

20

300a

80

v

20 0

0

50

100

150

200

250

40 20

300 0

150

0

100

150

T statistic

50

x

x

100

40

Q statistic

Q statistic

10

(a)…DPLS

T statistic

50

60

15

0

0

0

0

50

100

150

200

250

50

100

150

200

250

300

200

250

300

100 50

300 0

313

0

50

100

150

2018 IFAC ADCHEM 320 Shenyang, Liaoning, China, July 25-27, 2018 Xue Xu et al. / IFAC PapersOnLine 51-18 (2018) 315–320

12

x

Q statistic

10 8 6 4 2 0

0

50

100

150

200

250

300

(b) MDPLS

Fig.7 Fractionation process monitoring results (faulty 2) 5. CONCLUSIONS This paper proposed a modified dynamic PLS quality-related monitoring approach for fractionation processes. It provides a simple implementation to achieve comprehensive monitoring. The effectiveness of the proposed method is illustrated by the application results on a real fractionation processes. The corresponding fault diagnosis strategy of the proposed monitoring scheme can be developed for localizing the faults in the future work. ACKNOWLEDGEMENTS This work was financially supported in part by the National Natural Science Foundation of China under Grants 61590922, 61525302, 61673097, 61573022, and 61490704, the Project of Ministry of Industry and Information Technology of China under Grand 20171122-6, the Projects of Shenyang under Grand Y17-0-004, and the Fundamental Research Funds for the Central Universities under Grand N160801001, N161608001. REFERENCES (1) S. J. Qin, G. Cherry, R. Good, J. Wang, and C. A. Harrison, Semiconductor manufacturing process control and monitoring: A fabwide framework, Journal Process Control, 2006, 16(3), 179–191. (2) Yin S, Li X, Gao H, Data-based techniques focused on modern industry: An overview, IEEE Transactions on Industrial Electronics, 2015, 62(1), 657-667. (3) Z. Ge, Z. Song, and F. Gao, Review of recent research on data-based process monitoring, Industrial & Engineering Chemistry Research, 2013, 52(10), 3543–3562. (4) Ding, Jinliang, et al. Data-Based Multiobjective Plant-Wide Performance Optimization of Industrial Processes Under Dynamic Environments, IEEE Transactions on Industrial Informatics, 2016, 12(2), 454-465. (5) Ding J, Yang C, Chai T, Recent Progress on Data-Based Optimization for Mineral Processing Plants, Engineering, 2017, 3(2), 183-187. (6) Q. Liu, S.J. Qin, T.Y. Chai, Multi-block concurrent PLS for decentralized monitoring of continuous annealing processes, IEEE Transactions on Industrial Electronics, 2014, 61 (11), 6429–6437. (7) Tong C, Lan T, Shi X, Ensemble modified independent component analysis for enhanced non-Gaussian process monitoring, Control Engineering Practice, 2017, 58, 34-41. (8) Q. Liu, S.J. Qin, T.Y. Chai, Decentralized fault diagnosis of continuous annealing processes based on multi-level PCA, IEEE Transactions on Automation Science and Engineering, 2013, 10 (3), 687–698. (9) Tong C, Lan T, Shi X, Statistical process monitoring based on nonlocal and multiple neighborhoods preserving embedding model, Journal of Process Control, 2017, 65, 34-40. (10) Zhu Y, Zhong Z, Zheng W X, HMM-Based H∞ Filtering for DiscreteTime Markov Jump LPV Systems Over Unreliable Communication Channels, IEEE Transactions on Systems Man & Cybernetics Systems, 2017, 99, 1-12.

314

(11) Zhao, C. H, Gao, F. R, Niu, D. P, Wang, F. L, Enhanced process comprehension and quality analysis based on subspace separation for multiphase batch processes, AIChE Jouinal, 2011, 57, 388-403. (12) Wang G, Yin S, Quality-Related Fault Detection Approach Based on Orthogonal Signal Correction and Modified PLS. IEEE Transactions on Industrial Informatics, 2015, 11(2), 398-405. (13) Li G, Qin S J, Zhou D, A New Method of Dynamic Latent-Variable Modeling for Process Monitoring, IEEE Transactions on Industrial Electronics, 2014, 61(61), 6438-6445. (14) Chen T, Zhang J, On-line multivariate statistical monitoring of batch processes using Gaussian mixture model, Computers & Chemical Engineering, 2010, 34(4), 500-507. (15) Peng X, Tang Y, Du W, Multimode Process Monitoring and Fault Detection: A Sparse Modeling and Dictionary Learning Method, IEEE Transactions on Industrial Electronics, 2017, 64(6), 4866-4875. (16) Thomas C, Wada T, Seborg D E, Principal component analysis applied to process monitoring of an industrial distillation column. IFAC Proceedings Volumes, 1996, 29(1), 5859-5864. (17) Alghazzawi A, Lennox B, Monitoring a complex refining process using multivariate statistics. Control Engineering Practice, 2008, 16(3), 294307. (18) Gao Q, Chang Y, Xiao Z, Monitoring of Distillation Column Based on Indiscernibility Dynamic Kernel PCA. Mathematical Problems in Engineering, 2016, 4, 1-11. (19) Li G, Liu B, Qin S J, et al. Quality Relevant Data-Driven Modeling and Monitoring of Multivariate Dynamic Processes: The Dynamic T-PLS Approach, IEEE Transactions on Neural Networks, 2011, 22(12), 22622271. (20) Kresta J V, Macgregor J F, Marlin T E. Multivariate statistical monitoring of process operating performance, Canadian Journal of Chemical Engineering, 1991, 69(1), 35-47. (21) J. F. MacGregor and T. Kourti. Statistical process control of multivariate processes, Control Engineering Practice, 1995, 3(3), 403– 414. (22) Madakyaru M, Harrou F, Sun Y, Improved Data-based Fault Detection Strategy and Application to Distillation Columns, Process Safety & Environmental Protection, 2017, 107, 22-34. (23) Geladi P, Kowalski B R, Partial least-squares regression: a tutorial. Analytica Chimica Acta, 1986, 185(86), 1-17. (24) Hoskuldsson A, PLS regression methods. J Chemom. 1988, 2, 211–228. (25) Kristian HellandHans E. BerntsenOdd S. BorgenHarald Martens, Recursive algorithm for partial least squares regression, Chemometrics & Intelligent Laboratory Systems, 1992, 14(1), 129-137. (26) De Jong S, SIMPLS: an alternative approach to partial least squares regression, Chemometric and Intelligent Laboratory Systems, 1993, 18, 251–263. (27) Dong Y, Qin S J, Dynamic-Inner Partial Least Squares for Dynamic Data Modeling, IFAC-Papers OnLine, 2015, 48(8), 117-122. (28) S. J. Qin, Statistical process monitoring: Basics and beyond, Journal of Chemometrics, 2003, 17(8), 480–502. (29) Box G E P, Some Theorems on Quadratic Forms Applied in the Study of Analysis of Variance Problems, I. Effect of Inequality of Variance in the One-Way Classification, Annals of Mathematical Statistics, 1954, 25(2), 290-302.