CN102385706A - Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production - Google Patents

Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production Download PDF

Info

Publication number
CN102385706A
CN102385706A CN2011103108086A CN201110310808A CN102385706A CN 102385706 A CN102385706 A CN 102385706A CN 2011103108086 A CN2011103108086 A CN 2011103108086A CN 201110310808 A CN201110310808 A CN 201110310808A CN 102385706 A CN102385706 A CN 102385706A
Authority
CN
China
Prior art keywords
data
model
pta
training
soft sensing
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN2011103108086A
Other languages
Chinese (zh)
Inventor
纪彭
蒋鹏飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN2011103108086A priority Critical patent/CN102385706A/en
Publication of CN102385706A publication Critical patent/CN102385706A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a soft sensing method based on model learning and used in pure terephthalic acid (PTA) production. The traditional soft sensing can not effectively perform on-line soft sensing to a material consumption index and a quality index during a PTA production process. In the invention, a mechanism model of a PTA production apparatus which is verified is used to perform flow simulation. Under an offline condition, aiming at different typical working conditions, mass data sets are generated for one time. And then according to the different typical working conditions, the data sets generated through simulation are used to train and acquire a series of data models which are used for positive calculation. Under an on-line state, the data models are used to complete a soft sensing task of a key variable during the PTA production process. In the invention, on a real-time calculation aspect, a calculating speed is fast and accuracy is high. Simultaneously, robustness is good. And the method does not depend on one specific flow simulation method and the specific data model.

Description

A kind of flexible measurement method that is used for PTA production based on model learning
Technical field
The invention belongs to chemical technology field, particularly a kind of flexible measurement method that is used for PTA production based on model learning.
Background technology
Along with the rapid growth in China weaving market and polyester market, essence has obtained develop rapidly to benzene first diacid industry, and production capacity and output have all had significantly to be increased.Data show according to the relevent statistics, and China is PTA (pure terephthalic acid, smart in benzene first diacid) market the biggest in the world at present, estimates that 2010 annual requirements will reach 1,700 ten thousand tons.
In recent years domestic PTA industry develops rapidly, and production capacity and output all have significantly to be increased.But the control of domestic PTA production technology, especially energy consumption and material consumption has been compared certain distance with the advanced level of external same device.Such as the important judgment criteria of product carboxyl benzaldehyde concentration as product quality indicator; Very important in the PTA production run; But on-the-spot carboxyl benzaldehyde concentration must obtain through off-line analysis chemical examination once in four hours; Therefore concerning the site operation personnel, product quality indicator quantity of information wretched insufficiency and hysteresis quality are bigger.Therefore, need effectively the material consumption index and the quality index of PTA production run to be carried out online soft sensor, in time find the production operation problem, propose operations improvement and operating mode adjustment scheme, material consumption index and quality index are controlled at a reasonable levels.Traditional soft measurement is divided into two kinds: a kind of is to set up data-driven model; Calculate the predicted value of obtaining key variables through the model forward; But this method computing velocity is exceedingly fast is very high to the model training data demand, such as requires the training data frequency identical.Second kind is to set up target workshop section mechanism model, uses the method for flowsheeting to carry out the soft measurement to key variables, and the flowsheeting iterative computation is consuming time, and cost is huge because flowsheeting software copyright problem system puts into operation.Thereby be difficult to realize quick, the accurate and economic dispatch requirement in the calculating of industrial process online soft sensor.
Summary of the invention
The objective of the invention is deficiency, the flexible measurement method that provides a kind of PTA of being used for to produce based on model learning to prior art.
The inventive method utilizes the mechanism model of the PTA process units of verified to carry out flowsheeting; Under off-line state; To the large batch of data set of the disposable generation of different typical conditions, then according to different typical conditions, the data set training that simulation produces before utilizing draws a series of data models that forward calculates that are used for; Under presence, accomplish soft measuring task to the key variables in the PTA production run with this data model.
The concrete steps of the inventive method are:
Step (1). use flowsheeting software to set up the mechanism model of PTA production run device, and mechanism model is carried out verification according to the process knowledge of PTA production run;
Step (2). utilize the mechanism model of the PTA process units that verification is accomplished in the step (1)
Figure 2011103108086100002DEST_PATH_IMAGE002
planted different typical conditions carry out flowsheeting generation
Figure 829479DEST_PATH_IMAGE002
class frequency identical corresponding to the data set of typical condition
Figure 2011103108086100002DEST_PATH_IMAGE004
separately;
Figure 933571DEST_PATH_IMAGE002
is natural number, and training auxiliary variable is set;
Step (3). the characteristic according to the PTA process units is selected the raw data model under
Figure 2011103108086100002DEST_PATH_IMAGE008
individual typical condition;
Figure 2011103108086100002DEST_PATH_IMAGE010
data set that utilizes step (2) simulation to produce then, training draws and is used for the data model that forward calculates respectively;
Step (4). judge and whether equal
Figure 962575DEST_PATH_IMAGE002
: if
Figure 2011103108086100002DEST_PATH_IMAGE016
; Return step (3); Select the raw data model under
Figure 2011103108086100002DEST_PATH_IMAGE018
individual typical condition, training draws and is used for the data model
Figure 2011103108086100002DEST_PATH_IMAGE020
that forward calculates; If
Figure DEST_PATH_IMAGE022
then gets into step (5);
Step (5). that gets out from the training of historical floor data be individual to be used for data model that forward calculates; Draw and current working data matching model predication value according to formula
Figure DEST_PATH_IMAGE024
and
Figure DEST_PATH_IMAGE026
, with the output result of data model predicted value as soft measurement.
Described
Figure DEST_PATH_IMAGE028
is the data model vector that predicts the outcome, and
Figure DEST_PATH_IMAGE030
is
Figure 359107DEST_PATH_IMAGE002
the corresponding weight coefficient of individual different data model. for different data models
Figure DEST_PATH_IMAGE034
on the current condition of the data model calculations.
Beneficial effect of the present invention is:
1. the PTA production data that the flexible measurement method based on model learning that is used for PTA production of the present invention uses mechanism model simulation generation is as training data; Overcome training data frequency difference, training data and possibly receive the disappearance mistake etc. of external disturbance and influence of measurement error, training data to have a strong impact on the problem of model accuracy and have computing velocity and higher accuracy rate faster aspect the calculating in real time, had robustness preferably simultaneously.
2. the flexible measurement method based on model learning that the PTA of being used for of the present invention produces uses the data model that trains to replace the mechanism model that generally adopts at present.Therefore to have computing velocity fast for this method, and the low advantage that assesses the cost is highly suitable for the line operation.
3. the flexible measurement method based on model learning that is used for PTA production of the present invention is realized simple, does not rely on a certain concrete flowsheeting method and concrete data model, and both can select arbitrarily according to actual conditions, use very flexible.
4. the flexible measurement method based on model learning that the PTA of being used for of the present invention produces is compared with existing flexible measurement method; The method simple and clear principle that the present invention adopts be convenient to realize on the computing machine, and dirigibility is fine; Can select flowsheeting method and data model structure arbitrarily; Give full play to advantage separately, satisfy the rapidity of chemical-process real-time optimization better, the requirements such as convergence that accuracy is become reconciled.
Description of drawings
Fig. 1 instantiation system construction drawing of the present invention.
Fig. 2 instantiation process flow diagram of the present invention.
Embodiment
Below in conjunction with accompanying drawing the present invention is further described.
As shown in Figure 1; The present invention uses flowsheeting software Aspen Plus to come the PTA production procedure mechanism model that we set up is simulated; Historical data base has been preserved related data under the different typical conditions and has been used for sending into the mechanism model computing module and carries out analog computation by Aspen Plus, and data model substitution mechanism model that a neural network and least square regression (PLS) make up on-line operation is at the scene practised in a series of simulated datas training that utilize analog computation to obtain then.The data model module is regularly calculated the real time data of from real-time data base, obtaining, and data model output is the result export as soft measurement result.
The concrete steps of the inventive method are:
Step (1). use flowsheeting software to set up the mechanism model of PTA production run device, and mechanism model is carried out verification according to the process knowledge of PTA production run;
Step (2). utilize the mechanism model of the PTA process units that verification is accomplished in the step (1)
Figure 306204DEST_PATH_IMAGE002
planted different typical conditions carry out flowsheeting generation
Figure 760188DEST_PATH_IMAGE002
class frequency identical corresponding to the data set of typical condition
Figure 959088DEST_PATH_IMAGE004
separately;
Figure 627967DEST_PATH_IMAGE002
is natural number, and training auxiliary variable
Figure 769402DEST_PATH_IMAGE006
is set;
Step (3). the characteristic according to the PTA process units is selected the raw data model under
Figure 523731DEST_PATH_IMAGE008
individual typical condition;
Figure 198426DEST_PATH_IMAGE010
data set that utilizes step (2) simulation to produce then, training draws and is used for the data model
Figure 721811DEST_PATH_IMAGE012
that forward calculates respectively;
Step (4). judge
Figure 151656DEST_PATH_IMAGE014
and whether equal
Figure 580232DEST_PATH_IMAGE002
: if ; Return step (3); Select the raw data model under
Figure 436509DEST_PATH_IMAGE018
individual typical condition, training draws and is used for the data model that forward calculates; If
Figure 687548DEST_PATH_IMAGE022
then gets into step (5);
Step (5).
Figure 704046DEST_PATH_IMAGE002
that gets out from the training of historical floor data be individual to be used for data model that forward calculates; Draw and current working data matching model predication value according to formula
Figure 388974DEST_PATH_IMAGE024
and
Figure 98304DEST_PATH_IMAGE026
, with the output result of data model predicted value as soft measurement.
Described
Figure DEST_PATH_IMAGE036
is the data model vector that predicts the outcome, and
Figure 235893DEST_PATH_IMAGE030
is
Figure 118398DEST_PATH_IMAGE002
the corresponding weight coefficient of individual different data model.
Figure 408565DEST_PATH_IMAGE032
for different data models
Figure 351113DEST_PATH_IMAGE034
on the current condition of the data model calculations.
Described
Figure 913682DEST_PATH_IMAGE028
is the data model vector that predicts the outcome, and
Figure 334299DEST_PATH_IMAGE030
is
Figure 993819DEST_PATH_IMAGE002
the corresponding weight coefficient of individual different data model.
Figure 231902DEST_PATH_IMAGE032
for different data models
Figure 157133DEST_PATH_IMAGE034
on the current condition of the data model calculations.
Raw data model described in the step (3) has three kinds of different forms in the present invention: BP neural network model, multivariate statistical model and fitting of a polynomial formula, the raw data model is not limited to three kinds of selected forms of the present invention in actual use.
The training data that described data model uses is to utilize mechanism model to carry out the identical data acquisition of frequency that flowsheeting produces.
Described mechanism model is meant philosophy and the principle in related discipline field and the mathematical model that characteristic deduced of forming the parts of object.
As shown in Figure 2; The present invention at first sets up the mechanism model of PTA process units; And mechanism model carried out verification; From historical data base, select data under the different operating modes to send into then to carry out analog computation by Aspen-Plus in the mechanism model computing module to produce the identical calculation result data collection of frequency at last, then with the analog computation result under the different operating modes as training data to the training of data model, draw a series of data models that forward calculates that are used for.Under the presence, replace original mechanism model, from real-time data base, obtain the input data and calculate, the result of calculation of data model is exported as soft measurement result with data model.
The enforcement of The whole calculations flow process is to combine Aspen Plus12.1 OOMF Script Language and VB.net hybrid programming to realize.Wherein historical data base is Sqlite.
Embodiment one:
1, uses Aspen-Plus flowsheeting software that PTA oxidation reaction workshop section is carried out modelling by mechanism, and export tail oxygen concentration, carbonomonoxide concentration etc. through field instrument system and Model Calculation and compare and proofread and correct mechanism model.
2, from historical data base, obtain under the PTA oxidation reaction workshop section typical condition reactor feed; Air capacity; Temperature of reactor, the mechanism model that service datas such as pressure driving Aspen-Plus flowsheeting software finishes to adjustment carries out analog computation to produce large quantities of historical floor data collection.
3, to soft measurand P-xylene unit consumption and carboxyl benzaldehyde content and the stronger situation of associative operation variable nonlinear relationship in the actual implementation process of method; We have selected the BP neural network as data model to be trained, and calculate soft measurand P-xylene unit consumption and carboxyl benzaldehyde content with it.The characteristics of BP neural network model are to come constantly adjustment network parameter through backpropagation, make that finally the error sum of squares of network calculations result and training data is minimum.
4, from historical data base, take out the large quantities of historical data set pair BP neural network models that produce according to typical condition and train the BP neural network model that obtains each typical condition correspondence.
5, select different neural network models to calculate according to the real time data of field working conditions and the performance variable that from real-time data base, obtains, and result of calculation is exported as the soft measured value of key variables to accomplish forward.
Embodiment two:
1, uses Aspen-Plus flowsheeting software that PTA oxidation reaction workshop section is carried out modelling by mechanism, and export tail oxygen concentration, carbonomonoxide concentration etc. through field instrument system and Model Calculation and compare and proofread and correct mechanism model.
2, from historical data base, obtain reactor feed under the PTA oxidation reaction workshop section typical condition, air capacity, temperature of reactor, service datas such as pressure drive the mechanism model that Aspen-Plus finishes to adjustment and carry out analog computation to produce historical floor data collection.
3,, find to select PLS (PLS) model to have better model accuracy as data statistics model to be trained with BP neural network contrast back to acetic acid content and two variablees of terephthaldehyde's acid content in the soft measurand in the actual implementation process of method.
4, from historical data base, take out large quantities of historical data set pair PLSs (PLS) model training that produces according to typical condition and obtain corresponding PLS (PLS) model of each typical condition.
5, select different PLS (PLS) models to calculate according to field working conditions and the performance variable real time data from real-time data base, obtained to accomplish forward, and with result of calculation as the soft measured value output of key variables.
The present invention exists many different forms to implement, and therefore is not limited to the embodiment that instructions is listed.

Claims (1)

1. one kind is used for the flexible measurement method based on model learning that PTA produces, and it is characterized in that following steps:
Step (1). use flowsheeting software to set up the mechanism model of PTA production run device, and mechanism model is carried out verification according to the process knowledge of PTA production run;
Step (2). utilize the mechanism model of the PTA process units that verification is accomplished in the step (1)
Figure 2011103108086100001DEST_PATH_IMAGE001
planted different typical conditions carry out flowsheeting generation
Figure 464326DEST_PATH_IMAGE001
class frequency identical corresponding to the data set of typical condition
Figure 840950DEST_PATH_IMAGE002
separately;
Figure 740773DEST_PATH_IMAGE001
is natural number, and training auxiliary variable
Figure 2011103108086100001DEST_PATH_IMAGE003
is set;
Step (3). the characteristic according to the PTA process units is selected the raw data model under
Figure 244566DEST_PATH_IMAGE004
individual typical condition;
Figure DEST_PATH_IMAGE005
data set that utilizes step (2) simulation to produce then, training draws and is used for the data model
Figure 733185DEST_PATH_IMAGE006
that forward calculates respectively;
Step (4). judge
Figure 2011103108086100001DEST_PATH_IMAGE007
and whether equal
Figure 218393DEST_PATH_IMAGE001
: if
Figure 605512DEST_PATH_IMAGE008
; Return step (3); Select the raw data model under
Figure 2011103108086100001DEST_PATH_IMAGE009
individual typical condition, training draws and is used for the data model
Figure 912997DEST_PATH_IMAGE010
that forward calculates; If
Figure 2011103108086100001DEST_PATH_IMAGE011
then gets into step (5);
Step (5).
Figure 928226DEST_PATH_IMAGE001
that gets out from the training of historical floor data be individual to be used for data model that forward calculates; Draw and current working data matching model predication value according to formula
Figure 725281DEST_PATH_IMAGE012
and , with the output result of data model predicted value as soft measurement;
The
Figure 521068DEST_PATH_IMAGE014
predicted results for the data vector,
Figure 2011103108086100001DEST_PATH_IMAGE015
is
Figure 819194DEST_PATH_IMAGE001
different data models corresponding weights;
Figure 829875DEST_PATH_IMAGE016
for different data models
Figure 2011103108086100001DEST_PATH_IMAGE017
Data on the current condition of the model calculations.
CN2011103108086A 2011-10-14 2011-10-14 Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production Pending CN102385706A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN2011103108086A CN102385706A (en) 2011-10-14 2011-10-14 Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN2011103108086A CN102385706A (en) 2011-10-14 2011-10-14 Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production

Publications (1)

Publication Number Publication Date
CN102385706A true CN102385706A (en) 2012-03-21

Family

ID=45825107

Family Applications (1)

Application Number Title Priority Date Filing Date
CN2011103108086A Pending CN102385706A (en) 2011-10-14 2011-10-14 Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production

Country Status (1)

Country Link
CN (1) CN102385706A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105210089A (en) * 2013-05-22 2015-12-30 惠普发展公司，有限责任合伙企业 Production simulation
CN108197743A (en) * 2017-12-31 2018-06-22 北京化工大学 A kind of prediction model flexible measurement method based on deep learning
CN113516200A (en) * 2021-07-30 2021-10-19 盛景智能科技(嘉兴)有限公司 Method and device for generating model training scheme, electronic equipment and storage medium
CN113516200B (en) * 2021-07-30 2024-06-04 盛景智能科技(嘉兴)有限公司 Model training scheme generation method and device, electronic equipment and storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101598737A (en) * 2009-06-26 2009-12-09 华东理工大学 Among the pure terephthalic acid to the flexible measurement method of carboxyl benzaldehyde content
CN101620590A (en) * 2009-07-06 2010-01-06 浙江大学 Parameter estimation method for multi-condition and large-scale chemical process model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101598737A (en) * 2009-06-26 2009-12-09 华东理工大学 Among the pure terephthalic acid to the flexible measurement method of carboxyl benzaldehyde content
CN101620590A (en) * 2009-07-06 2010-01-06 浙江大学 Parameter estimation method for multi-condition and large-scale chemical process model

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
姚科田等: "基于数据驱动技术和工艺机理模型的PTA生产过程软测量建模方法", 《计算机与应用化学》, vol. 27, no. 10, 28 October 2010 (2010-10-28) *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105210089A (en) * 2013-05-22 2015-12-30 惠普发展公司，有限责任合伙企业 Production simulation
CN108197743A (en) * 2017-12-31 2018-06-22 北京化工大学 A kind of prediction model flexible measurement method based on deep learning
CN113516200A (en) * 2021-07-30 2021-10-19 盛景智能科技(嘉兴)有限公司 Method and device for generating model training scheme, electronic equipment and storage medium
CN113516200B (en) * 2021-07-30 2024-06-04 盛景智能科技(嘉兴)有限公司 Model training scheme generation method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN110378044B (en) Multi-time scale convolution neural network soft measurement method based on attention mechanism
CN108682023A (en) Close coupling Unscented kalman tracking filter algorithm based on Elman neural networks
CN110023850A (en) Method and control device for control technology system
CN104899135B (en) Software Defects Predict Methods and system
CN116822380B (en) Collaborative optimization method for tail gas recycling in copper smelting process based on digital twin
CN103674778B (en) The industrial melt index soft measurement instrument of RBF particle group optimizing and method
CN108664009A (en) Divided stages based on correlation analysis and fault detection method
CN106663086A (en) Apparatus and method for ensembles of kernel regression models
CN110009020A (en) A kind of non-linear process monitoring method based on multicore Principal Component Analysis Model
CN107782857A (en) Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN
CN108579929B (en) A kind of double-in and double-out tube mill control system and control method based on RBF neural PREDICTIVE CONTROL
CN100470418C (en) Polypropylene melting index softsensoring instrument based on blind signal analysis and method thereof
WO2020175418A1 (en) Prediction control development device, prediction control development method, and prediction control development program
CN107341359A (en) Stalk fermentation produces the flexible measurement method of ethanol process key parameters
CN103279030B (en) Dynamic soft measuring modeling method and device based on Bayesian frame
CN102385706A (en) Soft sensing method based on model learning and used in pure terephthalic acid (PTA) production
CN101285816A (en) Copper matte air refining procedure parameter soft sensing instrument and its soft sensing method
CN113858566A (en) Injection molding machine energy consumption prediction method and system based on machine learning
CN108204997A (en) Normal line oil flash point online soft sensor method
CN103675012B (en) The industrial melt index soft measurement instrument of BP particle group optimizing and method
CN107766880A (en) The flexible measurement method of Fermentation by Photosynthetic Bacteria process key parameter based on BA LSSVM
CN107544286A (en) A kind of system identifying method in evaporization process
CN106599391A (en) Relevance vector machine soft measurement modeling method based on triangular angle value dynamic weighting
CN104133437B (en) Continuous-type chemical-engineering device and performance indicator real-time evaluation method and device thereof
CN109933031A (en) A kind of system and method automatically correcting soft measuring instrument according to analysis data

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C02 Deemed withdrawal of patent application after publication (patent law 2001)
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20120321