CN114035529A - ATL-BMA-based low-cost modeling method for nonlinear industrial process - Google Patents

ATL-BMA-based low-cost modeling method for nonlinear industrial process Download PDF

Info

Publication number
CN114035529A
CN114035529A CN202111411517.6A CN202111411517A CN114035529A CN 114035529 A CN114035529 A CN 114035529A CN 202111411517 A CN202111411517 A CN 202111411517A CN 114035529 A CN114035529 A CN 114035529A
Authority
CN
China
Prior art keywords
industrial process
data
old
new
modeling
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.)
Granted
Application number
CN202111411517.6A
Other languages
Chinese (zh)
Other versions
CN114035529B (en
Inventor
褚菲
朱安强
丁珮宽
陆宁云
熊刚
王军
王福利
马小平
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.)
China University of Mining and Technology CUMT
Original Assignee
China University of Mining and Technology CUMT
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 China University of Mining and Technology CUMT filed Critical China University of Mining and Technology CUMT
Priority to CN202111411517.6A priority Critical patent/CN114035529B/en
Publication of CN114035529A publication Critical patent/CN114035529A/en
Application granted granted Critical
Publication of CN114035529B publication Critical patent/CN114035529B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41885—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by modeling, simulation of the manufacturing system
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32339—Object oriented modeling, design, analysis, implementation, simulation language
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Landscapes

  • Engineering & Computer Science (AREA)
  • Manufacturing & Machinery (AREA)
  • General Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention provides a low-cost modeling method for a nonlinear industrial process based on ATL-BMA, which comprises the steps of selecting N groups of similar old process modeling data; collecting an initial data set for modeling a new process; dividing new and old process data into two parts respectively, and performing normalization processing respectively; converting the N groups of old process data into N groups of old process data with new process information, mixing the N groups of old process data with corresponding old process data to obtain N groups of mixed data sets, and then training a support vector machine model to obtain N groups of old process basic models with new process information; mapping the input variables of the new process training set to the operation intervals of the input variables of the similar old process, and obtaining the fusion output of the N prediction models; and (3) taking the fusion output of the SVM model in the old process and the input data of the new process as the input data of the multi-model migration strategy, and training to obtain a new process model. The method can effectively solve the problems of high modeling cost, limited acquired modeling data and long modeling period of the complex industrial process.

Description

ATL-BMA-based low-cost modeling method for nonlinear industrial process
Technical Field
The invention belongs to the technical field of industrial process construction performance prediction models, and particularly relates to a nonlinear industrial process low-cost modeling method based on ATL-BMA.
Background
In order to meet the requirements of the market on products with multiple specifications, multiple varieties and high quality, the modern industrial process is moving towards large-scale, high-efficiency and integration. On one hand, with the gradual expansion of the production scale, new industrial production processes can be continuously added in the actual production process to meet different product requirements, which leads to the higher and higher complexity of the actual industrial production process. On the other hand, changes in the operating environment and increases in the operating time cause changes in the characteristics of the actual industrial process. Both of these aspects result in the characteristic of process data being variable. In this case, a troublesome problem needs to be solved when modeling an industrial process using a data-driven approach: due to various factors such as cost and the like, modeling data obtained from a new industrial process is seriously insufficient, an accurate process prediction model cannot be established by using a data-driven modeling method under the support of a small amount of modeling data, and meanwhile, the generalization capability of the obtained model is low. In the face of this situation, it is desirable to have available data or knowledge of the industrial process that has a long runtime to assist in building predictive models of new industrial processes. Although the characteristics of the new and old industrial process operation data are different to some extent, the physicochemical mechanism followed in the process is unchanged or very similar, so that the new industrial process data and the old industrial process data have the same or similar characteristic space and label space (the two input and output data dimensions are consistent). As shown in FIG. 1, a new industrial process and an old industrial process can be considered as a target domain and a source domain, respectively, and then a new industrial process prediction model can be established by using old industrial process data through a migration learning method. However, when the source domain data is much larger than the target domain data, a phenomenon of "negative migration" easily occurs when the target domain data is subjected to supplementary learning by using the source domain data under the conventional migration learning structure.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a low-cost modeling method for a nonlinear industrial process based on ATL-BMA, which can effectively solve the problems of high modeling cost, limited obtained modeling data and long modeling period of a complex industrial process, can solve the phenomenon of negative migration which possibly occurs when old process data is far more than new process data in migration learning, can fully utilize the information of the existing similar old industrial process model to assist in guiding and building a prediction model of the new industrial process, can effectively reduce the modeling cost, can accelerate the modeling speed and improve the modeling precision.
In order to achieve the aim, the invention provides a low-cost modeling method of a nonlinear industrial process based on ATL-BMA, which comprises the following steps:
step 1: selecting N groups of modeling data similar to the old industrial process, and determining the stable operation range of the input variable according to the information of the actual process to be modeled; simultaneously, selecting a Latin hypercube method for sampling and collecting a target nonlinear industrial process modeling initial data set; the method comprises the following specific steps:
step 1.1: selecting N groups of modeling data of similar old industrial processes and recording the data as
Figure BDA0003373986810000021
Modeling the ith old industrial process according to formula (1);
Figure BDA0003373986810000022
wherein X and X are old industrial process input data, y is old industrial process output data, kiRepresenting the modeling data quantity of the ith old industrial process, wherein n is the input variable dimension of the ith old industrial process, and the input variable dimensions of all the industrial processes are consistent and are n because the new industrial process and the old industrial process have certain similarity;
step 1.2: according to the factDetermining the stable operation range of input variables according to the information of the inter-process to be modeled, selecting discrete sparse data distribution points for sampling and collecting new industrial process modeling data, and acquiring collected new industrial process data D according to a formula (2)new；
Figure BDA0003373986810000023
In the formula, l represents the modeling data volume of the new industrial process;
step 2: dividing new industrial process data and old industrial process data into two parts respectively, namely a training data set and a testing data set in a new modeling process and an old modeling process, and respectively carrying out normalization processing on the new industrial process initial data set and the old industrial process modeling data; wherein, for the new industrial process data, the new industrial process data is divided into a new industrial process training data set
Figure BDA0003373986810000024
And new industrial process test data set
Figure BDA0003373986810000025
And maps the data to [0,1 ] using equation (3)]An interval;
Figure BDA0003373986810000031
in the formula, ziRepresenting the result after normalization of input or output data of an industrial process, xiIs the data before normalization, xmaxIs the maximum value, x, before data normalizationminIs the minimum value;
and step 3: converting N groups of old industrial process data into N groups of old industrial process data with new industrial process information by using a Cycle GANs-based new and old industrial process data migration algorithm; wherein the old industrial process training data set is
Figure BDA0003373986810000032
It is concretelyThe method comprises the following steps:
step 3.1: initializing parameters: g parameter thetaG，DoParameter omegaoF parameter θF，DnParameter omegan，ncritic＝5，α＝0.00005、β1＝0、β2＝0.7，m＝5，λ＝0.5，Epoch＝20000；
Wherein: g denotes a generator function of old to new industrial process data, DoA discriminator representing the correspondence of the old industrial process, F a generator function representing the new industrial process to the old industrial process, DnDiscriminators, n, corresponding to new industrial processescriticRepresenting the number of times of training the discriminant model after training the generator once, alpha, beta1And beta2The parameters of the Adam optimizer are shown, m is the sampling number, and Epoch is the number of times of model cyclic training;
step 3.2: will be derived from the ith old industrial process data by the generator G
Figure BDA0003373986810000033
M samples collected in
Figure BDA0003373986810000034
Converted into m new industrial process data, denoted Xo→n＝F(Xo) (ii) a Will be derived from new industrial process data by the generator F
Figure BDA0003373986810000035
M samples collected in
Figure BDA0003373986810000036
Converted into m old industrial process data, denoted Xn→o＝F(Xn)；
Step 3.3: obtaining the loss of the discriminator and the consistent loss of the two forward circulations according to a formula (4) and a formula (5);
Figure BDA0003373986810000037
Figure BDA0003373986810000038
step 3.4: updating the discriminator D by the formula (6) and the formula (7)oParameter omegaoAnd DnParameter omegan；
Figure BDA0003373986810000039
Figure BDA0003373986810000041
Step 3.5: repeating the step 3.2 to the step 3.4ncriticSecondly;
step 3.6: repeating the step 3.2;
step 3.7: calculating two forward cyclic consistent losses through formula (8) and formula (9);
Figure BDA0003373986810000042
Figure BDA0003373986810000043
step 3.8: calculating two generator losses by equation (10) and equation (11);
Figure BDA0003373986810000044
Figure BDA0003373986810000045
step 3.9: updating generator G parameter θ by equation (12) and equation (13)GAnd F parameter thetaF；
Figure BDA0003373986810000046
Figure BDA0003373986810000047
Step 3.10: repeating the steps 3.6 to 3.9Epoch times, converting the new industrial process data into the jth old industrial process data by using the trained F, and recording the jth old industrial process data as
Figure BDA0003373986810000048
Step 3.11: repeating the steps 3.1-3.9 by using each group of old industrial process data, transferring the new industrial process data into the old industrial process domain, obtaining N groups of old industrial process data with new industrial process information by the new industrial process data through antagonistic transfer learning, and recording the N groups of old industrial process data with the new industrial process information as
Figure BDA0003373986810000049
And 4, step 4: mixing the old industrial process data with the new industrial process information in the step 3 with the corresponding old industrial process data to obtain N groups of mixed data sets;
and 5: splitting a hybrid data set into a hybrid training set
Figure BDA00033739868100000410
And mixing the test data sets
Figure BDA00033739868100000411
At the same time, N old industrial process training data sets are combined
Figure BDA0003373986810000051
And a new industrial process prediction model y ═ f (x), training a Support Vector Machine (SVM) model respectively by utilizing N groups of mixed data sets to obtain N old industrial process basic models with new industrial process information, and recording the N old industrial process basic models as f1(·)-fN(·); wherein the content of the first and second substances,
Figure BDA0003373986810000052
ktrainis the size of the training data set and,
Figure BDA0003373986810000053
ktestis the test data set size, any ith old industrial process,
Figure BDA0003373986810000054
niis the ith old industrial process training set size; the method comprises the following specific steps:
step 5.1: initializing parameters;
step 5.2: new industrial process data are converted into N groups of old industrial process data carrying new industrial process information through a Cycle GANs-based new and old industrial process data migration algorithm
Figure BDA0003373986810000055
Mixing D according to formula (14)n→oAnd DoObtaining N groups of basic model training data DBasic；
Figure BDA0003373986810000056
Step 5.3: by using DBasicTraining N SVM to obtain N old industrial process basic models with new industrial process information, and recording as f1(·)-fN(·)；
Step 6: mapping input variables of the new industrial process training set to the operation interval of the input variables of the similar old industrial process through a model fusion formula (15), and recording input data of the converted new industrial process training set as
Figure BDA0003373986810000057
Obtaining the fusion output of the N prediction models by the Bayesian model average algorithm
Figure BDA0003373986810000058
Figure BDA0003373986810000059
And 7: fusing and outputting an old industrial process SVM model
Figure BDA00033739868100000510
And new industrial process input data
Figure BDA00033739868100000511
As input data of a multi-model migration strategy, training a new industrial process model by utilizing a least square support vector machine algorithm to obtain output of the new industrial process model
Figure BDA00033739868100000512
Completing the modeling of a new industrial process;
and 8: model verification, namely evaluating the effectiveness of the SVM model by utilizing the root mean square error and the determination coefficient according to a formula (16) and a formula (17), and finishing the modeling process if the prediction precision of the model obtained in the step 7 on the test data set meets an experiment set threshold; otherwise, repeating the step 3 to the step 7, adding N new groups of old industrial process data samples containing new industrial process information into the mixed sample, and continuing training the new industrial process model until the experiment stopping condition is met;
Figure BDA0003373986810000061
Figure BDA0003373986810000062
where N is the number of test data, yiIs the output of the predictive model and is,
Figure BDA0003373986810000063
is the mean of the predicted outputs, YiIs the real output of the new industrial process.
The method comprises the steps of firstly, collecting a small sample data set for modeling the nonlinear industrial process by using a Latin hypercube method, combining a plurality of similar old process data, and learning a conversion mapping function between new industrial process data and old industrial process data by a antagonism migration algorithm, so that a small amount of new process data is converted into a plurality of types of old industrial process data with new industrial process information; then, a plurality of 'old process models with new process information' are obtained through a regression algorithm of a support vector machine, and a foundation is established for the modeling of a subsequent new industrial process; and finally, migrating a plurality of trained 'old industrial process prediction models with new industrial process information' by using a multi-model migration strategy and a Bayesian model average theory, and combining a small amount of new industrial process data to obtain a final new industrial process performance prediction model. The method provided by the invention migrates useful information of a plurality of existing similar old industrial processes to help establish a new industrial process performance prediction model, thereby reducing the modeling cost of the new industrial process; meanwhile, in order to effectively solve the problem of negative migration which is possibly caused when the data of the old process is much more than the data of the new process, a new and old process data migration method based on the anti-migration learning is adopted, and the migration modeling effect is improved. The method effectively solves the problems of high modeling cost and long modeling period of the complex industrial process, fully utilizes useful information of the existing similar old industrial process model, simultaneously solves the phenomenon of negative migration which possibly occurs when the old process data is far more than the new process data in the migration learning, completes the modeling of the new industrial process, reduces the modeling cost, accelerates the modeling speed and improves the modeling precision.
Drawings
FIG. 1 is a flow chart of migration modeling;
FIG. 2 is a flow diagram of a non-linear industrial process low cost modeling method based on anti-migratory learning and Bayesian model averaging theory;
FIG. 3 is a graph of predicted values of the ATL-BMA model, the BMA model, and the SVM model on the compressor A test set;
FIG. 4 is an RMSE histogram of predicted values and true values for the ATL-BMA model, the BMA model, and the SVM model;
FIG. 5 isR of predicted value and true value of ATL-BMA model, BMA model and SVM model2A histogram.
Detailed Description
The invention is further illustrated by the following examples and figures.
As shown in fig. 1 to 5, the present invention provides a low-cost modeling method for nonlinear industrial process based on ATL-BMA (adaptive transfer Learning (ATL)) and Bayesian Model Averaging (BMA)), including the following steps:
step 1: selecting N groups of modeling data similar to the old industrial process, and determining the stable operation range of the input variable according to the information of the actual process to be modeled; meanwhile, selecting a Latin Hypercube Design (LHD) method for sampling and collecting a target nonlinear industrial process (new industrial process) modeling initial data set; the information of the actual process to be modeled comprises a parameter rated value, a performance curve and the like; the method comprises the following specific steps:
step 1.1: selecting N groups of modeling data of similar old industrial processes and recording the data as
Figure BDA0003373986810000071
Modeling the ith old industrial process according to formula (1);
Figure BDA0003373986810000072
wherein X is the old industrial process input data set, X is the old industrial process input data, y is the old industrial process output data, kiRepresenting the modeling data quantity of the ith old industrial process, wherein n is the input variable dimension of the ith old industrial process, and the input variable dimensions of all the industrial processes are consistent and are n because the new industrial process and the old industrial process have certain similarity;
step 1.2: determining the stable operation range of input variables according to the information of the actual process to be modeled, selecting discrete sparse data distribution points for sampling and collecting new industrial process modeling data, and acquiring acquisition data according to a formula (2)New industrial process data Dnew；
Figure BDA0003373986810000073
In the formula, l represents the modeling data volume of the new industrial process;
step 2: dividing new industrial process data and old industrial process data into two parts respectively, namely a training data set and a testing data set in a new modeling process and an old modeling process; in order to stabilize the subsequent training process and avoid adverse effects caused by data dimension difference, the data must be normalized, and the initial data set of the new industrial process and the modeling data of the old industrial process are respectively normalized; wherein, for the new industrial process data, the new industrial process data is divided into a new industrial process training data set
Figure BDA0003373986810000081
And new industrial process test data set
Figure BDA0003373986810000082
And mapping the data to [0,1 ] by using a maximum value and minimum value data normalization method according to formula (3)]An interval;
Figure BDA0003373986810000083
in the formula, ziRepresenting the result after normalization of input or output data of an industrial process, xiIs the data before normalization, xmaxIs the maximum value, x, before data normalizationminIs the minimum value;
and step 3: converting N groups of old industrial process data into N groups of old industrial process data with new industrial process information by using a Cycle GANs-based new and old industrial process data migration algorithm; wherein the old industrial process training data set is
Figure BDA0003373986810000084
The method comprises the following specific steps:
step 3.1: initializing parameters: g parameter thetaG，DoParameter omegaoF parameter θF，DnParameter omegan，ncritic＝5，α＝0.00005、β1＝0、β2＝0.7，m＝5，λ＝0.5，Epoch＝20000；
Wherein: g denotes a generator function of old to new industrial process data, DoA discriminator representing the correspondence of the old industrial process, F a generator function representing the new industrial process to the old industrial process, DnDiscriminators, n, corresponding to new industrial processescriticRepresenting the number of times of training the discriminant model after training the generator once, alpha, beta1And beta2The parameters of the Adam optimizer are shown, m is the sampling number, and Epoch is the number of times of model cyclic training;
step 3.2: will be derived from the ith old industrial process data by the generator G
Figure BDA0003373986810000085
M samples collected in
Figure BDA0003373986810000086
Converted into m new industrial process data, denoted Xo→n＝F(Xo) (ii) a Will be derived from new industrial process data by the generator F
Figure BDA0003373986810000087
M samples collected in
Figure BDA0003373986810000088
Converted into m old industrial process data, denoted Xn→o＝F(Xn)；
Step 3.3: obtaining the loss of the discriminator and the consistent loss of the two forward circulations according to a formula (4) and a formula (5);
Figure BDA0003373986810000091
Figure BDA0003373986810000092
step 3.4: updating the discriminator D by the formula (6) and the formula (7)oParameter omegaoAnd DnParameter omegan；
Figure BDA0003373986810000093
Figure BDA0003373986810000094
Step 3.5: repeating the step 3.2 to the step 3.4ncriticSecondly;
step 3.6: repeating the step 3.2;
step 3.7: calculating two forward cyclic consistent losses through formula (8) and formula (9);
Figure BDA0003373986810000095
Figure BDA0003373986810000096
step 3.8: calculating two generator losses by equation (10) and equation (11);
Figure BDA0003373986810000097
Figure BDA0003373986810000098
step 3.9: updating generator G parameter θ by equation (12) and equation (13)GAnd F parameter thetaF；
Figure BDA0003373986810000099
Figure BDA00033739868100000910
Step 3.10: repeating the steps 3.6 to 3.9Epoch times, converting the new industrial process data into the jth old industrial process data by using the trained F, and recording the jth old industrial process data as
Figure BDA00033739868100000911
Step 3.11: repeating the steps 3.1-3.9 by using each group of old industrial process data, transferring the new industrial process data into the old industrial process domain, obtaining N groups of old industrial process data with new industrial process information by the new industrial process data through antagonistic transfer learning, and recording the N groups of old industrial process data with the new industrial process information as
Figure BDA0003373986810000101
And 4, step 4: mixing the old industrial process data with the new industrial process information in the step 3 with the corresponding old industrial process data to obtain N groups of mixed data sets;
and 5: splitting a hybrid data set into a hybrid training set
Figure BDA0003373986810000102
And mixing the test data sets
Figure BDA0003373986810000103
At the same time, N old industrial process training data sets are combined
Figure BDA0003373986810000104
And a new industrial process prediction model y ═ f (x), respectively training a Support Vector Machine (SVM) (support Vector machine) model by utilizing N groups of mixed data sets to obtain N old industrial process basic models with new industrial process information, and recording as f1(·)-fN(·); wherein the content of the first and second substances,
Figure BDA0003373986810000105
ktrainis the size of the training data set and,
Figure BDA0003373986810000106
ktestis the test data set size, any ith old industrial process,
Figure BDA0003373986810000107
niis the ith old industrial process training set size; the method comprises the following specific steps:
step 5.1: initializing parameters;
step 5.2: new industrial process data are converted into N groups of old industrial process data carrying new industrial process information through a Cycle GANs-based new and old industrial process data migration algorithm
Figure BDA0003373986810000108
Mixing D according to formula (14)n→oAnd DoObtaining N groups of basic model training data DBasic；
Figure BDA0003373986810000109
Step 5.3: by using DBasicTraining N SVM to obtain N old industrial process basic models with new industrial process information, and recording as f1(·)-fN(·)；
Step 6: mapping input variables of the new industrial process training set to the operation interval of the input variables of the similar old industrial process through a model fusion formula (15), and recording input data of the converted new industrial process training set as
Figure BDA00033739868100001010
Obtaining the fusion output of the N prediction models by the Bayesian model average algorithm
Figure BDA0003373986810000111
Figure BDA0003373986810000112
And 7: fusing and outputting an old industrial process SVM model
Figure BDA0003373986810000113
And new industrial process input data
Figure BDA0003373986810000114
As input data of a multi-model migration strategy, a new industrial process model is trained by utilizing a Least Square Support Vector Machine (LSSVM) algorithm to obtain output of the new industrial process model
Figure BDA0003373986810000115
Completing the modeling of a new industrial process;
and 8: model verification using Root Mean Square Error (RMSE) and deterministic coefficient (R-Square, R) according to equation (16) and equation (17), respectively2) Evaluating the effectiveness of the SVM model, and if the prediction precision of the model obtained in the step 7 on the test data set meets an experiment set threshold, finishing the modeling process; otherwise, repeating the step 3 to the step 7, adding N new groups of old industrial process data samples containing new industrial process information into the mixed sample, and continuing training the new industrial process model until the experiment stopping condition is met;
Figure BDA0003373986810000116
Figure BDA0003373986810000117
where N is the number of test data, yiIs the output of the predictive model and is,
Figure BDA0003373986810000118
is the mean of the predicted outputs, YiIs the real output of the new industrial process.
The method comprises the steps of firstly, collecting a small sample data set for modeling the nonlinear industrial process by using a Latin hypercube method, combining a plurality of similar old process data, and learning a conversion mapping function between new industrial process data and old industrial process data by a antagonism migration algorithm, so that a small amount of new process data is converted into a plurality of types of old industrial process data with new industrial process information; then, a plurality of 'old process models with new process information' are obtained through a regression algorithm of a support vector machine, and a foundation is established for the modeling of a subsequent new industrial process; and finally, migrating a plurality of trained 'old industrial process prediction models with new industrial process information' by using a multi-model migration strategy and a Bayesian model average theory, and combining a small amount of new industrial process data to obtain a final new industrial process performance prediction model. The method provided by the invention migrates useful information of a plurality of existing similar old industrial processes to help establish a new industrial process performance prediction model, thereby reducing the modeling cost of the new industrial process; meanwhile, in order to effectively solve the problem of negative migration which is possibly caused when the data of the old process is much more than the data of the new process, a new and old process data migration method based on the anti-migration learning is adopted, and the migration modeling effect is improved. The method effectively solves the problems of high modeling cost and long modeling period of the complex industrial process, fully utilizes useful information of the existing similar old industrial process model, simultaneously solves the phenomenon of negative migration which possibly occurs when the old process data is far more than the new process data in the migration learning, completes the modeling of the new industrial process, reduces the modeling cost, accelerates the modeling speed and improves the modeling precision.
In order to verify the effectiveness of the method, experimental data is generated by using a laboratory centrifugal compressor mechanism model, and a performance prediction model of the centrifugal compressor is established to verify the effectiveness of the modeling method. A, B, C, D four different but similar compressor models were generated for simulation experiments by modifying the key geometry simulation of the compressor mechanics model. For the A, B, C and D four centrifugal compressors, where compressor A was the new compressor to be modeled, a small amount of new industrial process modeling data was generated, while the B, C and D centrifugal compressors were the old compressors with long run times, and a large amount of old industrial process modeling data was generated to assist in the creation of the new industrial process prediction model. The stable movement interval of the old and new compressors is shown in table 1.
TABLE 1 Stable operation interval of centrifugal compressor A, B, C, D and corresponding One-Hot code
Figure BDA0003373986810000121
And comparing the prediction effect of the established model with the prediction effects of the two groups of comparison experiment models, and further showing the superiority of the method. The three groups of comparison methods are specifically as follows:
the method comprises the following steps: converting a small amount of new industrial process data into old industrial process data through anti-migration learning, mixing the old industrial process data with each group of old industrial process data, training to obtain a plurality of SVM models, then establishing a new compressor prediction model through a multi-model migration strategy, and finally testing the model precision by using new compressor test data. The method is recorded as an ATL-BMA method in the analysis of experimental results.
The method 2 comprises the following steps: the method comprises the steps of training a plurality of old compressor SVM models by using a plurality of groups of old compressor data, establishing a new compressor prediction model by combining a plurality of new compressor training data through a multi-model migration strategy, and finally testing the model precision by using new compressor testing data. The BMA method is recorded in the analysis of experimental results.
The method 3 comprises the following steps: and establishing a new compressor SVM model by using only a small amount of new compressor training data, taking the model as a new compressor prediction model, and finally testing the model precision by using new compressor test data. And marking the test result as an SVM method in the analysis of the test result.
FIG. 3 shows the predicted values of the model of the three methods on the compressor A test set. It can be seen from the figure that the predicted value of the model established by the ATL-BMA method is the highest in coincidence degree with the test set, which shows that the ATL-BMA method can effectively utilize the useful information of the similar old industrial process to help the establishment of the new industrial process model, and simultaneously shows that the ATL-BMA method can more effectively utilize the information between the old industrial process and the new industrial process than a simple multi-model migration method.
To further compare the accuracy of the three models, FIGS. 4 and 5 show the RMSE and R of the predicted versus true values of the three models2As can be seen from the figure, the method provided in this chapter can fully utilize a large amount of existing old industrial process data and a small amount of new industrial process data, effectively improve the prediction accuracy of the model, and reduce the cost for establishing the model.
The method adopts an anti-migration learning method and a Bayes model average theory to establish a performance prediction model for the new industrial process, makes full use of the existing performance prediction model similar to the old industrial process in the industry, migrates the new and old industrial process data, establishes a plurality of old industrial process prediction models containing new industrial process information by using a support vector machine, and trains the old industrial process model by using the Bayes model average theory, thereby accelerating the modeling speed of the new industrial process, reducing the modeling cost, solving the negative migration effect caused by the fact that the old industrial process data is more than the new industrial process data when the new and old industrial process migrates the model, and obtaining the prediction model meeting the precision requirement. Meanwhile, the method can more effectively utilize the information between the old industrial process and the new industrial process than a pure multi-model migration method. And the method is closer to actual output, and a large amount of cost is reduced for industrial process modeling.

Claims (1)

1. A low-cost modeling method for a nonlinear industrial process based on ATL-BMA is characterized by comprising the following steps:
step 1: selecting N groups of modeling data similar to the old industrial process, and determining the stable operation range of the input variable according to the information of the actual process to be modeled; simultaneously, selecting a Latin hypercube method for sampling and collecting a target nonlinear industrial process modeling initial data set; the method comprises the following specific steps:
step 1.1: selecting N groups of modeling data of similar old industrial processes and recording the data as
Figure FDA0003373986800000011
Modeling the ith old industrial process according to formula (1);
Figure FDA0003373986800000012
wherein X and X are old industrial process input data, y is old industrial process output data, kiRepresenting the modeling data quantity of the ith old industrial process, wherein n is the input variable dimension of the ith old industrial process, and the input variable dimensions of all the industrial processes are consistent and are n because the new industrial process and the old industrial process have certain similarity;
step 1.2: determining the stable operation range of input variables according to the information of the actual process to be modeled, selecting discrete sparse data distribution points for sampling and collecting new industrial process modeling data, and obtaining the collected new industrial process data D according to the formula (2)new；
Figure FDA0003373986800000013
In the formula, l represents the modeling data volume of the new industrial process;
step 2: dividing new industrial process data and old industrial process data into two parts respectively, namely a training data set and a testing data set in a new modeling process and an old modeling process, and respectively carrying out normalization processing on the new industrial process initial data set and the old industrial process modeling data; wherein, for the new industrial process data, the new industrial process data is divided into a new industrial process training data set
Figure FDA0003373986800000014
And new industrial process test data set
Figure FDA0003373986800000015
And maps the data to [0,1 ] using equation (3)]An interval;
Figure FDA0003373986800000021
in the formula, ziRepresenting the result after normalization of input or output data of an industrial process, xiIs the data before normalization, xmaxIs the maximum value, x, before data normalizationminIs the minimum value;
and step 3: converting N groups of old industrial process data into N groups of old industrial process data with new industrial process information by using a Cycle GANs-based new and old industrial process data migration algorithm; wherein the old industrial process training data set is
Figure FDA0003373986800000029
The method comprises the following specific steps:
step 3.1: initializing parameters: g parameter thetaG，DoParameter omegaoF parameter θF，DnParameter omegan，ncritic＝5，α＝0.00005、β1＝0、β2＝0.7，m＝5，λ＝0.5，Epoch＝20000；
Wherein: g denotes a generator function of old to new industrial process data, DoA discriminator representing the correspondence of the old industrial process, F a generator function representing the new industrial process to the old industrial process, DnDiscriminators, n, corresponding to new industrial processescriticRepresenting the number of times of training the discriminant model after training the generator once, alpha, beta1And beta2The parameters of the Adam optimizer are shown, m is the sampling number, and Epoch is the number of times of model cyclic training;
step 3.2: will be derived from the ith old industrial process data by the generator G
Figure FDA0003373986800000022
M samples collected in
Figure FDA0003373986800000023
Is converted into m new industrial process data,is marked as Xo→n＝F(Xo) (ii) a Will be derived from new industrial process data by the generator F
Figure FDA0003373986800000024
M samples collected in
Figure FDA0003373986800000025
Converted into m old industrial process data, denoted Xn→o＝F(Xn)；
Step 3.3: obtaining the loss of the discriminator and the consistent loss of the two forward circulations according to a formula (4) and a formula (5);
Figure FDA0003373986800000026
Figure FDA0003373986800000027
step 3.4: updating the discriminator D by the formula (6) and the formula (7)oParameter omegaoAnd DnParameter omegan；
Figure FDA0003373986800000028
Figure FDA0003373986800000031
Step 3.5: repeating the step 3.2 to the step 3.4ncriticSecondly;
step 3.6: repeating the step 3.2;
step 3.7: calculating two forward cyclic consistent losses through formula (8) and formula (9);
Figure FDA0003373986800000032
Figure FDA0003373986800000033
step 3.8: calculating two generator losses by equation (10) and equation (11);
Figure FDA0003373986800000034
Figure FDA0003373986800000035
step 3.9: updating generator G parameter θ by equation (12) and equation (13)GAnd F parameter thetaF；
Figure FDA0003373986800000036
Figure FDA0003373986800000037
Step 3.10: repeating the steps 3.6 to 3.9Epoch times, converting the new industrial process data into the jth old industrial process data by using the trained F, and recording the jth old industrial process data as
Figure FDA0003373986800000038
Step 3.11: repeating the steps 3.1-3.9 by using each group of old industrial process data, transferring the new industrial process data into the old industrial process domain, obtaining N groups of old industrial process data with new industrial process information by the new industrial process data through antagonistic transfer learning, and recording the N groups of old industrial process data with the new industrial process information as
Figure FDA0003373986800000039
And 4, step 4: mixing the old industrial process data with the new industrial process information in the step 3 with the corresponding old industrial process data to obtain N groups of mixed data sets;
and 5: splitting a hybrid data set into a hybrid training set
Figure FDA00033739868000000310
And mixing the test data sets
Figure FDA00033739868000000311
At the same time, N old industrial process training data sets are combined
Figure FDA0003373986800000041
And a new industrial process prediction model y ═ f (x), training a Support Vector Machine (SVM) model respectively by utilizing N groups of mixed data sets to obtain N old industrial process basic models with new industrial process information, and recording the N old industrial process basic models as f1(·)-fN(·); wherein the content of the first and second substances,
Figure FDA0003373986800000042
ktrainis the size of the training data set and,
Figure FDA0003373986800000043
ktestis the test data set size, any ith old industrial process,
Figure FDA0003373986800000044
niis the ith old industrial process training set size; the method comprises the following specific steps:
step 5.1: initializing parameters;
step 5.2: new industrial process data are converted into N groups of old industrial process data carrying new industrial process information through a Cycle GANs-based new and old industrial process data migration algorithm
Figure FDA0003373986800000045
Mixing D according to formula (14)n→oAnd DoObtaining N groups of basic model training data DBasic；
Figure FDA0003373986800000046
Step 5.3: by using DBasicTraining N SVM to obtain N old industrial process basic models with new industrial process information, and recording as f1(·)-fN(·)；
Step 6: mapping input variables of the new industrial process training set to the operation interval of the input variables of the similar old industrial process through a model fusion formula (15), and recording input data of the converted new industrial process training set as
Figure FDA0003373986800000047
Obtaining the fusion output of the N prediction models by the Bayesian model average algorithm
Figure FDA0003373986800000048
Figure FDA0003373986800000049
And 7: fusing and outputting an old industrial process SVM model
Figure FDA00033739868000000410
And new industrial process input data
Figure FDA00033739868000000411
As input data of a multi-model migration strategy, training a new industrial process model by utilizing a least square support vector machine algorithm to obtain output of the new industrial process model
Figure FDA00033739868000000412
Completing new industrial process modeling；
And 8: model verification, namely evaluating the effectiveness of the SVM model by utilizing the root mean square error and the determination coefficient according to a formula (16) and a formula (17), and finishing the modeling process if the prediction precision of the model obtained in the step 7 on the test data set meets an experiment set threshold; otherwise, repeating the step 3 to the step 7, adding N new groups of old industrial process data samples containing new industrial process information into the mixed sample, and continuing training the new industrial process model until the experiment stopping condition is met;
Figure FDA0003373986800000051
Figure FDA0003373986800000052
where N is the number of test data, yiIs the output of the predictive model and is,
Figure FDA0003373986800000053
is the mean of the predicted outputs, YiIs the real output of the new industrial process.
CN202111411517.6A 2021-11-25 2021-11-25 ATL-BMA-based nonlinear industrial process low-cost modeling method Active CN114035529B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111411517.6A CN114035529B (en) 2021-11-25 2021-11-25 ATL-BMA-based nonlinear industrial process low-cost modeling method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111411517.6A CN114035529B (en) 2021-11-25 2021-11-25 ATL-BMA-based nonlinear industrial process low-cost modeling method

Publications (2)

Publication Number Publication Date
CN114035529A true CN114035529A (en) 2022-02-11
CN114035529B CN114035529B (en) 2023-09-08

Family

ID=80138766

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111411517.6A Active CN114035529B (en) 2021-11-25 2021-11-25 ATL-BMA-based nonlinear industrial process low-cost modeling method

Country Status (1)

Country Link
CN (1) CN114035529B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103500281A (en) * 2013-10-08 2014-01-08 广西大学 Nonlinear system modeling method of process of sugar boiling crystallization
CN106156434A (en) * 2016-07-11 2016-11-23 江南大学 Sliding window time difference Gaussian process regression modeling method based on the low and deep structure of local time
CN109902378A (en) * 2019-02-25 2019-06-18 中国矿业大学 Complex industrial process low cost modeling method based on multi-model migration and BMA theory
CN112750277A (en) * 2021-01-05 2021-05-04 武汉大学 Indoor falling detection system and method fusing track data and sensor posture

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103500281A (en) * 2013-10-08 2014-01-08 广西大学 Nonlinear system modeling method of process of sugar boiling crystallization
CN106156434A (en) * 2016-07-11 2016-11-23 江南大学 Sliding window time difference Gaussian process regression modeling method based on the low and deep structure of local time
CN109902378A (en) * 2019-02-25 2019-06-18 中国矿业大学 Complex industrial process low cost modeling method based on multi-model migration and BMA theory
CN112750277A (en) * 2021-01-05 2021-05-04 武汉大学 Indoor falling detection system and method fusing track data and sensor posture

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
FEI CHU: "A Minimum-Cost Modeling Method for Nonlinear Industrial Process Based on Multimodel Migration and Bayesian Model Averaging Method", 《IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING 》, vol. 17, no. 2, XP011782784, DOI: 10.1109/TASE.2019.2952376 *

Also Published As

Publication number Publication date
CN114035529B (en) 2023-09-08

Similar Documents

Publication Publication Date Title
CN112149316B (en) Aero-engine residual life prediction method based on improved CNN model
CN108897286B (en) Fault detection method based on distributed nonlinear dynamic relation model
CN108647272B (en) Method for predicting concentration of butane at bottom of debutanizer by expanding small samples based on data distribution
CN108596327B (en) Seismic velocity spectrum artificial intelligence picking method based on deep learning
CN104699894A (en) JITL (just-in-time learning) based multi-model fusion modeling method adopting GPR (Gaussian process regression)
CN112084727A (en) Transition prediction method based on neural network
CN116448419A (en) Zero sample bearing fault diagnosis method based on depth model high-dimensional parameter multi-target efficient optimization
CN104462850A (en) Multi-stage batch process soft measurement method based on fuzzy gauss hybrid model
CN113505477B (en) Process industry soft measurement data supplementing method based on SVAE-WGAN
CN111832839B (en) Energy consumption prediction method based on sufficient incremental learning
CN111310990A (en) Improved gray combination model-based track quality prediction method and system
CN112947300A (en) Virtual measuring method, system, medium and equipment for processing quality
CN111325264A (en) Multi-label data classification method based on entropy
CN106990768B (en) MKPCA batch process fault monitoring method based on Limited-DTW
CN110223342B (en) Space target size estimation method based on deep neural network
CN113255225B (en) Train motion state estimation method for few-sample-element lifting learning
CN110889207A (en) System combination model credibility intelligent evaluation method based on deep learning
CN114239397A (en) Soft measurement modeling method based on dynamic feature extraction and local weighted deep learning
CN115062528A (en) Prediction method for industrial process time sequence data
CN108734207B (en) Method for predicting concentration of butane at bottom of debutanizer tower based on model of double-optimization semi-supervised regression algorithm
CN112966429A (en) Non-linear industrial process modeling method based on WGANs data enhancement
CN114035529A (en) ATL-BMA-based low-cost modeling method for nonlinear industrial process
CN112947080A (en) Scene parameter transformation-based intelligent decision model performance evaluation system
CN109902378B (en) Complex industrial process low-cost modeling method based on multi-model migration and BMA theory
Yun et al. Quality evaluation and satisfaction analysis of online learning of college students based on artificial intelligence

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant