CN106709570A - Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method - Google Patents

Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method Download PDF

Info

Publication number
CN106709570A
CN106709570A CN201611236844.1A CN201611236844A CN106709570A CN 106709570 A CN106709570 A CN 106709570A CN 201611236844 A CN201611236844 A CN 201611236844A CN 106709570 A CN106709570 A CN 106709570A
Authority
CN
China
Prior art keywords
sample
test
learning machine
extreme learning
time dimension
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
CN201611236844.1A
Other languages
Chinese (zh)
Inventor
葛志强
李雨绅
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 CN201611236844.1A priority Critical patent/CN106709570A/en
Publication of CN106709570A publication Critical patent/CN106709570A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/10—Interfaces, programming languages or software development kits, e.g. for simulating neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)

Abstract

The invention discloses a time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method. A local weighting extreme learning machine-based method aims at a nonlinear process and has good intensive reading performance and relatively high calculation speed, but a data quantity of an industrial process is often short of generalization performance when an extreme learning machine is applied. An applied improvement method is that a sample is subjected to time dimension expansion in time dimension by adopting a just-in-time learning thought, and local weighting is performed in a new space, thereby aiming at a dynamic characteristic of the industrial process. According to the method, the advantages and the main shortcomings of the just-in-time learning thought and the extreme learning machine are combined and overcome, so that a high-speed high-precision industrial process soft measurement modeling method is obtained.

Description

A kind of soft sensor modeling that local weighted extreme learning machine is expanded based on time dimension Method
Technical field
The local weighted limit is expanded the invention belongs to industrial process prediction and control field, more particularly to a kind of time dimension The soft-measuring modeling method of learning machine.
Background technology
In traditional industrial process, exist and many for improve product quality and ensure that safety has and most important make used Parameter such as reaction rate, product composition content etc., but be much often difficult to or directly cannot be measured with sensor. Using needing the on-line analysis instrument of great amount of investment to be detected, often have larger delayed and cause that regulation is not prompt enough, from And make product quality be difficult to be guaranteed.We are called leading variable for the variable that industrial process plays an important roll, It is other some with the variable of measurement we term it auxiliary variable.Hard measurement essence thing by set up industrial process variable it Between Mathematical Modeling, realize by auxiliary variable prediction leading variable technical method.In recent years, hard measurement is in industrial process In apply many neural network algorithms, but due to neural network algorithm, to still suffer from convergence rate slower, is easily trapped into part The shortcomings of optimal solution and poor Generalization Capability.
For the above mentioned problem for overcoming neural network algorithm to exist, extreme learning machine is used as a kind of single hidden layer stochastic neural net Network, the convergence rate that the inverse iteration of traditional neural network algorithm can be overcome to produce is slow, is absorbed in locally optimal solution problem.But due to The randomness and hidden node number of its model sacrifice Generalization Capability commonly greater than other neutral nets.Asked to solve this Topic, it is proposed that local weighted extreme learning machine model, using the thought of instant learning, is expanded local weighted by time dimension Method, improve extreme learning machine model as a kind of nonlinear method be applied to industrial process extreme learning machine generalization ability compared with Poor shortcoming, for the problem of industrial process data dynamic, improves the performance that model is set up and predicted.
The content of the invention
The purpose of the present invention is to solve the shortcomings of the prior art, there is provided a kind of time dimension expands the local weighted limit The soft-measuring modeling method of learning machine.
The purpose of the present invention is achieved through the following technical solutions：It is a kind of based on the soft of local weighted extreme learning machine The foundation of measurement model, mainly including following steps：
(1) using system and off-line monitoring method under collecting and distributing control, the data of industrial processes are chronological Training sample set Xtr∈RN×nAnd Ytrain∈RN×mWith test sample collection Xte∈RK×nAnd Ytest∈RN×m.Wherein N is training sample Length, n is training sample dimension, and K is test sample collection length.Test sample collection is that these data are stored in into historical data base. And according to training sample set pre-treatment carried out to training sample set and test sample and normalization makes training sample set its average For 0 and variance be 1.
(2) it is temporally poor that former training sample and test sample are sampled, will XtrAnd XtTime dimension expansion is carried out respectively, It is the test set X of T-shaped Cheng Xin to expand parametertrain∈R(2T+1)N×nWith training set matrix Xtest∈R(2T+1)K×n。
(3) by test sample XtestRow is taken successively as query sample qi(i=1,2 ..., K), and correspondence test specimens reality Real-valued yi(i=1,2 ..., K).Carry out local weighted extreme learning machine modeling respectively by each query sample afterwards.
(4) query sample average is carried out to query sample and training set.This part go average value will model and Reduction goes to inquire about the data deviation of average on final result again after to modeling result.
(5) weight parameter w is determined apart from d according to query sample and historical sample (training set sample), to historical sample Weighting, obtains new training sample XwThe sample space to local weightedization.
(6) time dimension is expanded and the sample space of local weightedization carries out extreme learning machine modeling, obtain hard measurement As a result.
(7) whole test sample has been obtained after repeating the vectorial modeling of all of query sample and hard measurement result Hard measurement result.
(8) number of the local weighted extreme learning machine method to industrial process is expanded based on time dimension using derived above According to being modeled, the hard measurement of implementation process.
The beneficial effects of the invention are as follows：The present invention uses instant learning thought, and time dimension expansion is carried out on time dimension With the dynamic property of extraction model, spatially using local weighted method, with the dynamic sex chromosome mosaicism for data in industrial process, It is nonlinear for data in industrial process as a kind of ripe high speed high precision nonlinear method with reference to extreme learning machine Problem.In cohesive process, time dimension expands the big dynamic matrix of a large amount of correlations for producing and solves extreme learning machine most Big problem be that Generalization Capability is not enough, it is necessary to data sample is big and ignore the problem of sample correlations, local weighted method is being carried Extreme learning machine can be to a certain extent solved the problems, such as during taking system dynamic does not have a noise reduction process, and can be with yet The unstability of reduction and extreme learning machine in real process.Obtained at high speed high accuracy for industrial process nonlinear and The hard measurement device of dynamic.
Brief description of the drawings
Fig. 1 is extreme learning machine debutanizing tower hard measurement Error Graph；
Fig. 2 is local weighted extreme learning machine debutanizing tower hard measurement Error Graph；
Fig. 3 is that time dimension expands extreme learning machine debutanizing tower hard measurement Error Graph；
Fig. 4 is that time dimension expands local weighted extreme learning machine debutanizing tower hard measurement Error Graph.
Specific embodiment
The present invention be directed to industrial process nonlinear, dynamic sex chromosome mosaicism is opened up using instant learning thought by time dimension Exhibition method and local weighted method spatially solve the problems, such as dynamic, non-linear as one kind with reference to extreme learning machine algorithm Algorithm solves process nonlinear problem.Wherein, especially time dimension expanding method can greatly increase the side of process data amount Method, can cross largely solve extreme learning machine industrial process sample it is universal it is not big enough in the case of evolvement problem compared with Poor problem, and the big problem of sample correlations will not also cause very big burden to the accuracy and speed of extreme learning machine.It is local Method of weighting can reduce to a certain extent noise should around with reduce extreme learning machine unstability.This method can either be realized The extraction and modeling of time dimension expanding method and local method of weighting to industrial process dynamic, can also embody limit study Machine is used as a kind of nonlinear regression algo high precision, fireballing advantage.
The present invention is described in detail with instantiation below in conjunction with the accompanying drawings
The technical solution adopted by the present invention key step is as follows：
The first step is using system and off-line monitoring method under collecting and distributing control, the data composition modeling of industrial processes Training sample set and test sample collection.Wherein, training sample set includes Xtr∈RN×nAnd Ytrain∈RN×m, wherein, N is instruction Practice sample length, n is training sample dimension.Test sample collection is Xte∈RK×nAnd Ytest∈RN×mThese data are stored in history Database, wherein, K is test sample length, and n is training sample dimension.And to training sample set and test sample according to instruction Practice that sample set carries out pre-treatment and normalization makes that training sample set its average is 0 and variance is 1.
Second step is by XtrAnd XteCarry out time dimension to expand to form new test set and training set matrix, setting time dimension It is T that degree expands parameter, then by XtrAnd XteCarry out time point different sampling and form new matrix Xtrain=[Xtr(t=0) Xtr(t =1) Xtr(t=-1) ... Xtr(t=T) Xtr], (t=-T) Xtest=[Xte(t=0) Xte(t=1) Xte(t=-1) ... Xte(t= T)Xte(t=-T)], wherein, Xtrain∈R(2T+1)N×n, Xtest∈R(2T+1)K×n
3rd step is by test sample XtestRow is taken successively as query sample, as (q1,q2,…qi…qK), wherein qiFor Query sample, n dimension row vectors.Test sample actual value y is corresponded to respectivelyi(i=1,2 ..., K).From the first row (i=1) to finally A line carries out local weighted extreme learning machine modeling respectively.
4th step is q when query sampleiWhen (i=1,2 ..., K), query sample is carried out to query sample and training set Average.This part goes the value of average to model and reduction goes to inquire about the number of average on final result again after obtaining modeling result According to deviation.
Training sample is concentrated X by the 5th steptrainCapable amount form is decomposed into, willIn each xjAs history sample This (j=1,2 ..., N), you can to each historical sample and query sample qiSimilarity or the distance of point measuredThen weight can be formulated to similar sampleWhereinFor the part of setting adds Weight parameter.Obtain all weight W=[w1, w2,…wN].The input value for being input into neutral net is replaced byAfterwards Operating limit learning machine is modeled.
6th step extreme learning machine is a kind of single hidden layer probabilistic neural metanetwork, at random C hidden neuron node g of generation All parameter (a of (ai, bi)1,b1)(a2,b2)…(aC,bC) it is generated after, according to the output formula of monolayer neural networksThis formula can also be write as：
H β=T, wherein β are that neuron exports weight matrix, and T is neutral net output result,
Error | T-Y | is caused to pursue2Minimum, can obtain weight matrix computing formula β=H+T.The neutral net for obtaining The node parameter a of weight beta random generation plus beforei, biIt is exactly the neural network parameter of whole extreme learning machine.
7th step repeats above step, and a predicted value for the local weighted limit study of query sample is obtained every time, repeats Carry out all of query sample vector modeling and hard measurement result after obtained the hard measurement result of whole test sample.
8th step expands local weighted extreme learning machine method to industrial process using derived above based on time dimension Data be modeled, the hard measurement of implementation process.
Effectiveness of the invention is illustrated below in conjunction with a specific debutanizing tower example.For the process, adopt altogether 2394 data are collected, wherein 1197 data are used for modeling, and off-line analysis and mark have been carried out to its corresponding butane content value Note.1197 data samples of collection are used for verifying the validity of soft-sensing model in addition.In this process, 7 are have chosen altogether Individual process variable carries out soft sensor modeling to the butane content of the process, and this 7 process variables are respectively tower top temperature, tower top pressure Power, return flow, next stage flow, sensitive plate temperature, column bottom temperature and tower bottom pressure.Next the detailed process is combined to this The implementation steps of invention are set forth in：
1. data pre-processing, is pre-processed and normalizing to the process variable and butane content in 2394 modeling samples Change so that the average of each process variable and key variables is zero, and variance is 1, obtains new modeling data matrix.
2. the soft sensor modeling of local weighted extreme learning machine is expanded based on time dimension
Will choose procedure variable composition data matrix as soft-sensing model input, butane content data square Battle array, according to the method detailed provided in implementation steps, sets up related as the output of soft-sensing model to each test sample point Time dimension expand the soft-sensing model of local weighted extreme learning machine.
3. the data of pair whole sample sets are modeled and predict
In order to test the validity of new method, we carry out locally fine point to 2394 test samples and hard measurement is pre- Survey, obtain comparing after hard measurement result and studying.
4. the soft sensor modeling prediction of butane content is compared
Online soft sensor is carried out to 2394 checking samples, corresponding On-line Estimation value is obtained.Limits of application learning machine side For 2394 online soft sensors of checking sample, error is 0.130686 to method.Only carry out local weighted without the time of carrying out The extreme learning machine soft-sensing model that dimension is expanded carries out hard measurement to identical sample, as a result compares with actual value, error It is 0.0880488.Time dimension is only carried out to expand without carrying out local weighted extreme learning machine soft-sensing model to identical Sample carries out hard measurement, as a result compares with actual value, and error is 0.0549991.Pole is expanded using time dimension of the present invention The comparison of the result and actual value of the hard measurement that limit learning machine hard measurement is carried out to the sample, error is 0.0480588.Can see Go out time dimension expansion extreme learning machine and local weighted extreme learning machine all reduces the error of prediction, improve soft-side face mould The precision of type, and time dimension expands the error that local weighted extreme learning machine further reduces prediction, improves soft survey Measure the precision of model.
Above-described embodiment illustrates the present invention again, rather than limiting the invention, in spirit of the invention and In scope of the claims, to initial people of the invention and modifications and changes, protection scope of the present invention is both fallen within.

Claims (4)

1. a kind of soft-measuring modeling method that local weighted extreme learning machine is expanded based on time dimension, it is characterised in that including Following steps:
(1) using system and off-line monitoring method under collecting and distributing control, the chronological training of the data of industrial processes Sample set Xtr∈RN×nAnd Ytrain∈RN×mWith test sample collection Xte∈RK×nAnd Ytest∈RN×m.Wherein N is training sample length, N is training sample dimension, and K is test sample collection length.Test sample collection is that these data are stored in into historical data base.And it is right Training sample set and test sample carry out pre-treatment and normalization according to training sample set makes its average of training sample set be 0 and side Difference is 1.
(2) it is temporally poor that former training sample and test sample are sampled, will XtrAnd XtTime dimension expansion is carried out respectively, is expanded Parameter is the test set X of T-shaped Cheng Xintrain∈R(2T+1)N×nWith training set matrix Xtest∈R(2T+1)K×n。
(3) by test sample XtestRow is taken successively as query sample qi(i=1,2 ..., K), and correspondence test sample actual value yi(i=1,2 ..., K).Carry out local weighted extreme learning machine modeling respectively by each query sample afterwards.
(4) query sample average is carried out to query sample and training set.This part goes the value of average to model and built Reduction goes to inquire about the data deviation of average on final result again after mould result.
(5) determine weight parameter w apart from d according to query sample and historical sample (training set sample), historical sample weighted, Obtain new training sample XwThe sample space to local weightedization.
(6) time dimension is expanded and the sample space of local weightedization carries out extreme learning machine modeling, obtain hard measurement result.
(7) repeat all of query sample vector modeling and hard measurement result after obtained the soft survey of whole test sample Amount result.
(8) data of industrial process are entered based on time dimension expansion local weighted extreme learning machine method using derived above Row modeling, the hard measurement of implementation process.
2. a kind of industrial process hard measurement of the local weighted extreme learning machine model of time dimension expansion is built according to right 1 Mould, it is characterised in that in the step 2, by XtrAnd XtTime dimension is carried out to expand to form new test set and training set matrix, It is T that setting time dimension expands parameter, then by XtrAnd XteCarry out time point different sampling and form new matrix Xtrain=[Xtr (t=0) Xtr(t=1) Xtr(t=-1) ... Xtr(t=T) Xtr], (t=-T) Xtest=[Xte(t=0) Xte(t=1) Xte(t=- 1)…Xte(t=T) Xte(t=-T)], wherein, Xtrain∈R(2T+1)N×n, Xtest∈R(2T+1)K×n。
3. a kind of industrial process hard measurement of the local weighted extreme learning machine model of time dimension expansion is built according to right 1 Mould, it is characterised in that in the step 2, the test set of model data is being divided into the query sample q of test set pointsi(i= 1,2 ..., K), individually modeling is predicted each query sample during the foundation of ensuing model.
4. a kind of industrial process hard measurement of the local weighted extreme learning machine model of time dimension expansion is built according to right 1 Mould, it is characterised in that in the step 2, foundation query sample qi (i=1,2 ..., K) and historical sample dj (j=1,2 ..., N similarity)The distance between i.e. determine weightWhereinIt is the office of setting Portion's weighting parameters.By local weighted dataCarried out as the local weighted model training sample input of extreme learning machine Local weighted modeling.
CN201611236844.1A 2016-12-28 2016-12-28 Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method Pending CN106709570A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611236844.1A CN106709570A (en) 2016-12-28 2016-12-28 Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611236844.1A CN106709570A (en) 2016-12-28 2016-12-28 Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method

Publications (1)

Publication Number Publication Date
CN106709570A true CN106709570A (en) 2017-05-24

Family

ID=58896908

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611236844.1A Pending CN106709570A (en) 2016-12-28 2016-12-28 Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method

Country Status (1)

Country Link
CN (1) CN106709570A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107463994A (en) * 2017-07-07 2017-12-12 浙江大学 Semi-supervised flexible measurement method based on coorinated training extreme learning machine model
CN111768000A (en) * 2020-06-23 2020-10-13 中南大学 Industrial process data modeling method for online adaptive fine-tuning deep learning

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003330977A (en) * 2002-05-17 2003-11-21 Nec Corp Electromagnetic field analyzing device and electromagnetic field analyzing method
CN105205224A (en) * 2015-08-28 2015-12-30 江南大学 Modeling method for soft measurement of time difference gaussian process regression based on fuzzy curve analysis

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003330977A (en) * 2002-05-17 2003-11-21 Nec Corp Electromagnetic field analyzing device and electromagnetic field analyzing method
CN105205224A (en) * 2015-08-28 2015-12-30 江南大学 Modeling method for soft measurement of time difference gaussian process regression based on fuzzy curve analysis

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
GONZAGA J C B ETAL.: "ANN-based soft-sensor for real-time process monitoring and control of an industrial polymerization process", 《COMPUTERS & CHEMICAL ENGINEERING》 *
ZAMPROGNAE ETAL.: "Optimal selection of soft sensor inputs for batch distillation columns using principal component analysis", 《JOURNAL OF PROCESS CONTROL》 *
刘国海 等: "基于连续隐Maricov棋型的发酵过程关键状态变巧软测量", 《江苏大学学报(白然科学版)》 *
袁小锋 等: "基于时间差分和局部加权偏最小二乘算法的过程自适应软测量建模", 《化工学报》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107463994A (en) * 2017-07-07 2017-12-12 浙江大学 Semi-supervised flexible measurement method based on coorinated training extreme learning machine model
CN111768000A (en) * 2020-06-23 2020-10-13 中南大学 Industrial process data modeling method for online adaptive fine-tuning deep learning

Similar Documents

Publication Publication Date Title
CN111222549B (en) Unmanned aerial vehicle fault prediction method based on deep neural network
CN108897286B (en) Fault detection method based on distributed nonlinear dynamic relation model
CN104537415B (en) A kind of non-linear process industry failure predication and recognition methodss based on compressed sensing and DROS ELM
CN112101480B (en) Multivariate clustering and fused time sequence combined prediction method
CN109992921B (en) On-line soft measurement method and system for thermal efficiency of boiler of coal-fired power plant
CN113379164B (en) Load prediction method and system based on deep self-attention network
CN109635245A (en) A kind of robust width learning system
Tian et al. Short-term wind power prediction based on empirical mode decomposition and improved extreme learning machine
CN106682312A (en) Industrial process soft-measurement modeling method of local weighing extreme learning machine model
CN111178616A (en) Wind speed prediction method based on negative correlation learning and regularization extreme learning machine integration
CN106548270A (en) A kind of photovoltaic plant power anomalous data identification method and device
CN110070202A (en) A method of economic output is predicted by electricity consumption data
CN105846448A (en) Method for determining reactive compensation capacity of power distribution network based on random matrix theory
CN111324110A (en) Fermentation process fault monitoring method based on multiple shrinkage automatic encoders
CN106709570A (en) Time dimension expansion and local weighting extreme learning machine-based soft measurement modeling method
Drakoulas et al. FastSVD-ML–ROM: A reduced-order modeling framework based on machine learning for real-time applications
Zhang et al. An improved time delay neural network model for predicting dynamic heat and mass transfer characteristics of a packed liquid desiccant dehumidifier
CN106773697A (en) A kind of time dimension expands the industrial process soft-measuring modeling method of extreme learning machine model
Meng et al. An adaptive variational mode decomposition for wind power prediction using convolutional block attention deep learning network
CN103279030A (en) Bayesian framework-based dynamic soft measurement modeling method and device
Zheng et al. Dynamic rolling force prediction of reversible cold rolling mill based on BP neural network with improved PSO
CN112132328A (en) Photovoltaic output power ultra-short-term local emotion reconstruction neural network prediction method
Li et al. Research on the prediction method of stock price based on rbf neural network optimization algorithm
Wang et al. Continual residual reservoir computing for remaining useful life prediction
Chen et al. Prediction of gas emission based on grey-generalized regression neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20170524

WD01 Invention patent application deemed withdrawn after publication