CN104517162B - A kind of continuous annealing product hardness Online integration learns forecasting procedure - Google Patents

A kind of continuous annealing product hardness Online integration learns forecasting procedure Download PDF

Info

Publication number
CN104517162B
CN104517162B CN201410843307.8A CN201410843307A CN104517162B CN 104517162 B CN104517162 B CN 104517162B CN 201410843307 A CN201410843307 A CN 201410843307A CN 104517162 B CN104517162 B CN 104517162B
Authority
CN
China
Prior art keywords
sample
continuous annealing
product hardness
strip
production
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.)
Active
Application number
CN201410843307.8A
Other languages
Chinese (zh)
Other versions
CN104517162A (en
Inventor
唐立新
王显鹏
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.)
Northeastern University China
Original Assignee
Northeastern University China
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 Northeastern University China filed Critical Northeastern University China
Priority to CN201410843307.8A priority Critical patent/CN104517162B/en
Publication of CN104517162A publication Critical patent/CN104517162A/en
Application granted granted Critical
Publication of CN104517162B publication Critical patent/CN104517162B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • YGENERAL 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
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02PCLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30Computing systems specially adapted for manufacturing

Abstract

A kind of continuous annealing product hardness Online integration learns forecasting procedure, belongs to the automatic control technology field of iron and steel enterprise's continuous annealing production process.Using the history continuous annealing creation data sample of enterprise, the integrated study modeling method using LSSVM as sub- learning machine is used, its offline product hardness forecasting model is set up respectively for different temper strips;In actual production, continuous annealing production process data is read in real time, and real-time prediction is carried out to current belt steel product hardness using the offline product hardness forecasting model set up by integrated study;By the inspection of actual production data, the method of the present invention can significantly improve the precision and robustness of continuous annealing product hardness forecast result, site operation personnel is enabled to grasp the quality of current belt steel product in real time, and according to circumstances adjusted in good time, the deficiency of offline inspection large time delay is compensate for, so as to help continuous annealing production line to improve product quality, improve production operation level, increase economic benefit.

Description

A kind of continuous annealing product hardness Online integration learns forecasting procedure
Technical field
It is more particularly to a kind of continuous the invention belongs to the automatic control technology field of iron and steel enterprise's continuous annealing production process Annealing product hardness Online integration learns forecasting procedure.
Background technology
In the actual production process of iron and steel enterprise's cold rolling mill continuous annealing unit, the hardness of strip is to weigh product quality With the core index of Instructing manufacture.In actual production process, the hardness of strip can not also realize on-line checking, and scene is to pass through The head of strip, portion after interception annealing, then carry out test experiment analysis to determine the hardness of strip, so as to judge product quality Situation.But, the quality information of product is obtained by test experiment analysis typically all has regular hour hysteresis quality, also It is to say only after strip produces a period of time, could obtains its specific quality information, and strip in actual production Speed of production is very fast, and annealing can be just completed in minutes, so that it is very big often to occur strip hardness fluctuations Situation, causes to produce that hardness is not up to standard or even the quality problems such as waste product, has a strong impact on the economic benefit of cold rolling mill.
Paper " forecast of continuous annealing unit strip quality and process monitoring system design based on PLS are with realizing [D] " (Wang Source, Northeastern University, 2009) although being proposed for belt steel product hardness a kind of based on offset minimum binary (Partial Least Squares, PLS) data driven type modeling method, but the method proposed in this paper can not meet actual production The need for process, main cause has:(1) information related to strip hardness that the document is considered is less, only 20 or so, And procedural information related to strip hardness in actual production process then will up to 51;(2) the PLS methods that the document is proposed Monitoring and fault diagnosis main or for continuous annealing production process, and PLS methods belong to a kind of linear regression method, And actual production process is then nonlinear, so as to cause the precision of PLS methods not high;(3) there is input item in sample data It is numerous and intercouple and the problem of output result scope is narrower, cause to exist between sample input item difference it is larger and Output result is same or like so that traditional Data Modeling Method is difficult to obtain higher precision of prediction and robustness.
The content of the invention
In view of the deficienciess of the prior art, the present invention provides a kind of continuous annealing product hardness Online integration study forecast Method.
The technical scheme is that:
A kind of continuous annealing product hardness Online integration learns forecasting procedure, comprises the following steps:
Step 1:For each temper strip, the continuous annealing production in the nearest m periods is read from database Sample, obtains the continuous annealing containing B sample and produces sample set;
Step 2:Continuous annealing for each temper strip produces sample set, carries out data processing;
Each data item in sample set is normalized respectively for step 2.1, to eliminate between different pieces of information The different influence of dimension;
Step 2.2:The gross error detection method based on cluster is used for the sample set after normalized, is rejected The sample that A include human error in sample set;
Step 2.2.1:The minimum Eustachian distance d between each sample and other B-1 sample is calculated respectivelyi(i=1, 2,…,B);
Step 2.2.2:Calculate the average value d of the minimum Eustachian distance of all samplesavg, by davgQ times be used as first Sill;
Step 2.2.3:By the minimum Eustachian distance d in sample set between other B-1 samplesiMore than the first threshold The sample of value is rejected, while remaining sample is clustered, obtains P cluster;
Step 2.2.4:The minimum Eustachian distance D between each cluster and other P-1 cluster is calculated respectivelyu(u=1, 2,…,P);
Step 2.2.5:Calculate the average value D of the minimum Eustachian distance of all clustersavg, and by DavgR times be used as second Threshold;
Step 2.2.6:Minimum Eustachian distance between other P-1 are clustered is more than the second threshold value and clustered interior Sample size be less than t cluster and its comprising sample reject;
Step 3:According to the sample set after step 2 data processing, pivot transition matrix M is asked for and for every kind of temper Strip sets up its product hardness forecasting model using the integrated study modeling method based on LSSVM and forms offline continuous annealing band Product made from steel hardness forecasting model storehouse;
Step 3.1:Using PCA methods, all production process datas in the sample set after step 2 data processing are constituted Matrix carry out dimensionality reduction, keeping characteristics root is more than or equal to 1 pivot, and the pivot of reservation is stored in into pivot transition matrix M;
Step 3.2:According to the sample set after step 2 data processing, using the integrated study modeling side based on LSSVM Method, its product hardness forecasting model is set up for every kind of temper strip, forms offline continuously annealing steel strip product hardness forecast Model library;The method for setting up its product hardness forecasting model for every kind of temper strip, comprises the following steps:
Step 3.2.1:Identical weighted value is assigned for all sample standard deviations in the sample set after step 2 data processing, i.e., The weight w of each samplei=1/ (B-A);
Step 3.2.2:Based on current sample weights, set up what a belt steel product hardness was forecast using LSSVM methods Sub- learning machine model, calculates the root-mean-square error of the model and the training error of each sample;
Step 3.2.3:Training error according to each sample adjusts sample weights, and method is:Make training error larger first Sample weight increase and the less sample of training error weight reduce, new sample weights are obtained, then to new sample This weight is normalized;
Step 3.2.4:Step 3.2.2 is repeated to step 3.2.3, N number of continuously annealing steel strip product hardness is set up pre- The sub- learning machine model of report, and it is corresponding each to calculate the root-mean-square error and each sub- learning machine model of each sub- learning machine model The training error of sample;
Step 3.2.5:According to the root-mean-square error of the step 3.2.4 each sub- learning machine models calculated, each sub- study is calculated The weight of machine model;
Step 3.2.6:It is one by N number of sub- learning machine model integrated according to N number of sub- learning machine model and its respective weight Individual integrated study machine model, the integrated study machine model is this kind of temper belt steel product hardness forecasting model, and it is output as The weighting composite value of the belt steel product hardness predicted value of each sub- learning machine model output;
Step 3.2.7:This kind of temper belt steel product hardness forecasting model is stored in offline continuous annealing band product made from steel hard Spend forecasting model storehouse;
Step 4:According to the strip temper that continuous annealing is produced is treated, from continuously annealing steel strip product hardness forecasting model storehouse Middle selection correspondence temper belt steel product hardness forecasting model, starts continuous annealing production;
Step 5:The maximum sampling period T of point is up-sampled with continuous annealing production linemaxZ times be used as belt steel product hardness In the sampling period of forecasting model, continuous annealing production process data is obtained in real time, it is pre- as this kind of temper belt steel product hardness The process input vector of model is reported, while being produced to each sampled point on continuous annealing production line when every winding product made from steel head is passed through Raw associated production status data is recorded;The production process data is each sampled point on continuous annealing production line in every volume The synthesis of the associated production status data that belt steel product head is produced when passing through and corresponding strip information;The strip packet Include the specification information and hot rolling information (production information of the strip in hot-rolled process) of its own;
Step 6:Using clustering method to process input vector carry out gross error detection, if the process input to Amount includes human error, then deletes the process input vector, abandons the belt steel product hardness forecast in this cycle, and waits next The process input vector in sampling period;Otherwise, then step 7 is entered;
Step 7:Using pivot transition matrix M resulting in step 3, dimensionality reduction is carried out to current process input vector, Obtain the process input vector represented with pivot;
Step 8:Based on the process input vector after dimensionality reduction, using this kind of temper belt steel product hardness forecasting model to band Product made from steel hardness is forecast;
Step 9:After the completion of this coiled strip steel continuous annealing production, interception strip leader portion carries out offline inspection, obtains This coiled strip steel product hardness actual value, and recorded;
Step 10:The corresponding product hardness of the continuous annealing production process data of the every coiled strip steel recorded is actual Value is matched and synthesized, and forms a new continuous annealing production sample;Each continuous annealing production sample includes H data ;The H data item include several related continuous annealing production status data item, several strips information datas and Belt steel product hardness actual value;
Step 11:Continuous annealing production sample is updated, method is:For a certain temper strip, when its nearest O time section After the completion of interior continuous annealing production sample collection, then the continuous annealing production sample in the nearest O time section is added to enterprise In industry database, while the continuous annealing production sample in the database in earliest O time section is deleted;
Step 12:According to the method for step 1 to step 3, sample is produced according to the continuous annealing after renewal, respectively to each Temper belt steel product hardness forecasting model carries out re -training and renewal;
Step 13:When new strip continuous annealing produces beginning, step 4 is repeated to step 13, realizes known adjust The online forecasting of matter degree belt steel product hardness.
Beneficial effects of the present invention:The continuous annealing product hardness Online integration study forecasting procedure of the present invention is used LSSVM is integrated the result of each sub- learning machine as sub- learning machine maker, and using integrated study framework, as final Belt steel product hardness predicted value.By the inspection of actual production data, method of the invention can significantly improve continuous annealing The precision and robustness of product hardness forecast result so that site operation personnel can grasp the matter of current belt steel product in real time Amount, and according to circumstances adjusted in good time, the deficiency of offline inspection large time delay is compensate for, so as to help continuous annealing production line to carry High yield quality, improvement production operation level, increase economic benefit.
Brief description of the drawings
Fig. 1 illustrates for the continuous annealing product hardness Online integration study forecasting procedure principle of one embodiment of the present invention Figure;
Fig. 2 learns forecasting procedure flow chart for the continuous annealing product hardness Online integration of one embodiment of the present invention;
Fig. 3 learns the pre- press of forecasting procedure for the continuous annealing product hardness Online integration of one embodiment of the present invention Face figure;
Fig. 4 is the continuous annealing product hardness integrated study predicted value and corresponding test sample of one embodiment of the present invention Comparative result figure;
Fig. 5 monitors surface chart on-line for the continuous annealing product hardness of one embodiment of the present invention;
Embodiment
The invention will be further described with reference to the accompanying drawings and detailed description.
Continuous annealing production is one of important procedure of iron and steel enterprise's cold rolling mill, as shown in figure 1, continuous annealing production line is pressed It can be divided into following 9 stages according to function:Heating furnace (HF), soaking pit (SF), leer (SCF), 1# black furnaces (1C), 1# are out-of-date Imitate stove (1OA), 2# overaging stove (2OA), 2# cooling furnaces (2C), water quenching oven (WQ), planisher.In process of production, cold rolling strap Steel is walked with certain speed in each stove of production line, it is completed the heat such as heating, cooling according to the annealing process route of setting Handling process process, so as to eliminate internal stress of the strip caused by cold rolling, then after smooth, obtains high-quality strip.
For in actual production process, the hardness of strip can not also realize the present situation of on-line checking, and the present invention provides one Plant continuous annealing product hardness Online integration study forecasting procedure.The operation principle of this method is as shown in Figure 1:First, enterprise is utilized The history continuous annealing creation data sample of industry, use with LSSVM (Least Square Support Vector Machine, Least square method supporting vector machine) as the integrated study modeling method of sub- learning machine, set up respectively for different temper strips Its offline product hardness forecasting model, and form offline continuously annealing steel strip product hardness forecasting model storehouse;Secondly, in actual life During production, mould is forecast according to strip temper that continuous annealing is produced is treated from the continuously annealing steel strip product hardness set up offline Corresponding product hardness forecasting model is selected in type storehouse;Then, from the sampling of continuous annealing production process with being obtained in control system Each sampled point is produced when every winding product made from steel head is passed through on continuous annealing production line associated production status data is simultaneously carried out Data prediction and carry out synthesizing and obtaining continuous annealing Real-time Production Process data with strip information;Then, using clustering Method carries out gross error detection to continuous annealing Real-time Production Process data;It is real for the continuous annealing not comprising human error When production process data, after PCA (Principle Component Analysis, pivot analysis) method dimension-reduction treatment, Selected product hardness forecasting model is passed to, so as to export the belt steel product hardness under current working to site operation personnel Predicted value, and then it is easy to site operation personnel to grasp product quality in time and carried out according to current belt steel product hardness predicted value Corresponding operating condition adjustment.In addition, can be stored per the production process data at winding steel taping head, when belt steel product hardness After offline inspection result comes out, after being synthesized with it, new continuous annealing creation data sample is formed, enterprise database is arrived in storage In.After the completion of the new samples collection of nearest one month, the re -training process of belt steel product hardness forecasting model can be started, from And the newest operating mode of continuous annealing unit can be tracked by ensureing forecasting model.
Present embodiment is based on the continuous annealing production technology shown in Fig. 1, the belt steel product that production temper is T5, and root Learn forecasting procedure principle according to continuous annealing product hardness Online integration, continuous annealing product hardness proposed by the invention is existed Line integrated study method and application is in present embodiment, as shown in Fig. 2 comprising the following steps:
S1:For the strip that temper is T5, the continuous annealing that nearest 3 months are gathered from enterprise database produces sample This, obtains 350 samples altogether;Wherein each sample is made up of following 52 data item, including 14 strip Information Numbers According to item, respectively 9 strip steel specification information data items are that inlet thickness, throat width, carbon content, sulfur content, phosphorus content, manganese contain Amount, nitrogen content, silicone content and total aluminium content, and 5 hot rolling information data items are tapping temperature, Average curl temperature, are averaged Final rolling temperature, average finishing temperature and continuous acid-washing-rolling elongation percentage (CDCM elongation percentage);37 continuous annealing associated production shapes State data item, respectively central sections speed, the area's furnace temperature of HF stoves 1, the area's furnace temperature of HF stoves 2, the area's furnace temperature of HF stoves 3, the area's furnace temperature of HF stoves 4, HF The area's furnace temperature of stove 5, HF outlet of stills belt steel temperature, the area's furnace temperature of SF stoves 1, the area's furnace temperature of SF stoves 2, SF outlet of stills belt steel temperature, the area of SCF stoves 1 Furnace temperature, the area's furnace temperature of SCF stoves 2, SCF outlet of stills belt steel temperature, the area's furnace temperature of 1C stoves 1, the area's furnace temperature of 1C stoves 2, the area's furnace temperature of 1C stoves 3,1C stoves Cooling gas temperature, 1C outlet of stills belt steel temperature, the area's furnace temperature of 1OA stoves 1, the area's furnace temperature of 1OA stoves 2,1OA outlet of stills belt steel temperature, 2OA Stove 1-1 areas furnace temperature, 2OA stove 1-2 areas furnace temperature, 2OA stove 2-1 areas furnace temperature, 2OA stove 2-2 areas furnace temperature, 2OA outlet of stills belt steel temperature, 2C Stove furnace temperature, 2C cooling gas temperature, 2C outlet of stills belt steel temperature, the area's water temperature of WQ stoves 1, the area's water temperature of WQ stoves 2, planisher elongation percentage, Planisher strip steel at entry tension force, strip tension in the middle of planisher, planisher outlet strip tension, planisher 1# frames roll-force and 2# frame roll-forces;1 belt steel product hardness actual value data item.The continuous annealing production sample set such as institute of table 1 gathered Show, each of which row represents a sample, it is each to be classified as a data item, and successively according to from strip steel specification information data, Hot rolling information data, continuous annealing associated production status data to last belt steel product hardness actual value data item it is suitable Sequence is arranged.With SiRepresent sample, Si=(si1,si2,…,sik,…,si52), wherein i=1,2 ..., 350, sikRepresent i-th K-th of data item of individual sample.
In table 1. nearest 3 months temper for T5 belt steel product continuous annealing production sample set
S2:Data processing is carried out to the continuous annealing production sample set shown in table 1;
S2.1:Each column data in table 1 is normalized respectively, to eliminate dimension between different pieces of information Different influences;
S2.2:The gross error detection method based on cluster is used for the sample set after normalized, bag is rejected Sample containing human error, comprises the following steps:
S2.2.1:Each sample S is calculated respectivelyiWith the minimum Eustachian distance between other 349 samplesWherein i=1,2 ..., 350;
S2.2.2:Calculate the average value of the minimum Eustachian distance of all samplesBy davg3 times, That is 3davgIt is used as the first threshold value (3davgFor the threshold value obtained by experiment);
S2.2.3:It is d by the minimum Eustachian distance in sample set with other 349 samplesi>3davgSample reject, By remaining sample, (minimum Eustachian distance with other 349 samples is d simultaneouslyi≤3davgSample) carried out according to nearby principle Cluster, obtains 36 clusters;
S2.2.4:Each cluster C is calculated respectivelyuMinimum Euclidean between (u=1,2 ..., 36) and other 35 clusters away from FromWhereinRepresent cluster CuCenter (cu1,…, cu52) in k-th of data item value, i.e., the arithmetic mean of instantaneous value of k-th of data item in all samples that the cluster is included, wherein | Cu| represent cluster CuThe quantity of middle sample;
S2.2.5:Calculate the average value D of the minimum Eustachian distance of 36 clustersavg, and by Davg3 times, i.e. 3DavgAs Second threshold;
S2.2.6:Minimum Eustachian distance between other 35 are clustered is more than the second threshold value and clusters interior sample Quantity be less than 10 cluster and its comprising sample reject;Wherein 3DavgIt is to be obtained by experiment with the sample size 10 in cluster Threshold value;
Continuous annealing shown in table 1 produces sample set after S2.1 and S2.2 data processing, has 5 and included The sample of error difference is removed, as shown in table 2.
5 comprising the human error sample that table 2. is removed
S3:Using PCA methods to preceding 14 strips information data of remaining 345 samples in sample set and 37 The associated production status data item matrix that totally 51 data item are constituted carries out dimensionality reduction, and keeping characteristics root is more than or equal to 1 11 masters Member, is stored in pivot transition matrix M;Meanwhile, always according to remaining 345 samples in sample set, using the collection based on LSSVM Into learning model building method, its product hardness forecasting model is set up for this kind of temper strip, and it is quenched by this kind set up Degree belt steel product hardness forecasting model is stored in offline continuously annealing steel strip product hardness forecasting model storehouse;It is described to be directed to this kind The method that temper strip sets up its product hardness forecasting model, comprises the following steps:
S3.1:Identical weighted value, i.e., the weight w of each sample are assigned for 345 sample standard deviationsi=1/345, and is set Habit machine indexes l=1;
S3.2:Based on current sample weights, trained using LSSVM methods and obtain what a belt steel product hardness was forecast Sub- learning machine model;Method is:Radial basis kernel function is used in LSSVM, i.e.,Wherein xiRepresent Preceding 51 input datas of i-th of training sample, σ2For kernel functional parameter, specific step is as follows:
S3.2.1:In span, 20 initial LSSVM parameter combinations are randomly generated.Remember that each parameter combination is Pj=(γj2 j), wherein γjWith σ2 jParameter combination P is represented respectivelyjIn to the penalty coefficient and kernel functional parameter of error.γjWith σ2 jSpan be respectively [5,50] and [0.1,2.0];
S3.2.2:For each parameter combination Pj=(γj2 j), solve a system of linear equations as follows:
Wherein n=345 is total sample number amount, yi(i=1 ..., 345) represents the output of i-th of sample, i.e. strip hardness. So as to obtain the model parameter α of the LSSVM corresponding to the parameter combinationi(i=1 ..., 345) and b, further according to each sample data Preceding 51 input datas use set up LSSVM models to predict its hardness, i.e. Hardness Prediction valueAnd then calculate parameter combination PjEvaluation index, that is, the root-mean-square error predicted the outcome
S3.2.3:Using the minimum LSSVM models of current RMSE value as best LSSVM models, LSSVM is designated asbest
S3.2.4:Based on 20 current parameter combinations, 20 new parameter combinations are obtained in accordance with the following methods:First, Randomly choose two parameter combination Pf=(γf2 f) and Pg=(γg2 g);Secondly, a new parameter combination P is producedj= (γj2 j), γjIn scope [γmin–0.5I1max+0.5I1] in randomly generate, wherein γmin=min { γfg, γmax =max { γfg, I1max–γmin;Similarly, σ2 jIn scope [σ2 min–0.5I22 max+0.5I2] in randomly generate, Wherein σ2 min=min { σ2 f2 g, σ2 max=max { σ2 f2 g, I22 max–σ2 min
S3.2.5:For 20 parameter combinations newly produced, its evaluation index is calculated using the S3.2.2 methods proposed. Less preceding 20 parameter combinations of evaluation index are taken from current 40 parameter combinations, if having obtained more preferable LSSVM models, Then update LSSVMbest, delete remaining 20 second-rate parameter combinations.
S3.2.6:S3.2.4 and S3.2.5 are repeated, (repeats to be able to for 500 times in given fortune for 500 times altogether The preferable LSSVM models of quality are obtained in row time range), by resulting LSSVMbestStore, learned as l-th of son Habit machine, and calculate the corresponding each sample x of the modeliTraining error
S3.3:According to each sample xiTraining error ζiSample weights are adjusted, method is:First, ε=0 is set, for institute There is sample, if ζi>0.01, then ε=ε+wi;Secondly, β is setl=ε × ε;Again, if sample xiTraining error ζi> 0.01, then the weight for adjusting the sample is wi=wi×βl, otherwise its weight be adjusted to wi=1;So that training error is larger Sample weight increase, and the less sample weights of training error diminish;Finally, to all weight wiPlace is normalized Reason.
S3.4:The method that above-mentioned S3.2 and S3.3 are proposed is reused, son of 5 belt steel product hardness forecast is set up Habit machine model, the best parameter group and its corresponding β of each sub- learning machine modellValue, average training error are as shown in table 3.
The relevant parameter of the pre- man who brings news of appointment's learning machine model of the belt steel product hardness of table 3.
Numbering l γ σ2 βl Average relative error (%) RMSE
1 23.408 0.806 0.0125 0.50059 4.12081
2 20.9353 0.845 0.778 0.539676 4.34022
3 49.6097 0.731 0.570 0.533521 4.30655
4 25.0794 0.702 0.563 0.604213 4.95296
5 21.8188 0.712 0.181 0.480041 3.86035
S3.5:Based on the 5 sub- learning machine models set up, the strip of integrated study machine model is determined according to equation below Product hardness predicts the outcome:WhereinIt is to be directed to sample x Preceding 51 input items, the predicted value calculated by l-th of belt steel product hardness forecasting model.5 sub- learning machine Models Sets The training result of T5 temper belt steel product hardness forecasting models after is as shown in Figure 3.
In addition to 345 training samples of used nearest 3 months, from enterprise database acquisition time earlier 1 41 belt steel product hardness actual values of individual month as T5 temper belt steel product hardness forecasting models test sample, with examine The validity of the continuous annealing product hardness forecasting model based on integrated study proposed.It is hard by T5 temper belt steel products The comparative result of belt steel product Hardness Prediction result and this 41 belt steel product hardness actual values obtained by degree forecasting model is such as Shown in table 4 and Fig. 4, it can be seen that method proposed by the invention is relative not have in the estimated performance of continuous annealing product hardness The consensus forecast deviation for having the test sample learnt is 0.585%, and can follow the change of belt steel product actual hardness Trend, with preferable generalization ability and robustness.
The test sample of table 4. and the contrast that predicts the outcome of the continuous annealing product hardness forecasting model based on integrated study
S3.6:The integrated study model that the temper belt steel product hardness is forecast is stored in offline continuous annealing band product made from steel Hardness forecasting model storehouse;
S4:When actual continuous annealing produces the strip that temper is T5, mould is forecast from continuously annealing steel strip product hardness Above-mentioned T5 tempers belt steel product hardness forecasting model is selected in type storehouse, starts continuous annealing production;
S5:The maximum sampling period using continuous annealing production line up-sampling point took its 2 times as T5 temper bands as 1 second It is the sampling period of product made from steel hardness forecasting model, each with being read in control system from the sampling of continuous annealing production process every 2 seconds The associated production status data that sampled point is produced when every winding product made from steel head is passed through, and it is synthesized into company with strip information Continuous annealing production process data, i.e. first 51 of continuous annealing production sample, referred to as process input vector, meanwhile, to continuously moving back The associated production status data that each sampled point is produced when every winding product made from steel head is passed through on fiery production line is recorded;
S6:Gross error detection is carried out to process input vector using clustering method, method is:Calculating process first Minimum Euclidean in input vector and current database between first 51 of each continuous annealing production sample of T5 tempers strip Distance, then minimum Eustachian distance average value is calculated, if the average value is more than the first threshold value calculated in S2.2.2 3davg, then it is assumed that it is normal sample, then enters in next step S7;Otherwise it is assumed that the sample includes human error, delete The process input vector, abandons the hardness forecast in this cycle, and waits the process input vector in next sampling period;
S7:Using pivot transition matrix M resulting in step 3, dimensionality reduction is carried out to current process input vector, obtained The process input vector represented with pivot;
S8:Based on the process input vector after dimensionality reduction, correspondence is forecast using T5 temper belt steel product hardness forecasting model Strip hardness number, and shown by operating platform display device, as shown in figure 5, the interface upper left corner can be shown in real time Strip hardness predicted value corresponding to current production status, facilitates site operation personnel to be carried out by the predicted value to production process Monitoring, so as to ensure product quality in acceptability limit.
S9:After the completion of current strip production, interception strip leader portion carries out offline inspection, obtains this winding product made from steel hard It is 64 to spend actual value, and is recorded;
S10:By the continuous annealing production process data of this coiled strip steel recorded and this coiled strip steel product hardness actual value 64 are matched and are synthesized, and form a new continuous annealing production sample, and recorded;
S11:T5 temper strips continuous annealing production sample is updated, method is:For T5 temper strips, if when it After the completion of the continuous annealing production sample collection of nearest one month, then the continuous annealing production sample of this month is added first Into enterprise database, then all continuous annealings production sample standard deviation of that month by the time in the database earliest is deleted Remove,
S12:According to S2 to S4 method, sample is produced according to the T5 temper strips continuous annealing after renewal, T5 is adjusted Matter degree belt steel product hardness forecasting model carries out re -training with updating, so that T5 temper belt steel products hardness forecasts mould Type can follow current continuous annealing actual production operating mode to be adjusted correspondingly.
S13:When new T5 temper strips continuous annealing produces beginning, S4 to S13 is repeated, is realized quenched to T5 Spend the online forecasting of belt steel product hardness.
Present embodiment learns the company that forecast system realizes present embodiment by continuous annealing product hardness Online integration Continuous annealing product hardness Online integration study forecasting procedure, the system is included with lower module:Database maintenance module, continuous annealing Production process modeling module, data preprocessing module and product hardness online forecasting module.Database maintenance module is mainly used to The deletion to enterprise's historical data and the importing of new samples data etc. is completed to operate;Data preprocessing module is mainly used to completion and is directed to Gross error detection and PCA dimensionality reductions of production process sample etc. are operated.Continuous annealing production process modeling module is to complete base In LSSVM integrated study modeling function, and show training and the test result of set up belt steel product hardness forecasting model. Product hardness online forecasting module is mainly used to realize continuously annealing steel strip product hardness online forecasting function.Present embodiment Continuous annealing product hardness Online integration study forecast system, be and shown in Fig. 1 continuous annealing production process sampling with control System processed matches what is used, and the sampling of continuous annealing production process and control system place can be attached it to when using at the scene Computer in, the computer as the inventive method hardware platform.The company that the inventive method will be installed from the computer Associated production state of the continuous annealing production process sampling with reading each sampled point of continuous annealing production process in control system in real time Data, after being pre-processed, form continuous annealing production process data, reuse integrated study model to current belt steel product Hardness carries out real-time prediction, and result is shown on this computer screen, facilitates site operation personnel to grasp band in real time The quality information of steel, and take in quality fluctuation the regulating measures of correlation.

Claims (5)

1. a kind of continuous annealing product hardness Online integration learns forecasting procedure, it is characterised in that:Comprise the following steps:
Step 1:For each temper strip, the continuous annealing production sample in the nearest m periods is read from database, Obtain the continuous annealing containing B sample and produce sample set;
Step 2:Continuous annealing for each temper strip produces sample set, carries out data processing;
Step 3:According to the sample set after step 2 data processing, pivot transition matrix M is asked for and for every kind of temper strip Its product hardness forecasting model is set up using the integrated study modeling method based on LSSVM and offline continuously annealing steel strip production is formed Product hardness forecasting model storehouse;
Step 4:According to the strip temper that continuous annealing is produced is treated, selected from continuously annealing steel strip product hardness forecasting model storehouse Correspondence temper belt steel product hardness forecasting model is selected, starts continuous annealing production;
Step 5:The maximum sampling period T of point is up-sampled with continuous annealing production linemaxZ times be used as belt steel product hardness forecast In the sampling period of model, continuous annealing production process data is obtained in real time, mould is forecast as this kind of temper belt steel product hardness The process input vector of type, while each sampled point on continuous annealing production line is produced when every winding product made from steel head is passed through Associated production status data is recorded;The production process data is each sampled point on continuous annealing production line in every coiled strip steel The synthesis of the associated production status data that product head is produced when passing through and corresponding strip information;The strip information includes it The specification information and hot rolling information of itself;
Step 6:Gross error detection is carried out to process input vector using clustering method, if the process input vector bag Containing human error, then the process input vector is deleted, abandon the belt steel product hardness forecast in this cycle, and wait next sampling The process input vector in cycle;Otherwise, then step 7 is entered;
Step 7:Using pivot transition matrix M resulting in step 3, dimensionality reduction is carried out to current process input vector, obtained The process input vector represented with pivot;
Step 8:Based on the process input vector after dimensionality reduction, strip is produced using this kind of temper belt steel product hardness forecasting model Product hardness is forecast;
Step 9:After the completion of this coiled strip steel continuous annealing production, interception strip leader portion carries out offline inspection, obtains this volume Belt steel product hardness actual value, and recorded;
Step 10:The corresponding product hardness actual value of the continuous annealing production process data of the every coiled strip steel recorded is entered Row matching and synthesis, form a new continuous annealing production sample;Each continuous annealing production sample includes H data item; The H data item includes several related continuous annealing production status data item, several strip information datas and band Product made from steel hardness actual value;
Step 11:Update continuous annealing production sample;
Step 12:According to the method for step 1 to step 3, sample is produced according to the continuous annealing after renewal, respectively to each quenched Degree belt steel product hardness forecasting model is updated;
Step 13:When new strip continuous annealing produces beginning, step 4 is repeated to step 13, realizes known temper The online forecasting of belt steel product hardness.
2. continuous annealing product hardness Online integration according to claim 1 learns forecasting procedure, it is characterised in that:It is described Step 2 in for each temper strip continuous annealing production sample set carry out data processing, comprise the following steps:
Step 2.1:Each data item in sample set is normalized respectively, to eliminate dimension between different pieces of information Different influences;
Step 2.2:The gross error detection method based on cluster, Rejection of samples are used for the sample set after normalized The sample that A include human error in set;
Step 2.2.1:The minimum Eustachian distance d between each sample and other B-1 sample is calculated respectivelyi(i=1,2 ..., B);
Step 2.2.2:Calculate the average value d of the minimum Eustachian distance of all samplesavg, by davgQ times be used as the first threshold;
Step 2.2.3:By the minimum Eustachian distance d in sample set between other B-1 samplesiMore than the first threshold value Sample is rejected, while remaining sample is clustered, obtains P cluster;
Step 2.2.4:The minimum Eustachian distance D between each cluster and other P-1 cluster is calculated respectivelyu(u=1,2 ..., P);
Step 2.2.5:Calculate the average value D of the minimum Eustachian distance of all clustersavg, and by DavgR times be used as the second threshold;
Step 2.2.6:Minimum Eustachian distance between other P-1 are clustered is more than the second threshold value and clusters interior sample Quantity be less than t cluster and its comprising sample reject.
3. continuous annealing product hardness Online integration according to claim 1 learns forecasting procedure, it is characterised in that:It is described Step 3 in pivot transition matrix M is asked for according to the sample set after step 2 data processing method be:Using PCA methods, Dimensionality reduction is carried out to the matrix that all production process datas are constituted in the sample set after step 2 data processing, keeping characteristics root is big Pivot transition matrix M is stored in the pivot equal to 1, and by the pivot of reservation.
4. continuous annealing product hardness Online integration according to claim 2 learns forecasting procedure, it is characterised in that:It is described Step 3 according to the sample set after step 2 data processing for every kind of temper strip use integrated based on LSSVM The method that modeling method sets up its product hardness forecasting model is practised, is comprised the following steps:
Step 3.2.1:Assign identical weighted value for all sample standard deviations in the sample set after step 2 data processing, i.e., it is each The weight w of samplei=1/ (B-A);
Step 3.2.2:Based on current sample weights, son that a belt steel product hardness is forecast is set up using LSSVM methods Habit machine model, calculates the root-mean-square error of the model and the training error of each sample;
Step 3.2.3:Training error according to each sample adjusts sample weights, and method is:The sample for making training error larger first This weight increase and the weight of the less sample of training error reduce, and obtain new sample weights, then new sample is weighed It is normalized again;
Step 3.2.4:Step 3.2.2 to step 3.2.3 is repeated, N number of continuously annealing steel strip product hardness forecast is set up Sub- learning machine model, and calculate the root-mean-square error and the corresponding each sample of each sub- learning machine model of each sub- learning machine model Training error;
Step 3.2.5:According to the root-mean-square error of the step 3.2.4 each sub- learning machine models calculated, each sub- learning machine mould is calculated The weight of type;
Step 3.2.6:It is a collection by N number of sub- learning machine model integrated according to N number of sub- learning machine model and its respective weight Into learning machine model, the integrated study machine model is this kind of temper belt steel product hardness forecasting model, and it is output as each son The weighting composite value of the belt steel product hardness predicted value of learning machine model output;
Step 3.2.7:This kind of temper belt steel product hardness forecasting model is stored in offline continuously annealing steel strip product hardness pre- Report model library.
5. continuous annealing product hardness Online integration according to claim 1 learns forecasting procedure, it is characterised in that:It is described Step 11 in update continuous annealing production sample method be:For a certain temper strip, when in its nearest O time section After the completion of continuous annealing production sample collection, then the continuous annealing production sample in the nearest O time section is added to enterprise's number According in storehouse, while the continuous annealing production sample in the database in earliest O time section is deleted.
CN201410843307.8A 2014-12-30 2014-12-30 A kind of continuous annealing product hardness Online integration learns forecasting procedure Active CN104517162B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410843307.8A CN104517162B (en) 2014-12-30 2014-12-30 A kind of continuous annealing product hardness Online integration learns forecasting procedure

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410843307.8A CN104517162B (en) 2014-12-30 2014-12-30 A kind of continuous annealing product hardness Online integration learns forecasting procedure

Publications (2)

Publication Number Publication Date
CN104517162A CN104517162A (en) 2015-04-15
CN104517162B true CN104517162B (en) 2017-09-05

Family

ID=52792422

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410843307.8A Active CN104517162B (en) 2014-12-30 2014-12-30 A kind of continuous annealing product hardness Online integration learns forecasting procedure

Country Status (1)

Country Link
CN (1) CN104517162B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107541597B (en) * 2016-06-29 2019-03-05 宝山钢铁股份有限公司 Strip running deviation monitoring and diagnostic method and the system of continuous annealing unit soaking pit
CN106345823B (en) * 2016-08-31 2018-02-27 北京首钢自动化信息技术有限公司 The method of online real-time estimate mechanical performance based on coils of hot-rolled steel production procedure
CN106845825B (en) * 2017-01-18 2020-04-28 西安交通大学 Strip steel cold rolling quality problem tracing and control method based on improved PCA
CN106886799B (en) * 2017-03-17 2019-08-02 东北大学 A kind of continuous annealing band steel quality online test method based on hybrid integrated study
CN108034804B (en) * 2017-12-08 2019-04-23 中国地质大学(武汉) A kind of method and system of continuous annealing unit furnace area's energy consumption modeling
CN108281192B (en) * 2017-12-29 2022-03-22 一诺仪器(中国)有限公司 Human body component prediction method and system based on ensemble learning algorithm
CN108229072A (en) * 2018-02-10 2018-06-29 东北大学 Continuous annealing soaking pit strip tension On-line Measuring Method based on data parsing
CN110321658B (en) * 2019-07-10 2023-09-01 江苏金恒信息科技股份有限公司 Method and device for predicting plate performance
CN110989522B (en) * 2019-12-06 2022-09-09 东北大学 Multi-steel-coil-oriented optimal setting method for technological parameters in continuous annealing production process
TWI718928B (en) * 2020-04-10 2021-02-11 中國鋼鐵股份有限公司 Method and system for evaluating a width of a steel strip
CN111563686A (en) * 2020-05-13 2020-08-21 中冶南方工程技术有限公司 Cold-rolled silicon steel quality judgment method based on full-process data
CN112598026B (en) * 2020-12-07 2023-09-22 东北大学 Continuous annealing production process fault diagnosis method based on unbalanced data deep learning
CN114943364A (en) * 2022-03-31 2022-08-26 苏州新凌电炉有限公司 Intelligent visual management method and device for mesh belt furnace

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008099457A1 (en) * 2007-02-09 2008-08-21 Toshiba Mitsubishi-Electric Industrial Systems Corporation Process line control unit and method of controlling the line
CN101899563A (en) * 2009-06-01 2010-12-01 上海宝钢工业检测公司 PCA (Principle Component Analysis) model based furnace temperature and tension monitoring and fault tracing method of continuous annealing unit
CN102033523A (en) * 2009-09-25 2011-04-27 上海宝钢工业检测公司 Strip steel quality forecasting, furnace condition early-warning and fault diagnosis method based on partial least square
CN102564644A (en) * 2011-12-20 2012-07-11 东北大学 Temperature online measuring method for plate blank in production process of heating furnace

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008099457A1 (en) * 2007-02-09 2008-08-21 Toshiba Mitsubishi-Electric Industrial Systems Corporation Process line control unit and method of controlling the line
CN101899563A (en) * 2009-06-01 2010-12-01 上海宝钢工业检测公司 PCA (Principle Component Analysis) model based furnace temperature and tension monitoring and fault tracing method of continuous annealing unit
CN102033523A (en) * 2009-09-25 2011-04-27 上海宝钢工业检测公司 Strip steel quality forecasting, furnace condition early-warning and fault diagnosis method based on partial least square
CN102564644A (en) * 2011-12-20 2012-07-11 东北大学 Temperature online measuring method for plate blank in production process of heating furnace

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于平均轨迹的连续退火过程带钢硬度预测方法;彭俊 等;《控制工程》;20130930;第20卷(第5期);864-868 *
基于正交信号校正的连续退火机组带钢质量预测方法;王超 等;《仪表技术》;20121231;1-4,43 *

Also Published As

Publication number Publication date
CN104517162A (en) 2015-04-15

Similar Documents

Publication Publication Date Title
CN104517162B (en) A kind of continuous annealing product hardness Online integration learns forecasting procedure
CN107377634B (en) A kind of hot-strip outlet Crown Prediction of Media method
CN106660090B (en) Rolling simulation device
CN101391268B (en) Reverse optimization method of steel plate rolling and cooling controlling-process temperature institution
CN106345823B (en) The method of online real-time estimate mechanical performance based on coils of hot-rolled steel production procedure
CN105740467B (en) A kind of method for digging of C-Mn steel industry big data
CN108311544B (en) A kind of roll-force Parameter Self-learning method and device
CN106413931A (en) Method for optimally producing metal steel and iron alloys in hot-rolled and thick plate factories using a microstructure simulator, monitor, and/or model
CN107179749A (en) Hot dip zinc product whole process method of quality control
CN109389238B (en) Ridge regression-based short-term load prediction method and device
CN108655186A (en) Roll-force presetting method based on artificial neural network and mathematical model
CN104657602A (en) Method and system for forecasting thickness of band steel in hot continuous rolling production process
CN106886799B (en) A kind of continuous annealing band steel quality online test method based on hybrid integrated study
CN103447314B (en) Rough rolling temperature-measuring feedback control method
CN104714519B (en) A kind of continuous annealing unit production process technology parameter setting and method for on-line optimization
CN111832215A (en) Method for on-line predicting steel plate structure performance
CN108921232B (en) Hot-rolled strip steel cooling historical data clustering and similarity measuring method
WO2020148918A1 (en) Metal material design support method and design support device
CN1641356B (en) Rigid line product microscopic structure and mechanical property forecasting system
CN105095273A (en) Fuzzy BP neural network based glass tempering process parameter setting method
CN109840313B (en) Low-salt water intake judging method for estuary area
CN114564879A (en) Continuous annealing strip steel hardness online prediction method based on multi-scale data fusion
Vorel et al. Material-Technological Modelling of Controlled Cooling of Closed die Forgings from Finish Forging Temperature
RU2655398C2 (en) Method of rolled products production
CN102260781B (en) On-line early warming method of elongation rate and flattening process state of strip steel of continuous annealing unit

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant