CN105108097B - Hybrid model for continuous casting breakout prediction - Google Patents
Hybrid model for continuous casting breakout prediction Download PDFInfo
- Publication number
- CN105108097B CN105108097B CN201510447779.6A CN201510447779A CN105108097B CN 105108097 B CN105108097 B CN 105108097B CN 201510447779 A CN201510447779 A CN 201510447779A CN 105108097 B CN105108097 B CN 105108097B
- Authority
- CN
- China
- Prior art keywords
- model
- thermocouple
- temperature
- continuous casting
- neural network
- 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
Links
Classifications
-
- 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
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B22—CASTING; POWDER METALLURGY
- B22D—CASTING OF METALS; CASTING OF OTHER SUBSTANCES BY THE SAME PROCESSES OR DEVICES
- B22D11/00—Continuous casting of metals, i.e. casting in indefinite lengths
- B22D11/16—Controlling or regulating processes or operations
- B22D11/18—Controlling or regulating processes or operations for pouring
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Mechanical Engineering (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Continuous Casting (AREA)
Abstract
The invention discloses a hybrid model for continuous casting breakout prediction and belongs to the technical field of monitoring of metallurgical continuous casting. The hybrid model comprises a single couple sequence model based on a GA (Genetic Algorithm)-BP (Back Propagation) neural network and a group couple space model based on logical judgment. The prediction method comprises the following steps: (1) monitoring the temperature of a crystallizer and saving data; (2) inputting the data to the single couple sequence model, judging whether the time-varying temperature of each thermocouple conforms to the temperature variation waveform during sticking, and saving a judgment result in an array Y (i, j, t); (3) when the array Y (i, j, t) is within the range of [theta min, theta max], marking that the thermocouple is abnormal, and calculating the number of abnormal thermocouples in the ith row and the (i-1)th row to be respectively m and n; and (4) comparing m+n with the threshold values of the numbers of sticking-alarming and sticking-warning thermocouples for sticking judgment. The hybrid model and the hybrid-model-based continuous casting bleed-out prediction method achieve the goal of improving the sticking breakout recognition precision.
Description
Technical field
The invention belongs to monitoring technology field in metallurgical continuous casting, more particularly, it relates to one kind is for continuous casting bleed-out forecast
Mixed model.
Background technology
Continuous casting bleed-out is the pernicious quality thing that the cc billet surface quality development of defects such as bonding or crackle are produced to a certain extent
Therefore, conticaster can be caused to stop production, the production schedule of casting process continuity and whole steel-making, and damage equipment are affected, affect casting
The operating rate and yield of machine, reduces the recovery rate of metal, causes huge economic loss.In actual production process, bonding
Property bleed-out occurrence frequency highest, accounts for the 70%～80% of all kinds of bleed-out events.Especially with the efficient slab continuous casting technology of modernization
Development, not only will cast crack sensitivity steel grade, and steel grade wide ranges of casting, casting difficulty is big, and the raising of pulling rate causes
More complicated Billet, friction and the problems such as lubricate, cool down initial solidification shell and solidify, the stability that covering slag is flowed into and
Uniformity is remarkably decreased, and strand bonding phenomenon increases in crystallizer, and then caused sticker type breakout problem is projected very much.Bonding
Property bleed-out be main bleed-out form, study and solve sticker type breakout to ensure continuous casting direct motion and improve slab quality have
Significance.
From the seventies in last century, the detection method of a variety of sticker type breakout signs is developed both at home and abroad, it is most effective
Method be thermocouple thermometry, general principle is by burying certain amount thermocouple, detection crystallization on copper plate of crystallizer
Device copper coin different parts temperature variations, using the localized heat transfer shape inside copper plate temperature situation of change monitor in real time crystallizer
Condition and identification strand rupture location and its mobile message.At present, the breakout prediction method based on thermocouple temperature measurement mainly has two classes,
One class is, by logic judgment model, to carry out extracting appropriate after qualitative and quantitative analysis according to bleed-out Analysis on Mechanism and bleed-out data
Logical condition carries out breakout prediction, and its principle is according to each electric thermo-couple temperature amplitude of variation, rate temperature change, upper and lower heat extraction
The parameters such as the galvanic couple temperature difference, temperature change time delay and the threshold values multilevel iudge of setting, make the warning of bleed-out degree.Logic is sentenced
Disconnected model depends on specific technique and device parameter etc., and the selection of model parameter needs substantial amounts of manpower and time test, from
Often there is higher rate of false alarm in adaptability and poor robustness, Jing, and frequently reporting by mistake can equally affect the quality of strand and casting machine
High-efficiency reform, and reduce reporting by mistake to increase failing to report.Another kind of is by intellectual technology (such as neutral net, supporting vector
Machine isotype recognizer) sticker type breakout is forecast, be characterized in have very strong adaptivity, self-learning capability,
Fault-tolerance and robustness, can preferably process complexity nonlinear problem, can further improve the accuracy of breakout prediction, into
For current study hotspot.Model of mind belongs to black-box model, and its deficiency is undue dependence data, such as neural network model instruction
White silk must rely on enough effective samples, if the incomplete or inaccurate generalization ability that can all affect network of sample data.Even
Casting machine is gone into operation the initial stage, due to lacking data effective enough, it is necessary to bleed-out is forecast and avoided by logic judgment model.
Method with regard to being forecast to sticker type breakout using neutral net, existing related art scheme in prior art
Disclosure, such as patent publication No.：The A of CN 101850410, publication date：On October 6th, 2010, invention and created name is：One kind is based on
The continuous casting breakout prediction method of neutral net, this application case discloses a kind of continuous casting breakout prediction method based on neutral net,
The method includes, step 1：The temperature data of online acquisition continuous casting scene thermocouple simultaneously stores the temperature data；Step 2：To institute
State temperature data to be pre-processed；Step 3：By through the pretreated temperature data gathered from any one thermocouple
It is input to single idol sequential network breakout prediction model, and the output valve to single idol sequential network breakout prediction model and maximum differentiation
Threshold value is compared, if the output valve of single idol sequential network breakout prediction model is more than maximum discrimination threshold, forecast leakage
Steel can occur；Meanwhile, the connection weight and threshold value of the list idol sequential network breakout prediction model are initialized using genetic algorithm.
The method can improve the recognition effect and forecast precision that bleed-out process is cohered to continuous casting, so as to reducing rate of false alarm and failing to report
Rate.But, this application case is disadvantageous in that：From single idol temporal model to a group structure for even spatial model, god is fully relied on
Jing network technologies, although the intellectual technology such as neutral net has obvious advantage in dynamic waveform pattern-recognition, but utilizes it
Foundation group idol spatial model is simultaneously improper；Output valve in the method step 3 by single idol sequential network breakout prediction model is big
In maximum discrimination threshold, just forecast bleed-out can occur, and unreasonable, because the single electric thermo-couple temperature Jing of actual production process often goes out
Existing larger temperature fluctuation, it is easy to be close to sticking temperature pattern, and cause false alarm；The group idol space being related in the method
The input of network breakout prediction model be respectively simultaneously from meet a thermocouple determining whether and it is corresponding under
The temperature data of row's three thermocouples in left, center, right is input to the output valve obtained after single idol temporal model, it can thus be appreciated that the group idol
Spatial model does not account for electric thermo-couple temperature Spatial Variation when bonding V-type tearing port is propagated, and only selects described
Four thermocouples judge the two-dimentional dissemination of bonding, when there is multiple thermocouple failures and temperature in actual production process
When fluctuation is larger, the group idol spatial model is easy to generation and fails to report and false alarm, and robustness will during its practical application
It is deteriorated.
In sum, how to overcome it is existing sticker type breakout is carried out by logic judgment model or neural network model it is pre-
The weak point of report, is the technical barrier of urgent need to resolve in prior art.
The content of the invention
1. the invention technical problem to be solved
By logic judgment model or neural network model sticker type breakout is forecast instant invention overcomes existing
A kind of weak point, there is provided mixed model for continuous casting bleed-out forecast, realizes and improves sticker type breakout accuracy of identification
Target.
2. technical scheme
To reach above-mentioned purpose, the technical scheme that the present invention is provided is：
The mixed model for continuous casting bleed-out forecast of the present invention, mainly including following two parts：Single idol temporal model and
The even spatial model of group；
(1), single idol temporal model；
The structure of single idol temporal model includes：The determination of mode input variable, the determination of model output variable, data it is pre-
Process, the foundation of GA-BP neutral nets；
(2) even spatial model, is organized；
1), all electric thermo-couple temperatures on crystallizer are changed over after pattern is identified using single idol temporal model,
Output result is saved in three-dimensional array Y (i, j, t), wherein, Y (i, j, t) represents the i-th row j row thermocouple in single idol of t
Sequence Model Identification result；
2), when Y (i, j, t) is in threshold range [θmin,θmax] it is interior when, it is believed that thermocouple TC (i, j) temperature change meets
Sticking temperature pattern, marks the thermocouple exception；
3) and then the Y (i, j, t) of all thermocouples of the i-th row is checked, is counted in threshold range [θmin,θmax] in exception
Thermocouple number is m, while checking the Y (i-1, j, t) of all thermocouples of the i-th -1 row, is counted in threshold range [θmin,θmax] in
Abnormal thermocouple number be n, wherein i be more than 1；
If 4), m and n are all higher than being equal to 2, check in past 10 seconds menisci row (the first row) exception thermoelectricity even number
Whether mesh is more than or equal to 2, if it is satisfied, then alerting thermoelectricity with strike-alarm and bonding respectively using abnormal thermocouple sum m+n
Even number mesh threshold values compares, and carries out the judgement of strike-alarm and bonding warning.
Used as further improvement of the present invention, mode input variable is defined below；
Single thermocouple 30 temperature sampling points continuous in time are selected as mode input variable；
Wherein, the collection period of temperature sampling point is 1 second.
Used as further improvement of the present invention, model output variable is defined below；
Model output variable is single thermocouple strike-alarm signal, and model output variable is by above-mentioned 30 temperature samplings point
Corresponding temperature change waveform pattern is determined with the degree of closeness of sticking temperature pattern；
Wherein, model output variable is the number between -1～2, when the corresponding temperature change fluted mould of 30 temperature sampling points
When formula is identical with sticking temperature pattern, model output variable is labeled as 1；When the corresponding temperature change of 30 temperature sampling points
During curve held stationary, model output variable is labeled as 0.
Used as further improvement of the present invention, the preparation and pretreatment of data is as follows；
Sample is extracted from historical data, above-mentioned effective sample is normalized between [- 1,1] using formula (1)；
In formula (1), x' be normalization after sample data, x be normalization before sample data, xmaxBefore being normalization
The maximum of sample data, xminIt is the minimum of a value of sample data before normalization；
Sample data after normalizing by more than is divided into two parts, and a portion is made as training sample, another part
For test sample.
Used as further improvement of the present invention, the foundation of GA-BP neural network models is as follows；
BP neural network input layer number is 30, representative model input variable；Output layer nodes are 1, representative model
Output variable；One discrimination threshold scope is determined according to the result of network training, when model output variable is not less than presetting
Threshold range when be considered as detecting sticking temperature pattern；Training network adopts 3 layers of BP networks, hidden layer excitation function to use
S type tangent transfer functions, output layer uses linear transfer function, training process to adopt LM optimized algorithms；Selected the number of hidden nodes
For 12, the BP neural network model that structure is 30 × 12 × 1 is obtained；
Wherein：BP neural network learning process includes information forward-propagating and error back propagation, and according to giving sample is trained
This input and output vector constantly learn and adjust the connection weight between neuron and threshold value, make network constantly approach sample defeated
Mapping relations between entering and exporting；The maximum frequency of training of BP neural network is set to 2000, and learning rate is 0.05, performance error
For 0.0001；
Using genetic algorithm optimization BP neural network, GA-BP neural network models, genetic algorithm optimization BP nerve net are set up
The basic step of network includes：1) all weights of BP neural network and threshold value are carried out real coding by, initialization of population, are produced individual
Body, arranges population scale and evolutionary generation；2), each individual fitness function is calculated according to formula (2), is with minimum of a value
It is optimum；
In formula (2), n is network output layer nodes；yiFor the reality output of i-th node；y'iFor i-th node
Desired output；3), carry out the selection of genetic algorithm, intersect and mutation operation, produce a new generation individuality population；4) new one, is evaluated
For population, judge whether evolutionary generation reaches requirement or whether network error meets condition, if meet obtaining current population most
Excellent fitness value correspondence is individual；Wherein, Population in Genetic Algorithms scale is set to 50, and crossover probability is 0.7, and mutation probability is 0.06,
Evolutionary generation is 200.
As further improvement of the present invention, the threshold range [θmin,θmax] it is [0.6,1.3].
Used as further improvement of the present invention, strike-alarm thermocouple number threshold values is 6, bonding warning thermocouple number
Threshold values is 3；When abnormal thermocouple sum m+n is more than or equal to 3, bonding warning is sent；When abnormal thermocouple sum m+n more than etc.
When 6, strike-alarm is sent.
The continuous casting breakout prediction method based on mixed model of the present invention, comprises the steps：
Step (1), multiple rows of high density thermocouple is arranged in copper plate of crystallizer, monitor mould temperature situation of change, and
Gather and store live all electric thermo-couple temperature real time datas, be saved in three-dimensional array T (i, j, t)；Wherein, T (i, j, t) is represented
Temperature value of the i-th row j row thermocouple in t；
Step (2), by the single idol temporal model of all thermo-electric couple temperature datas input, in single idol temporal model, each heat
The time series data of galvanic couple temperature, after the conversion and data processing of shift register, is input into GA-BP neural network models
Calculate, temperature change waveform when judging whether change of each electric thermo-couple temperature in time series meets bonding will determine that
As a result in being saved in three-dimensional array Y (i, j, t)；
Step (3) if, Y (i, j, t) setting threshold range [θmin,θmax] it is interior when, then it is assumed that current thermocouple TC
(i, j) temperature change meets sticking temperature pattern, marks the thermocouple exception, then then carry out a group judgement for even spatial model,
Calculate current thermocouple to be expert at and lastrow exception thermocouple number；
Step (4), the abnormal thermocouple sum for exporting group even spatial model alert heat with strike-alarm and bonding respectively
Galvanic couple number threshold values compares, and carries out the judgement of strike-alarm and bonding warning.
3. beneficial effect
The technical scheme provided using the present invention, compared with prior art, with following remarkable result：
(1), the present invention considers advantage of the artificial intelligence technology on wave pattern cognition, crystallizer when boning with strand
The spatial and temporal variation of electric thermo-couple temperature is foundation in copper coin, using GA-BP neural networks list idol temporal model, for knowing
Whether change of the not single electric thermo-couple temperature in time series meets temperature change waveform during bonding, belongs to dynamic waveform mould
Formula recognizes problem.Wherein, genetic algorithm optimization BP neural network, by ability of searching optimum BP neural network best initial weights are determined
And threshold value, improve accuracy of identification of single idol temporal model to sticking temperature waveform pattern.And in view of bonding V-type tearing port
Two-dimentional dissemination, a group even spatial model is established on the basis of single idol temporal model by effective logic rules, is judged
Whether neighboring thermocouple has the propagation of bonding, improves the accuracy of identification of sticker type breakout, can especially reduce actual production
During multiple thermocouple failures or larger electric thermo-couple temperature failing to report and false alarm when fluctuating.
(2), by the mixed model for continuous casting bleed-out forecast proposed by the present invention, GA-BP neutral nets are made full use of
Advantage in wave pattern cognition, and effective logic rules judgement is coupled, not only realize the even space-time of single even summation group and sentence
Break, and overcome simple logic judgment model parameter and determine that difficult or inaccurate and simple model of mind lacks technique and instructs
Deficiency, reached preferable breakout prediction performance, can be promptly and accurately quote whole bondings, it is to avoid sticker type breakout accident,
And false alarm frequency is minimized into level.
Description of the drawings
Fig. 1 is copper plate of crystallizer thermocouple arrangement schematic diagram, mm in embodiment 1；
Fig. 2 is the flow chart of the continuous casting breakout prediction method in embodiment 1 based on mixed model；
Fig. 3 is sticking temperature pattern in embodiment 1；
Fig. 4 is BP neural network topological structure and learning process schematic diagram in embodiment 1；
Fig. 5 is BP neural network and GA-BP neutral net test result figure in embodiment 1.
Specific embodiment
The present invention proposes a kind of mixed model and the continuous casting breakout prediction method based on mixed model, it is intended to solve slab
This technical barrier of sticker type breakout in casting process.Present invention is primarily based on strand bond when copper plate of crystallizer in thermocouple temperature
The spatial and temporal variation of degree, initially with genetic algorithm optimization BP neural network (setting up GA-BP neural network models), sets up
Single idol temporal model, the time dependent dynamic waveform of single electric thermo-couple temperature, is then built using logic rules during identification bonding
Vertical group idol spatial model, differentiates whether vertical and horizontal neighboring thermocouple has sticking temperature waveform, identification bonding two dimension propagation row
Thus to constitute GA-BP neutral nets and logic judgment mixed model.Determine BP neural network using genetic algorithm in the present invention
Best weight value and threshold value, improve the accuracy of identification of single idol temporal model.Continuous casting based on mixed model proposed by the present invention leaks
Steel forecasting procedure, can timely and accurately quote strand and all bond, it is to avoid sticker type breakout accident, and reduce bonding wrong report
Probability.
To further appreciate that present disclosure, in conjunction with the accompanying drawings and embodiments the present invention is described in detail.
Embodiment 1
Conticaster in the present embodiment adopts High Efficiency Slab Caster, two machines two to flow, slab section be 230 × (900～
2150)mm2, pulling rate is 0.80～2.03m/min, and using Combined vertical crystallizer, length of mould is 900mm, crystallizer width
Degree and thickness are according to slab cross section regulation.As shown in figure 1, burying multiple rows of high density thermocouple in copper plate of crystallizer, affixed side is (outward
Arc) and the wide face of active side (inner arc) 6 rows 12 are respectively installed and arrange totally 72 thermocouples, left side and right side leptoprosopy respectively install 6 rows 2 and arrange totally 12
168 thermocouples are installed altogether in individual thermocouple, copper plate of crystallizer.This conticaster is often poured due to steel grade wide ranges of casting, and Jing
Casting crack sensitivity steel grade, casting difficulty is big, while pulling rate is higher, therefore sticker type breakout problem is projected very much, single current bleed-out
Rate reaches 0.0392%, and sticker type breakout accounts for 72% or so of whole bleed-outs, is main bleed-out form, so reducing caking property
Bleed-out is the key for reducing breakout ratio, and it is the reliability for reducing sticker type breakout that sticker type breakout is timely and accurately forecast
Ensure, above-mentioned target can be reached using the continuous casting breakout prediction method based on mixed model of the present embodiment.
As shown in Fig. 2 the mixed model of the present embodiment mainly includes following two parts：List based on GA-BP neutral nets
The group idol spatial model that even temporal model and logic-based judge.
(1), single idol temporal model；
The structure of single idol temporal model includes：The determination of mode input variable, the determination of model output variable, the standard of data
Standby and pretreatment, the foundation of GA-BP neutral nets.The selection of mode input variable and model output variable is most important, directly
Affect predicting the outcome for the list idol temporal model.It can be seen from history bonding or sticker type breakout actual measurement sample, bond process list
At 30 seconds or so, the temperature acquisition cycle was 1 second to the temperature anomaly change in time of individual thermocouple, therefore when selecting single thermocouple
Between upper continuous 30 temperature samplings point (select single thermocouple in 30 continuous temperature acquisitions as mode input variable
Data on cycle are used as mode input variable), can completely characterize the typical ripple of the single electric thermo-couple temperature change of bonding process
Shape pattern, the typical waveform pattern of the single electric thermo-couple temperature change of bonding process is referred to as sticking temperature pattern, concrete such as Fig. 3 institutes
Show.In the present embodiment, model output variable is single thermocouple strike-alarm signal, and model output variable is by 30 temperature samplings
The corresponding temperature change waveform pattern of point is determined with the degree of closeness of sticking temperature pattern；Model output variable is between -1～2
Number, when the corresponding temperature change waveform pattern of 30 temperature sampling points is identical with sticking temperature pattern, model output
Variable label is 1；When the corresponding temperature variation curve held stationary of 30 temperature sampling points, i.e., when temperature change is completely normal,
Model output variable is labeled as 0.
The preparation and pretreatment of data is as follows, according to mode input variable and model output variable, from bonding and positive reason
Sample is extracted in the historical datas such as condition, the temperature data of imperfect and apparent error is rejected, 611 groups of effective samples are obtained altogether (to be had
Effect sample is 611 groups of single thermocouples, 30 temperature sampling points continuous in time), there are 141 groups to be viscous in 611 groups of effective samples
Junction temperature pattern sample；Above-mentioned 611 groups of effective samples are normalized between [- 1,1] using formula (1), is normalized by more than
Sample data afterwards is divided into two parts, wherein selecting 502 groups of sample datas for training pattern, this 502 groups of sample datas are referred to as
Training sample, has 131 groups for sticking temperature pattern sample in training sample；Remaining 109 groups of sample datas are used for test model,
This 109 groups of sample datas are referred to as test sample, have 30 groups in test sample for sticking temperature pattern sample.
In formula, x' be normalization after sample data, x be normalization before sample data, xmaxIt is sample number before normalization
According to maximum, xminIt is the minimum of a value of sample data before normalization.
The foundation of GA-BP neutral nets is as follows, analyzes from more than, and BP neural network input layer number is 30, generation
List thermocouple 30 temperature sampling points (i.e. mode input variable) continuous in time；BP neural network output layer nodes
For 1, output result is single thermocouple strike-alarm signal (i.e. model output variable), represents 30 samplings on current thermocouple
The corresponding temperature change waveform pattern of point and sticking temperature pattern degree of closeness.
In actual test, a discrimination threshold scope is determined according to the result of network training, when model output variable not
It is considered as detecting sticking temperature pattern during more than discrimination threshold scope set in advance.Training network adopts 3 layers of BP networks, hidden
Excitation function containing layer uses S type tangent transfer functions, output layer to use linear transfer function, training process to adopt Levenberg-
Marquardt (LM) optimized algorithm.Through multiple trial, it is 12 to select the number of hidden nodes, obtains structure for 30 × 12 × 1
BP neural network model, as shown in Figure 4.BP neural network learning process includes information forward-propagating and error back propagation, root
According to training sample input and output vector constantly learn and adjust the connection weight between neuron and threshold value, make network not
The disconnected mapping relations approached between sample input and output.The maximum frequency of training of BP neural network is set to 2000, and learning rate is
0.05, performance error is 0.0001.
To improve the generalization ability of BP neural network, using genetic algorithm (GA) Optimized BP Neural Network, basic step bag
Include：The all weights of neutral net and threshold value are carried out real coding by 1., initialization of population, produce it is individual, arrange population scale and
Evolutionary generation；2. each individual fitness function, is calculated, such as shown in formula (2), with minimum of a value as optimum；3., lost
The selection of propagation algorithm, intersect and mutation operation, produce a new generation individuality population；4. population of new generation, is evaluated, evolutionary generation is judged
Whether reach requirement or whether network error meets condition, if meet obtaining current population adaptive optimal control angle value correspondence individuality,
The BP neural network weights and threshold value of optimum are corresponded to.Wherein, Population in Genetic Algorithms scale is set to 50, and crossover probability is 0.7,
Mutation probability is 0.06, and evolutionary generation is 200.Using genetic algorithm optimization BP neural network, by the training of effective sample and
After test, the neural network structure of optimum is obtained, for identification of single idol temporal model to temperature pattern.
In formula, n is network output layer nodes；yiFor the reality output of i-th node of BP；y'iFor the phase of i-th node
Hope output.
Using above-mentioned 502 groups of training samples, BP neural network model and GA-BP neural network models are set up respectively, and divide
It is other that above-mentioned 109 groups of test samples (wherein 30 groups is sticking temperature pattern sample) are predicted, predict the outcome and be shown in Table 1 and Fig. 5.
As shown in Figure 5, the predicted value and desired value degree of closeness of GA-BPNN models (i.e. GA-BP neural network models) is compared with BPNN models
(i.e. BP neural network model) is high, illustrates that GA Optimized BP Neural Networks improve the generalization ability of network, while from GA-BPNN moulds
Recognition result of the type to 30 groups of sticking temperature pattern samples, the threshold range that also can determine that single idol sticking temperature pattern-recognition is
[0.6,1.3] is more suitable.As shown in Table 1, GA-BPNN models are high compared with the accuracy of identification of BPNN models, also illustrate that by losing
Propagation algorithm (GA) Optimized BP Neural Network improves recognition effect of single idol temporal model to sticking temperature pattern.
The accuracy of identification of the single idol temporal model of table 1
(2) even spatial model, is organized；
As shown in Fig. 2 change over pattern to all electric thermo-couple temperatures on crystallizer using single idol temporal model carrying out
After identification, output result is saved in three-dimensional array Y (i, j, t), and it represents single idol sequential mould of the i-th row j row thermocouple in t
Type recognition result (alarm signal).Here threshold range [θ set in advancemin,θmax] be as the above analysis [0.6,
1.3], when Y (i, j, t) is in the range of this, it is believed that thermocouple TC (i, j) temperature change meets sticking temperature pattern, mark
Remember the thermocouple exception.A group judgement for even spatial model is so then carried out, current thermocouple is calculated and is expert at and lastrow exception
Thermocouple number；The Y (i, j, t) of all thermocouples of the i-th row is specially checked, is counted in threshold range [θmin,θmax] in it is different
Often thermocouple number is m, while checking the Y (i-1, j, t) of the i-th row lastrow (i.e. the i-th -1 row) all thermocouples, is counted in valve
Value scope [θmin,θmax] in abnormal thermocouple number be n, wherein i be more than 1.If m and n are all higher than being equal to 2, inspection is needed
Whether look in past 10 seconds menisci row (the first row) exception thermocouple number more than or equal to 2, if it is satisfied, then using exception
Thermocouple sum (m+n) compares respectively with strike-alarm and bonding warning thermocouple number threshold values, carries out strike-alarm and bonding
The judgement of warning.From the foregoing, it will be observed that the group idol spatial model that the present embodiment is proposed is by patrolling on the basis of single idol temporal model
Collect whether rule judgment neighboring thermocouple has the propagation of bonding, substantially increase the reliability that mixed model forecasts to continuous casting bleed-out
Property.
The continuous casting breakout prediction method based on mixed model of the present embodiment, comprises the steps：1., in copper plate of crystallizer
It is interior to arrange multiple rows of high density thermocouple, mould temperature situation of change is monitored, and gather and store live all electric thermo-couple temperatures
Real time data, is saved in three-dimensional array T (i, j, t)；2., by the single idol temporal model of all thermo-electric couple temperature datas input, in list
In even temporal model, the time series data of each electric thermo-couple temperature is defeated after the conversion and data processing of shift register
Enter the calculating of GA-BP neural network models, when judging whether change of each electric thermo-couple temperature in time series meets bonding
Temperature change waveform, will determine that result is saved in three-dimensional array Y (i, j, t)；If 3., threshold values models of the Y (i, j, t) in setting
Enclose [θmin,θmax] it is interior when, then it is assumed that current thermocouple TC (i, j) temperature change meets sticking temperature pattern, marks the thermocouple
It is abnormal, then then to carry out a group judgement for even spatial model, calculate current thermocouple and be expert at and lastrow exception thermocouple number；
4., a group abnormal thermocouple sum for even spatial model output is alerted into thermocouple number threshold values ratio with strike-alarm and bonding respectively
Compared with, carry out strike-alarm and bonding warning judgement.Wherein, T (i, j, t) represents temperature of the i-th row j row thermocouple in t
Value, Y (i, j, t) represents single idol temporal model recognition result of the i-th row j row thermocouple in t, reflects current thermocouple temperature
The degree of closeness of degree changing pattern and sticking temperature pattern.
In the present embodiment, the even spatial model of group is tested to 97, scene of continuous casting heat and (actually occurs 14 times to glue
Knot), it is determined that optimal strike-alarm thermocouple number threshold values is 6, and bonding warning thermocouple number threshold values is 3.And will test
As a result compare with existing Danieli systems breakout prediction method, as shown in table 2.Wherein, rate=true warning time is quoted
Number/(failing to report number of times+true alarm times), forecast accuracy=true alarm times/(fail to report number of times+true alarm times+wrong report time
Number).As shown in Table 2, organizing even spatial model can all quote bonding, and nothing is failed to report, and it is 1 to report number of times by mistake, and forecast accuracy reaches
To 93.33%, better than existing Danieli systems breakout prediction method.It can thus be appreciated that the present embodiment propose based on mixed model
Continuous casting breakout prediction method reached preferable breakout prediction performance, wrong report can be reduced and avoid failing to report, be it is a kind of effectively
Breakout prediction method.
The test result of the even spatial model of 2 groups of table
The present invention considers advantage of the artificial intelligence technology on wave pattern cognition, copper plate of crystallizer when boning with strand
The spatial and temporal variation of interior electric thermo-couple temperature is foundation, using GA-BP neural networks list idol temporal model, for recognizing list
Whether change of the individual electric thermo-couple temperature in time series meets temperature change waveform during bonding, belongs to the knowledge of dynamic waveform pattern
Other problem.Wherein, genetic algorithm optimization BP neural network, by ability of searching optimum BP neural network best initial weights and threshold are determined
Value, improves accuracy of identification of single idol temporal model to sticking temperature waveform pattern.And in view of the two dimension of bonding V-type tearing port
Dissemination, a group even spatial model is established on the basis of single idol temporal model by effective logic rules, judges adjacent
Whether thermocouple has the propagation of bonding, improves the accuracy of identification of sticker type breakout, can especially reduce actual production process
In multiple thermocouple failures or larger electric thermo-couple temperature failing to report and false alarm when fluctuating；By hybrid guided mode proposed by the present invention
Type, makes full use of advantage of the GA-BP neutral nets in wave pattern cognition, and couples effective logic rules judgement, not only
Realize the even space-time of single even summation group to judge, and overcome simple logic judgment model parameter determine it is difficult or inaccurate and
Simple model of mind lacks the deficiency that technique is instructed, and has reached preferable breakout prediction performance, can be promptly and accurately quote whole
Bonding, it is to avoid sticker type breakout accident, and false alarm frequency is minimized into level.
Below schematically to the present invention and embodiments thereof be described, the description does not have restricted, institute in accompanying drawing
What is shown is also one of embodiments of the present invention, and actual structure is not limited thereto.So, if the common skill of this area
Art personnel enlightened by it, in the case of without departing from the invention objective, is designed and the technical scheme without creative
Similar frame mode and embodiment, all should belong to protection scope of the present invention.
Claims (7)
1. it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that mainly including following two parts：Single idol sequential mould
Type and the even spatial model of group；
(1), single idol temporal model；
The structure of single idol temporal model includes：The determination of mode input variable, the determination of model output variable, the pre- place of data
Reason, the foundation of GA-BP neutral nets；
(2) even spatial model, is organized；
1), all electric thermo-couple temperatures on crystallizer are changed over after pattern is identified using single idol temporal model, output
As a result three-dimensional array Y (i, j, t) is saved in, wherein, Y (i, j, t) represents single idol sequential mould of the i-th row j row thermocouple in t
Type recognition result；
2), when Y (i, j, t) is in threshold range [θmin,θmax] it is interior when, it is believed that thermocouple TC (i, j) temperature change meets bonding
Temperature model, marks the thermocouple exception；
3) and then the Y (i, j, t) of all thermocouples of the i-th row is checked, is counted in threshold range [θmin,θmax] in abnormal thermoelectricity
Even number mesh is m, while checking the Y (i-1, j, t) of all thermocouples of the i-th -1 row, is counted in threshold range [θmin,θmax] in it is different
Often thermocouple number is n, and wherein i is more than 1；
If 4), m and n are all higher than being equal to 2, check past 10 seconds menisci row exception thermocouple number whether more than etc.
In 2, if it is satisfied, then alerting thermocouple number threshold values ratio with strike-alarm and bonding respectively using abnormal thermocouple sum m+n
Compared with, carry out strike-alarm and bonding warning judgement.
2. it is according to claim 1 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：Mode input becomes
That what is measured is defined below；
Single thermocouple 30 temperature sampling points continuous in time are selected as mode input variable；
Wherein, the collection period of temperature sampling point is 1 second.
3. it is according to claim 2 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：Model output becomes
That what is measured is defined below；
Model output variable is single thermocouple strike-alarm signal, and model output variable is by above-mentioned 30 temperature samplings point correspondence
The degree of closeness of temperature change waveform pattern and sticking temperature pattern determine；
Wherein, model output variable is the number between -1～2, when the corresponding temperature change waveform pattern of 30 temperature sampling points with
When sticking temperature pattern is identical, model output variable is labeled as 1；When the corresponding temperature variation curve of 30 temperature sampling points
During held stationary, model output variable is labeled as 0.
4. it is according to claim 1 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：The preparation of data
It is as follows with pre-processing；
Sample is extracted from historical data, above-mentioned effective sample is normalized between [- 1,1] using formula (1)；
In formula (1), x' be normalization after sample data, x be normalization before sample data, xmaxIt is sample before normalization
The maximum of data, xminIt is the minimum of a value of sample data before normalization；
Sample data after normalizing by more than is divided into two parts, and, used as training sample, another part is used as survey for a portion
Sample sheet.
5. it is according to claim 1 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：GA-BP is neural
The foundation of network model is as follows；
BP neural network input layer number is 30, representative model input variable；Output layer nodes are 1, representative model output
Variable；One discrimination threshold scope is determined according to the result of network training, when model output variable is not less than threshold set in advance
It is considered as detecting sticking temperature pattern during value scope；Training network adopts 3 layers of BP networks, hidden layer excitation function to use S types
Tangent transfer function, output layer uses linear transfer function, training process to adopt LM optimized algorithms；Selected the number of hidden nodes is
12, obtain the BP neural network model that structure is 30 × 12 × 1；
Wherein：BP neural network learning process includes information forward-propagating and error back propagation, defeated according to given training sample
Enter and output vector constantly learn and adjust the connection weight between neuron and threshold value, make network constantly approach sample input and
Mapping relations between output；The maximum frequency of training of BP neural network is set to 2000, and learning rate is 0.05, and performance error is
0.0001；
Using genetic algorithm optimization BP neural network, GA-BP neural network models are set up, genetic algorithm optimization BP neural network
Basic step includes：1) all weights of BP neural network and threshold value are carried out real coding by, initialization of population, produce individuality, if
Put population scale and evolutionary generation；2), each individual fitness function is calculated according to formula (2), with minimum of a value as optimum；
In formula (2), n is network output layer nodes；yiFor the reality output of i-th node；y'iFor the expectation of i-th node
Output；3), carry out the selection of genetic algorithm, intersect and mutation operation, produce a new generation individuality population；4), evaluate a new generation to plant
Group, judges whether evolutionary generation reaches requirement or whether network error meets condition, fits if meeting and obtaining current population optimum
Answer angle value correspondence individual；Wherein, Population in Genetic Algorithms scale is set to 50, and crossover probability is 0.7, and mutation probability is 0.06, evolves
Algebraically is 200.
6. it is according to claim 1 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：The threshold values model
Enclose [θmin,θmax] it is [0.6,1.3].
7. it is according to claim 1 and 2 it is a kind of for continuous casting bleed-out forecast mixed model, it is characterised in that：Bonding report
Alert thermocouple number threshold values is 6, and bonding warning thermocouple number threshold values is 3；When abnormal thermocouple sum m+n is more than or equal to 3,
Send bonding warning；When abnormal thermocouple sum m+n is more than or equal to 6, strike-alarm is sent.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710206080.XA CN106980729B (en) | 2015-07-24 | 2015-07-24 | A kind of continuous casting breakout prediction method based on mixed model |
| CN201510447779.6A CN105108097B (en) | 2015-07-24 | 2015-07-24 | Hybrid model for continuous casting breakout prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510447779.6A CN105108097B (en) | 2015-07-24 | 2015-07-24 | Hybrid model for continuous casting breakout prediction |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710206080.XA Division CN106980729B (en) | 2015-07-24 | 2015-07-24 | A kind of continuous casting breakout prediction method based on mixed model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105108097A CN105108097A (en) | 2015-12-02 |
| CN105108097B true CN105108097B (en) | 2017-05-10 |
Family
ID=54656329
Family Applications (2)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710206080.XA Active CN106980729B (en) | 2015-07-24 | 2015-07-24 | A kind of continuous casting breakout prediction method based on mixed model |
| CN201510447779.6A Active CN105108097B (en) | 2015-07-24 | 2015-07-24 | Hybrid model for continuous casting breakout prediction |
Family Applications Before (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710206080.XA Active CN106980729B (en) | 2015-07-24 | 2015-07-24 | A kind of continuous casting breakout prediction method based on mixed model |
Country Status (1)
| Country | Link |
|---|---|
| CN (2) | CN106980729B (en) |
Families Citing this family (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105522133B (en) * | 2015-12-21 | 2018-04-06 | 安徽工业大学 | A kind of continuous casting billet bonds the decision method of behavior with copper plate of crystallizer |
| CN108801320B (en) * | 2017-04-28 | 2020-12-01 | 中国石油天然气股份有限公司 | Diagnosis method of natural gas measurement system |
| JP6591509B2 (en) * | 2017-11-06 | 2019-10-16 | 株式会社東芝 | Mold temperature abnormality sign detection device and program |
| CN108097905B (en) * | 2017-12-28 | 2020-02-28 | 北京首钢自动化信息技术有限公司 | Steel leakage early warning method |
| CN108436050B (en) * | 2018-04-16 | 2019-06-21 | 大连理工大学 | A method of continuous cast mold bleed-out is forecast using space density clustering DBSCAN |
| CN109396375B (en) * | 2018-12-11 | 2019-09-27 | 大连理工大学 | A kind of crystallizer bleedout prediction electric thermo method based on feature vector and hierarchical clustering |
| CN109365769B (en) * | 2018-12-18 | 2021-03-30 | 重庆邮电大学 | Crystallizer bleed-out forecasting method based on mixed model judgment |
| CN110523941B (en) * | 2019-08-06 | 2023-03-17 | 中冶南方连铸技术工程有限责任公司 | Continuous casting bonding bleed-out multistage risk control method and control device |
| CN110918973B (en) * | 2019-12-11 | 2021-04-27 | 东北电力大学 | Crystallizer thermal image abnormal region marking method based on run |
| CN111112567B (en) * | 2019-12-16 | 2021-09-21 | 河钢股份有限公司 | Comprehensive judgment method for breakout prediction of slab caster |
| CN111570748B (en) * | 2020-04-28 | 2021-08-06 | 中冶南方连铸技术工程有限责任公司 | Crystallizer bleed-out forecasting method based on image processing |
| CN112926622B (en) * | 2021-01-22 | 2022-10-18 | 大连理工大学 | Crystallizer breakout prediction method for generating countermeasure network based on feature vector and SWGAN-GP |
| CN113579190B (en) * | 2021-06-24 | 2022-10-28 | 邯郸钢铁集团有限责任公司 | Slab continuous casting breakout prediction method based on regional characteristics |
| CN113814369B (en) * | 2021-09-28 | 2022-11-18 | 大连理工大学 | Crystallizer bleed-out forecasting method based on eigenvector and Adaboost integrated model |
| CN113935416B (en) * | 2021-10-15 | 2024-06-25 | 东北大学 | Continuous casting breakout prediction method of multi-state crystallizer |
| CN115446276B (en) * | 2022-10-05 | 2024-07-02 | 大连理工大学 | Continuous casting steel leakage early warning method based on convolutional neural network recognition crystallizer copper plate V-shaped bonding characteristics |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5755782A (en) * | 1991-01-24 | 1998-05-26 | Autogenics | Stents for autologous tissue heart valve |
| CN101477374B (en) * | 2009-01-20 | 2011-11-23 | 东北大学 | Continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network |
| CN101934353A (en) * | 2009-06-30 | 2011-01-05 | 上海宝信软件股份有限公司 | Device and method for breakout prediction during slab continuous casting |
| CN101985166B (en) * | 2009-07-29 | 2012-04-25 | 宝山钢铁股份有限公司 | Continuous casting slab longitudinal-crack steel leakage prediction method |
| CN101850410B (en) * | 2010-06-22 | 2012-06-20 | 攀钢集团钢铁钒钛股份有限公司 | Continuous casting breakout prediction method based on neural network |
| CN102151814B (en) * | 2010-12-30 | 2012-11-28 | 中冶连铸技术工程股份有限公司 | Bonding alarm method and system during continuous casting production |
| CN102554171B (en) * | 2011-12-21 | 2014-04-23 | 天津钢铁集团有限公司 | Breakout prediction method for continuous casting |
| CN102699302B (en) * | 2012-07-10 | 2014-01-22 | 中冶赛迪电气技术有限公司 | Bleed-out forecasting system and forecasting method of slab continuous casting crystallizer |
| CN103878335B (en) * | 2014-02-10 | 2015-11-04 | 河北钢铁股份有限公司邯郸分公司 | A kind of method of preventing conticaster crystallizer cohesiveness bleed-out |
| CN104484715A (en) * | 2014-11-28 | 2015-04-01 | 江苏大学 | Neural network and particle swarm optimization algorithm-based building energy consumption predicting method |
-
2015
- 2015-07-24 CN CN201710206080.XA patent/CN106980729B/en active Active
- 2015-07-24 CN CN201510447779.6A patent/CN105108097B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN105108097A (en) | 2015-12-02 |
| CN106980729A (en) | 2017-07-25 |
| CN106980729B (en) | 2018-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105108097B (en) | Hybrid model for continuous casting breakout prediction | |
| He et al. | Mold breakout prediction in slab continuous casting based on combined method of GA-BP neural network and logic rules | |
| CN107054410B (en) | The intelligent diagnosis system and diagnostic method of point machine | |
| US11105758B2 (en) | Prediction method for mold breakout based on feature vectors and hierarchical clustering | |
| CN105328155B (en) | A kind of bleed-out visual feature forecasting procedure based on improvement neutral net | |
| CN115657617A (en) | Implementation method of intelligent monitoring alarm system for thermal power plant | |
| CN108436050B (en) | A method of continuous cast mold bleed-out is forecast using space density clustering DBSCAN | |
| CN109365769A (en) | A kind of crystallizer bleedout prediction electric thermo method based on mixed model judgement | |
| CN102554171B (en) | Breakout prediction method for continuous casting | |
| CN113570200B (en) | Power grid running state monitoring method and system based on multidimensional information | |
| CN107159862B (en) | A kind of crystallizer bleedout prediction electric thermo method that logic-based judges | |
| CN107742008A (en) | A kind of fault early warning method of gearbox of wind turbine | |
| CN106384186A (en) | Distributed new energy grid-connected power quality monitoring method and system | |
| CN107096899B (en) | A kind of crystallizer bleedout prediction electric thermo system that logic-based judges | |
| CN108454653A (en) | A kind of train control system communicated based on satellite communication and public cellular | |
| CN104267610B (en) | High-precision blast furnace smelting process abnormal data detection and repair method | |
| CN105522133A (en) | Judgment method of bonding behavior between continuous casting slab and crystallizer copper plate | |
| CN117892882A (en) | Huang Heba bank monitoring alarm risk reporting system and method based on network communication platform | |
| CN112434887B (en) | Water supply network risk prediction method combining network kernel density estimation and SVM | |
| CN109902265A (en) | A kind of underground method for early warning based on hidden Markov model | |
| CN115446276B (en) | Continuous casting steel leakage early warning method based on convolutional neural network recognition crystallizer copper plate V-shaped bonding characteristics | |
| CN113935416B (en) | Continuous casting breakout prediction method of multi-state crystallizer | |
| CN113435794B (en) | Intelligent monitoring method for post-processing of nozzle casting based on image processing | |
| CN104697585A (en) | Cooling tower tower-barrel strength predicating control system and method thereof | |
| CN114925867A (en) | Internet of things maintenance prediction system for key equipment in seamless steel tube production |
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 |