CN108388762A - Sinter chemical composition prediction technique based on depth confidence network - Google Patents
Sinter chemical composition prediction technique based on depth confidence network Download PDFInfo
- Publication number
- CN108388762A CN108388762A CN201810188530.1A CN201810188530A CN108388762A CN 108388762 A CN108388762 A CN 108388762A CN 201810188530 A CN201810188530 A CN 201810188530A CN 108388762 A CN108388762 A CN 108388762A
- Authority
- CN
- China
- Prior art keywords
- data
- chemical composition
- sinter
- prediction
- dbn
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/30—Prediction of properties of chemical compounds, compositions or mixtures
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Crystallography & Structural Chemistry (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of sinter chemical composition prediction techniques based on depth confidence network.This method predicts sinter chemical composition according to sinter mixture chemical composition, using based on the prediction technique of DBN algorithms；Specifically include following steps：The historical data of sintering plant actual production, rejecting abnormalities data and normalized are obtained first；Then determining influences the input/output argument of sinter quality, and the reasonability of input parameter is examined using gray relative analysis method；Resettle the sinter chemical composition prediction model based on DBN, and the training of usage history Data Data, optimal prediction model；Sinter chemical composition is finally predicted with this prediction model, and to result anti-normalization processing, obtains sinter chemical composition predicted value.Compared with prior art, the present invention is based on the prediction models of DBN can more accurately realize approaching for complex nonlinear function, improves sinter chemical composition precision of prediction, has application and popularization value in actual production.
Description
Technical field
The invention belongs to iron and steel smelting technology field, be related to a kind of sinter chemical composition prediction technique, more particularly to one
Sinter chemical composition prediction technique of the kind based on depth confidence network.
Background technology
Sinter is the primary raw material of blast furnace ironmaking, and the chemical composition of sinter is to evaluate and test the important finger of sinter quality
Mark.Dispensing is the first working procedure of sintering production, has significant impact to the chemical composition of sinter.Due to sinter chemical raw material
Source is wide, wide in variety, complicated component, and sintering process has the characteristics that lag, close coupling, non-linear so that being sintered mineralising for a long time
Study point is difficult to accurately control.In matching formulation process in dispensing, sinter chemical composition is accurately predicted, in time
Raw material proportioning is adjusted, sinter quality is improved and is of great significance.
In sinter chemical composition forecasting research, with the development of computer technology, some shallow-layer intelligent forecast models
Wide research and application had been obtained in recent years, improved precision of prediction.Dragon it is red it is bright et al. using with momentum term it is linear again
Step-varied back propagation BP neural network algorithm is encouraged, the sinter chemical composition forecast mould based on multicycle operational mode is established
Type；Model knows intelligent et al. the advantages of combining gray prediction and neural net prediction method, establishes the burning based on grey neural network
Tie mine chemical composition prediction model；Song Qiang et al. proposes the soft of the sinter chemical composition based on least square method supporting vector machine
The forecasting model of sinter chemical composition is established in the research of measurement model using support vector machines.The god mentioned in above-mentioned document
Through network, gray theory and support vector machines etc. belong to shallow-layer learning algorithm, sample of the shallow-layer learning algorithm in given limited quantity
This when is it is difficult to effectively indicate that non-linear complicated function, generalization ability are restricted, and then influence sinter chemical composition
Prediction result.
Deep learning is a kind of Multilayer Perception structure algorithm of simulation human brain.Relative to shallow-layer learning method, depth
Study can more accurately realize approaching for complex nonlinear function, in recent years by learning a kind of deep layer nonlinear network structure
Through having obtained effective application in many fields.Depth confidence network (Deep Belief Network, abbreviation DBN) is a kind of normal
The frame of deep learning.
Invention content
It is an object of the present invention to overcome above-mentioned the deficiencies in the prior art, a kind of burning based on depth confidence network is provided
Tie mine chemical composition prediction technique.This method can fully excavate the substantive characteristics of sintering process by establishing DBN prediction models,
Improve sinter chemical composition precision of prediction.
The purpose of the present invention is what is be achieved through the following technical solutions：
One kind being based on the sinter chemical composition prediction technique of depth confidence network (i.e. DBN), and this method is：It is former in sintering
After the completion of expecting dispensing, mixture is obtained, according to mixture chemical composition, using based on depth confidence network (i.e. DBN) algorithm
Prediction technique predicts sinter chemical composition, and to examine the accuracy matched in sintering burden process, adjustment proportioning, reaches in time
To the purpose for improving sinter quality.This method specifically includes following steps：
S1：The historical data for obtaining sintering plant actual production, pre-processes the data of acquisition, rejecting abnormalities data are simultaneously
It is normalized；
S2：Determining influences the input/output argument of sinter quality, and input parameter is examined using gray relative analysis method
Reasonability；
S3：Establish the sinter chemical composition prediction model based on DBN, using the data in step S1 to prediction model into
Row training, optimal prediction model；
S4：The prediction model obtained with step S3 predicts that sinter chemical composition, the data obtained after prediction exist
Between [0,1], then anti-normalization processing is carried out to the data that these are obtained, obtains sinter chemical composition predicted value.
Further, in the step S1：The historical data for obtaining sintering plant actual production includes mixture chemical composition
And its corresponding sinter chemical composition.Abnormal data is divided into two classes, and one kind is due to system communication mistake or software error
The abnormal gathered data Deng caused by.This when sintering process may normal operation, but the data that arrive of system acquisition are not
With reference significance, so rejecting this partial data.Another kind of is abnormal data caused by unusual service condition.The prediction mould of the present invention
Type is built upon the data acquired under nominal situation, and unusual service condition can have an impact sintering process, the change under unusual service condition
Relationship between mixture and sinter cannot normally be reflected by testing result (abnormal data).Above-mentioned two classes abnormal data is required for picking
It removes.
For the above situation, the data under nominal situation are obtained after rejecting above-mentioned two classes abnormal data.In view of input and
The difference of data level and data dimension between output data, in order to ensure the efficiency and precision of prediction model, it is unified to data into
Row normalized.Normalizing formula is：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor data sequence
Maximum value in row, XminFor minimum value in data sequence.
Data are normalized, are transformed data between [0,1].DBN algorithms are using softmax functions
When, to avoid network from being operated in the flat site of function, [0.1,0.9] is transformed data to, with following formula：
Further, in the step S2：According to practical production experience and inspection information, selection directly affects sintering mineral
The chemical composition of amount determines output parameter and influences the input parameter of output parameter.
To increase the reliability of prediction model input parameter selection, the present invention is using Grey Incidence Analysis to above-mentioned ginseng
Whether number and Prediction Parameters are analyzed, reasonable to verify above-mentioned routine testing parameter, and remove unreasonable parameter to determine
Input parameter.Grey Incidence Analysis directly ignores the information being unable to get, and different system is weighed at any time using the degree of association
Between the correlation size that changes, reflection is variation tendency between two sequences, if two sequence variation trend have it is consistent
Property, then the two correlation degree is higher, otherwise correlation degree is then relatively low.Grey Incidence Analysis be mode input parameter really
Surely theoretical foundation is provided.Grey correlation analysis formula is as follows：
In formula, r represents the degree of association of different data sequence, and between value 0 to 1, that reflects the phases of two data sequences
Pass degree, r values illustrate that correlation is better closer to 1.Y (k) is reference data sequence,To compare data sequence, ρ is point
Distinguish coefficient, degree of association distortion effect when data sequence differs greatly compared with for weakening reference data sequence.
Sequences y (k) andRespectively original reference data sequence x (k) and original relatively data sequenceAt equalization
Data sequence after reason, equalization handle the influence that can eliminate data dimension in data sequence, and it is as follows that equalization handles formula：
Further, in the step S3, the method for establishing the sinter chemical composition prediction model based on DBN is as follows：
DBN model network structure is determined first.DBN model there is no the relevant theory of optimum network structure, actual experiment to grind
It is preferable using 3-6 layers of RBM (restricted Boltzmann machine) effect structure to study carefully discovery, optimal mould can get by Commissioning Contrast repeatedly
The type number of plies.
DBN model parameter is followed by set.Model parameter includes the number of hidden nodes, iterations, batch size, study effect
Rate, train epochs and excitation function etc., parameter designing are to consider the result of precision of prediction and training time.
Finally DBN model is trained.DBN model is trained using the data in step S1, training method is：
First use unsupervised greedy algorithm pre-training model；Then by the entire model of BP algorithm backpropagation optimization, (i.e. weights are micro- again
It adjusts).That is,：The model training process is broadly divided into following 2 step, and the detailed process of each step is as follows：
Step1：Model pre-training:
First layer RBM is trained up using the Hinton unsupervised greedy algorithms proposed and obtains weight w1, one can be obtained
The output of a distribution similar with input data, this process is feature extraction.Keep w1 constant, by first layer RBM hidden layers
The input as second layer RBM visual layers is exported, second RBM of training is continued.This process is repeated until all RBM have been trained
At obtaining weight w2 and w3, it is ensured that as more as possible when maps feature vectors are to different feature spaces during Feature Mapping
Keeping characteristics information.Finally be mapped to output layer, output layer excitation function is softmax functions, by with data label pair
Than random initializtion output layer weight w.
Step2：Weights are finely tuned:
After Step1 pre-training, the initial weight w1, w2, w3, w of network have been obtained.By DBN last
Layer setting BP networks finely tune initial weight with having supervision according to sample data label backpropagation.In trim process, use is following
Object function：
Wherein, θ={ w1 w2 w3 w }, that is, need the weights finely tuned, yiFor sample label, yi' it is prediction result.Because of w
It is random initializtion, so needing first to adjust the value of w in trim process, several times after iteration, then finely tunes w1, the value of w2, w3.
When DBN trains each layer of RBM, only ensure that mapping weights optimize in the RBM, but regardless of other RBM weights, because
This, predicted value and sample data tag error are propagated back to whole network, hierarchical optimization DBN model by BP algorithm.DBN is trained
Process is considered as the initialization to a deep layer BP network weight parameter, the process prevent DBN networks occur BP networks because
Random initializtion weighting parameter and be easily trapped into local optimum and the training time length disadvantage.
Further, in the step S4, after being predicted with the DBN model of step S3, obtained data are in [0,1]
Between, then by carrying out anti-normalization processing to these data, obtain the predicted value (prediction data) of sinter chemical composition.
Anti-normalization processing formula is：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor data sequence
Maximum value in row, XminFor minimum value in data sequence.
Further, prediction model finally is verified using the test sample data in step S1, predicted value is respectively adopted
The precision of prediction of prediction model is verified with original value matched curve, Relative Error curve.
Beneficial effects of the present invention：
Compared with prior art, advantage of the invention is as follows：
(1) by analyzing mechanism of sintering process, rejecting abnormal data and normalized is carried out to training data, improved
The reliability of training data.
(2) it uses Grey Incidence Analysis to analyze input and output parameter degree of being associated, demonstrates input
The reasonability of parameter selection.
(3) prediction technique based on depth learning technology DBN, can be more approximate using the more traditional shallow-layer network of deep layer network
Ground is fitted the advantage of complex nonlinear function, improves the precision of prediction of sinter chemical composition.
Description of the drawings
Fig. 1 is the sinter chemical composition prediction technique flow chart of the present invention；
Fig. 2 is the sinter chemical composition forecasting system structure chart in the present invention；
Fig. 3 is the DBN prediction model figures in the present invention；
Fig. 4 is the DBN prediction result TFe matched curve figures in the present invention；
Fig. 5 is the DBN prediction result TFe error opposing curves figures in the present invention.
Specific implementation mode
The invention will be further described in the following with reference to the drawings and specific embodiments.The present embodiment is with technical solution of the present invention
Premised on implemented, but do not limit the scope of the invention.
As shown in Figure 1 and Figure 2, the embodiment of the present invention provides a kind of sinter chemistry based on depth confidence network (i.e. DBN)
Ingredient prediction method, this method are：After the completion of raw materials for sintering dispensing, obtaining mixture, (i.e. raw materials for sintering mixture, also known as burns
Tie mixture), according to mixture chemical composition, using predicting sinter chemical composition based on the prediction technique of DBN algorithms, with
Examine the accuracy matched in sintering burden process.This method is in sintering process, according to sinter mixture ingredient come pre- in advance
Sinter chemical composition is reported, to adjust dispensing proportioning in time, improves sinter quality.This method is as follows：
S1：The historical data for obtaining sintering plant actual production, pre-processes the data of acquisition, rejecting abnormalities data are simultaneously
It is normalized.
Obtain the historical data of certain domestic large-scale steel mill sintering plant actual production, including sinter mixture chemical composition and its
Corresponding sinter chemical composition.By actual production collection in worksite to data include abnormal data, need to reject.Abnormal data
It is divided into two classes, one kind is the data acquisition abnormity caused by system communication mistake or software error etc..This when is sintered
The possible normal operation of process, but the data that system acquisition arrives have not had reference significance, so rejecting this partial data.Separately
One kind is data exception caused by unusual service condition.The prediction model of the present invention is built upon the data acquired under nominal situation,
Unusual service condition can have an impact sintering process, and the result of laboratory test (abnormal data) under unusual service condition cannot normally reflect mixture
Relationship between sinter.Unusual service condition includes following several：The data acquired under equipment fault in sintering process belong to different
Normal floor data；Supervisor's work negative pressure normal range (NR) be 16.5~17kPa, if there is fluctuation, not herein within the scope of
Belong to unusual service condition；It is abnormal operating mode that machine speed, which is less than 1m/s, when sintering, is considered as parking；One mix water be 20 tons/t~
Between 25 tons/t, if not belonging to unusual service condition data in the data of this range；Two mix water normal range (NR) be 5 tons/t~
15 tons/t, it is more than 15 tons/t and belongs to unusual service condition data less than 5 tons/t；Firing temperature just belongs to not just less than 800 DEG C
Normal range, the data acquired under this unusual service condition need to reject.After data screening, 905 groups of data, part number is obtained
According to as shown in table 1 below.Wherein 700 groups of data it will to be used for model training, remaining 205 groups of data is used for test sample.
1 part modeling of table emulates data
The data under nominal situation are obtained after rejecting two class abnormal datas for the above situation.In view of outputting and inputting number
The difference of data level and data dimension between uniformly carries out normalizing to ensure the efficiency and precision of prediction model to data
Change is handled.Normalizing formula is：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor data sequence
Maximum value in row, XminFor minimum value in data sequence.
Normalized is transformed data between [0,1].DBN algorithms are when using softmax functions, to avoid net
Network is operated in the flat site of function, [0.1,0.9] is transformed data to, with following formula：
S2：Determining influences the input/output argument of sinter quality, and input parameter is examined using gray relative analysis method
Reasonability.
According to practical production experience and related data selection input/output argument is consulted, selection directly affects sinter quality
Sinter chemical composition as output parameter.The TFe contents of sinter are the important indicators of blast fumance, are directly related to height
Stove production capacity, energy consumption consumption etc., play an important role to blast fumance.The fluctuation of TFe contents will have a direct impact on blast furnace process production
The fluctuation of the working of a furnace, therefore, using sinter TFe contents as one of Prediction Parameters of prediction model.Sinter basicity is that evaluation is burnt
Another important indicator of mineral amount is tied, the height of sinter basicity and the basicity of clinker are closely bound up, the calculation formula of basicity
For CaO/SiO2, sinter basicity is codetermined by CaO content and SiO2 contents, therefore, the present invention by CaO content and
Other two Prediction Parameters of SiO2 contents as prediction model.And in the production specification of sintering plant, all it is distinctly claimed control
The stability of all iron content processed and basicity, TFe contents, CaO content and SiO2 contents can be than more comprehensively reflecting sintering mineral
The quality of amount.Therefore, using TFe contents, CaO content and SiO2 contents in sinter as sinter chemical composition prediction model
Prediction Parameters are suitable.The chemical composition of sinter determines by sinter mixture ingredient, according to actual production data and
Related data is consulted it is found that TFe, FeO, CaO, SiO2, charcoal, MgO, Al2O3, the mixtures ingredient such as S, P it is practical raw as sintering plant
TFe, CaO, SiO in routine testing project and sinter during production2Content is related.
In order to increase the reliability of prediction model input parameter selection, the present invention is using Grey Incidence Analysis to above-mentioned
Whether parameter and Prediction Parameters are analyzed, reasonable to verify above-mentioned routine testing parameter, and are removed unreasonable parameter and come really
Determine input parameter.Grey Incidence Analysis directly ignores the information being unable to get, using the degree of association come weigh different system with
The correlation size of time change, reflection is variation tendency between two sequences, if two sequence variation trend have one
Cause property, then the two correlation degree is higher, otherwise correlation degree is then relatively low.Grey Incidence Analysis is mode input parameter
Determination provides theoretical foundation.Grey correlation analysis formula is as follows：
In formula, r represents the degree of association of different data sequence, and that reflects the tightness degree of two data sequences, y (k) is
Reference data sequence,To compare data sequence, ρ is resolution ratio, for weakening reference data sequence data sequence compared with
Degree of association distortion effect when row differ greatly, the present embodiment ρ take 0.5.Wherein, sequences y (k) andRespectively original reference
Data sequence x (k) and original relatively data sequenceEqualization treated data sequence, it is as follows that equalization handles formula：
In the present embodiment, reference data sequence y (k) be respectively sinter chemical composition TFe contents, CaO content and
The data sequence that SiO2 content original reference data sequences obtain after equalization is handled, compares data sequenceThen distinguish
It is mixture ingredient TFe contents, FeO contents, CaO content, SiO2 contents, carbon content, content of MgO, Al2O3 contents, S contents, P
The original data sequence that relatively data sequence obtains after homogenize process of content.
Grey relational grade (r) is calculated according to above formula using the data sequence in S1 after screening, as a result such as 2 institute of table
Show：
2 mixture ingredient of table and sinter chemical composition grey relational grade (r)
| TFe | FeO | CaO | SiO2 | Charcoal | MgO | Al2O3 | S | P | |
| TFe | 0.7602 | 0.7324 | 0.7102 | 0.7239 | 0.7530 | 0.636 | 0.6101 | 0.5767 | 0.5465 |
| CaO | 0.7210 | 0.7027 | 0.7859 | 0.7268 | 0.7159 | 0.6403 | 0.5769 | 0.5018 | 0.4755 |
| SiO2 | 0.7176 | 0.7015 | 0.7286 | 0.7963 | 0.7097 | 0.6451 | 0.5658 | 0.4638 | 0.4963 |
As can be seen from the above table, TFe contents, FeO contents, CaO content, SiO in mixture ingredient2Content, carbon content and
Content of MgO respectively with TFe contents, CaO content and SiO in sinter chemical composition2The association angle value of content all up to 0.7 or more,
Degree of correlation is larger, and the association angle value of remaining ingredient is less than 0.7.Gray system theory thinks, when the degree of association reaches 0.7 or more
When, it is believed that the two is strong correlation, therefore in the present embodiment, select TFe contents, FeO contents, CaO in mixture ingredient to contain
Amount, SiO2Content and carbon content are as prediction model input parameter.
S3：The sinter chemical composition prediction model (abbreviation DBN prediction models) based on DBN is established, uses step S1's
Data are trained model, optimal prediction model.
The method for establishing the sinter chemical composition prediction model based on DBN is as follows：
First, prototype network structure is determined.DBN model there is no the relevant theory of optimum network structure, actual experiment research
It was found that preferable using 3-6 layers of RBM effect structures.The present invention has found in debugging process repeatedly, under identical sintering data qualification,
It is promoted obviously, and worked as compared with using the DBN model precision of prediction of 2 layers of RBM structures using the DBN model precision of prediction of 3 layers of RBM structures
When using 4 layers or more layer RBM structures, precision of prediction, which has almost no change, even to be declined, and can increase the training time.Therefore this mould
Type uses 3 layers of RBM structures, DBN prediction models as shown in Figure 3.
DBN model parameter is followed by set.This model carries out emulation experiment using MATLAB, and DBN model parameter designing is such as
Under, hidden layer number of nodes dbn.sizes is set as [5 5 5], i.e., 3 layers of RBM altogether, and every layer of hidden layer node number is 5；Iteration
Number dbn.numepochs is set as 1；Batch size dbn.batchsize is set as 50；Learning rate dbn.alpha is set as 0.1；Swash
Function is encouraged to be selected as being particularly suited for polytypic softmax functions.Parameter designing is to consider precision of prediction and training time
Result.
Finally, model is trained using 700 groups of training datas in step S1.Using the data in step S1 to model
It is trained, which is broadly divided into 2 steps：
Step1：Model pre-training:
First layer RBM is trained up using the Hinton unsupervised greedy algorithms proposed and obtains weight w1, one can be obtained
The output of a distribution similar with input data, this process is feature extraction.Keep w1 constant, by first layer RBM hidden layers
The input as second layer RBM visual layers is exported, second RBM of training is continued.This process is repeated until all RBM have been trained
At obtaining weight w2 and w3, it is ensured that as more as possible when maps feature vectors are to different feature spaces during Feature Mapping
Ground keeping characteristics information.Finally be mapped to output layer, output layer excitation function is softmax functions, by with data label pair
Than random initializtion output layer weight w.
Step2：Weights are finely tuned:
After Step1 model pre-training, the initial weight w1, w2, w3, w of network have been obtained.By DBN most
BP networks are arranged in later layer, and initial weight is finely tuned with having supervision according to sample data label backpropagation.In trim process, use
Following object function：
Wherein, θ={ w1w2w3w }, that is, need the weights finely tuned, yiFor sample label, yi' it is prediction result.Because w is
Random initializtion, so needing first to adjust the value of w in trim process, several times after iteration, then w1 is finely tuned, the value of w2, w3.
When DBN trains each layer of RBM, only ensure that mapping weights optimize in the RBM, but regardless of other RBM weights, because
This, BP algorithm is by predicted value and sample label error back propagation to whole network, and hierarchical optimization DBN model (successively weigh by fine tuning
Value).DBN training process is considered as the initialization to a deep layer BP network weight parameter, which prevents DBN networks to go out
Existing BP networks are easily trapped into the disadvantage of local optimum and training time length because of random initializtion weighting parameter.
S4：The prediction model obtained with step S3 predicts that sinter chemical composition, the data obtained after prediction exist
Between [0,1], then anti-normalization processing is carried out to the data that these are obtained, obtains sinter chemical composition predicted value.
Anti-normalization processing formula is：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor data sequence
Maximum value in row, XminFor minimum value in data sequence.
To illustrate to use the test sample in step S1 with the precision of prediction and validity of comparison prediction model, the present embodiment
Data verify prediction model, predicted value and original value matched curve are respectively adopted, Relative Error curve verifies prediction
The result of model.By taking TFe (TFe is special iron, refers to the ferroalloy containing other alloying elements) prediction result as an example, predicted value
It is as shown in Figure 4 and Figure 5 with original value matched curve, Relative Error curve.
From fig. 5, it can be seen that the Relative Error absolute value of TFe is within 2%, it is seen that the precision of prediction of TFe is higher.
Further to verify the validity of proposed DBN prediction models, using same data by this DBN prediction models
Prediction result compared with BP neural network model and support vector machines (SVM) model.
Wherein, the setting of BP neural network model parameter is as follows：Factor of momentum is set as 0.30, and study initial rate is set as
0.70, initial weight and Initial Hurdle are all set as 0.50, and excitation function usesFunction.For SVM moulds
Type, present invention selection obtain most widely used RBF kernel functions in SVMAnd it uses
Cross-validation method Select Error penalty factor and RBF kernel function width optimum values.
After prediction, the mean square error (MSE) and average absolute percentage of three kinds of prediction model results have been calculated separately
Both more common evaluation indexes of ratio error (MAPE).MSE and MAPE are defined as follows：
(1) mean square error (MSE)
Wherein ykFor original value,For predicted value.MSE can reflect the influence of big error, can evaluate the variation of data
Degree, the smaller then error of MSE values is smaller, which is one of most common index of evaluation model performance.
(2) mean absolute percentage error (MAPE)
Wherein ekFor relative error.What MAPE reflected is the precision of prediction of model, and value is smaller to illustrate that precision of prediction is higher.
By taking TFe prediction results as an example, each the model calculation comparison is as shown in the table：
3 TFe prediction model performance indicator contrast tables of table
| Model | DBN | BP neural network | SVM |
| MSE | 0.1796 | 0.6201 | 0.2310 |
| MAPE | 0.0059 | 0.0131 | 0.0083 |
As can be seen from the above table, the deep layer Network Prediction Model error criterion proposed by the present invention based on DBN is better than SVM
With the shallow-layer Network Prediction Model of BP neural network, this shows validity of the DBN in sinter chemical composition prediction, and predicts
It works well.
Claims (10)
1. a kind of sinter chemical composition prediction technique based on depth confidence network, which is characterized in that in raw materials for sintering dispensing
After the completion, mixture is obtained, according to mixture chemical composition, using the prediction technique based on depth confidence network, that is, DBN algorithms
Predict sinter chemical composition, to examine the accuracy matched in sintering burden process, adjustment proportioning, improves sinter in time
Quality；
This method specifically includes following steps：
S1：The historical data for obtaining sintering plant actual production, pre-processes the data of acquisition, rejecting abnormalities data simultaneously carry out
Normalized；
S2：Determining influences the input/output argument of sinter quality, and the conjunction of input parameter is examined using gray relative analysis method
Rationality；
S3：The sinter chemical composition prediction model based on DBN is established, prediction model is instructed using the data in step S1
Practice, optimal prediction model；
S4：The prediction model obtained with step S3 predicts sinter chemical composition, and the data obtained after prediction are in [0,1]
Between, then anti-normalization processing is carried out to the data that these are obtained, obtain sinter chemical composition predicted value.
2. the sinter chemical composition prediction technique according to claim 1 based on depth confidence network, which is characterized in that
In the step S1：
The historical data for obtaining sintering plant actual production includes mixture chemical composition and its corresponding sinter chemical composition；It is different
Regular data includes two classes, and one kind is the abnormal gathered data caused by system communication mistake or software error；It is another kind of to be
Abnormal data caused by unusual service condition；Above-mentioned two classes abnormal data will reject；
Data are normalized, are transformed data between [0,1]；DBN algorithms using softmax functions when,
To avoid network from being operated in the flat site of function, [0.1,0.9] is transformed data to, with following formula：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor in data sequence
Maximum value, XminFor minimum value in data sequence.
3. the sinter chemical composition prediction technique according to claim 1 based on depth confidence network, which is characterized in that
In the step S2：
According to practical production experience and inspection information, selection directly affects the chemical composition of sinter quality, determines input, output
Parameter；It is whether reasonable that input parameter is verified using following grey correlation analysis formula：
In formula, r represents the degree of association of different data sequence, and that reflects the tightness degree of two data sequences, y (k) is reference
Data sequence,To compare data sequence, ρ is resolution ratio, poor for weakening reference data sequence data sequence compared with
Degree of association distortion effect when different larger；
Sequences y (k) andRespectively original reference data sequence x (k) and original relatively data sequenceAfter equalization processing
Data sequence, equalization, which is handled, can eliminate the influence of data dimension in data sequence, and it is as follows that equalization handles formula：
4. the sinter chemical composition prediction technique according to claim 1 based on depth confidence network, which is characterized in that
In the step S3, the method for establishing the sinter chemical composition prediction model based on DBN is as follows：
DBN model network structure is determined first；The DBN model number of plies is obtained by debugging repeatedly, using 3-6 layers of RBM structures；
DBN model parameter is followed by set；Model parameter include the number of hidden nodes, iterations, batch size, learning efficiency,
Train epochs and excitation function, parameter designing are to consider the result of precision of prediction and training time；
Finally, DBN model is trained using the data in step S1：First use unsupervised greedy algorithm pre-training model；
Then entire model is optimized by BP algorithm backpropagation again.
5. the sinter chemical composition prediction technique according to claim 4 based on depth confidence network, which is characterized in that
In step S3, the DBN model network structure, using 3 layers of RBM structures.
6. the sinter chemical composition prediction technique according to claim 5 based on depth confidence network, which is characterized in that
Method using unsupervised greedy algorithm pre-training model is as follows：First layer is trained up using unsupervised greedy algorithm
RBM obtains weight w1, obtains the output of a distribution similar with input data, this process is feature extraction；Keep w1 not
Become, first layer RBM hidden layers are exported into the input as second layer RBM visual layers, continues second RBM of training；Repeat this mistake
RBM training all Cheng Zhizhi are completed, and weight w2 and w3 are obtained, it is ensured that maps feature vectors are to different during Feature Mapping
Feature space when, keeping characteristics information as much as possible；It is finally mapped to output layer, output layer excitation function is softmax
Function, by being compared with data label, random initializtion output layer weight w；After above-mentioned pre-training model, net has been obtained
The initial weight w1, w2, w3, w of network.
7. the sinter chemical composition prediction technique according to claim 6 based on depth confidence network, which is characterized in that
The method for optimizing entire model by BP algorithm backpropagation is as follows：By last layer in DBN, BP networks, root are set
Initial weight is finely tuned according to sample data label backpropagation with having supervision；In trim process, following object function is used：
Wherein, θ={ w1 w2 w3 w }, that is, need the weights finely tuned, yiFor sample label, yi' it is prediction result；Because w be with
Machine initialization, so needing first to adjust the value of w in trim process, several times after iteration, then w1 is finely tuned, the value of w2, w3；
When DBN trains each layer of RBM, only ensure that mapping weights optimize in the RBM, but regardless of other RBM weights；BP algorithm will
Predicted value propagates back to whole network, hierarchical optimization DBN model with sample data tag error.
8. the sinter chemical composition prediction technique according to claim 1 based on depth confidence network, which is characterized in that
Anti-normalization processing formula in the step S4 is：
In formula, X ' is each data sequence after normalized, and X is each data sequence before normalized, XmaxFor in data sequence
Maximum value, XminFor minimum value in data sequence.
9. the sinter chemical composition prediction technique according to claim 3 based on depth confidence network, which is characterized in that
In step S2, using TFe contents, CaO content and SiO2 contents in sinter as sinter chemical composition prediction model
Prediction Parameters select TFe contents, FeO contents, CaO content, SiO in mixture2Content and carbon content are inputted as prediction model
Parameter.
10. the sinter chemical composition prediction technique according to claim 1 based on depth confidence network, feature exist
In finally, prediction model being verified using the test sample data in step S1, predicted value is respectively adopted and original value fitting is bent
Line, Relative Error curve verify the precision of prediction of prediction model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810188530.1A CN108388762A (en) | 2018-03-07 | 2018-03-07 | Sinter chemical composition prediction technique based on depth confidence network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810188530.1A CN108388762A (en) | 2018-03-07 | 2018-03-07 | Sinter chemical composition prediction technique based on depth confidence network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108388762A true CN108388762A (en) | 2018-08-10 |
Family
ID=63066893
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810188530.1A Pending CN108388762A (en) | 2018-03-07 | 2018-03-07 | Sinter chemical composition prediction technique based on depth confidence network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108388762A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109596649A (en) * | 2018-11-29 | 2019-04-09 | 昆明理工大学 | A kind of method and device that host element concentration is influenced based on convolutional network coupling microalloy element |
| CN109784276A (en) * | 2019-01-16 | 2019-05-21 | 东南大学 | A kind of voltage dip feature extraction based on DBN and temporarily drop source discrimination method |
| CN110070217A (en) * | 2019-04-11 | 2019-07-30 | 武汉科技大学 | A kind of Forcasting Sinter Quality method of Kernel-based methods parameter |
| CN110218862A (en) * | 2019-05-09 | 2019-09-10 | 山东钢铁股份有限公司 | A kind of staflux is with addition of adjusting method |
| CN110533082A (en) * | 2019-08-09 | 2019-12-03 | 东北大学秦皇岛分校 | A kind of sintering, mixing and water adding control method based on dual model collaborative forecasting |
| CN112070408A (en) * | 2020-09-14 | 2020-12-11 | 唐山学院 | Agglomerate composition forecasting model based on big data and deep learning |
| CN113095561A (en) * | 2021-04-06 | 2021-07-09 | 上海交通大学 | Method for predicting FeO content of sinter by multi-source information and long-and-short time memory network |
| CN113517037A (en) * | 2021-03-25 | 2021-10-19 | 中南大学 | Method and system for predicting sintering ore FeO by fusing data and knowledge |
| US11221610B1 (en) | 2019-09-30 | 2022-01-11 | General Electric Company | Optimized support design for sintering parts with complex features |
| CN114697012A (en) * | 2020-12-30 | 2022-07-01 | 科大国盾量子技术股份有限公司 | Real-time wide area user node quantum key secret communication method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104757992A (en) * | 2015-03-16 | 2015-07-08 | 广东工业大学 | Cardiac sound diagnostic system based on depth confidence network and diagnostic method |
| CN105825230A (en) * | 2016-03-15 | 2016-08-03 | 中国农业大学 | Soil moisture status prediction method and system based on DBN (Deep Belief Network) model |
| CN106202946A (en) * | 2016-07-18 | 2016-12-07 | 燕山大学 | Clinker free calcium levels Forecasting Methodology based on degree of depth belief network model |
| CN106326634A (en) * | 2016-08-09 | 2017-01-11 | 浙江工业大学 | Hospital outpatient quantity predicting method based on deep belief network |
| CN106407690A (en) * | 2016-09-28 | 2017-02-15 | 厦门理工学院 | Outpatient number prediction method and system based on automatic deep belief network |
| CN106769048A (en) * | 2017-01-17 | 2017-05-31 | 苏州大学 | Self adaptation depth confidence network Method for Bearing Fault Diagnosis based on Nesterov momentum methods |
| CN106802977A (en) * | 2016-12-14 | 2017-06-06 | 同济大学 | One kind is used for sintering performance index prediction and Quality evaluation method |
-
2018
- 2018-03-07 CN CN201810188530.1A patent/CN108388762A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104757992A (en) * | 2015-03-16 | 2015-07-08 | 广东工业大学 | Cardiac sound diagnostic system based on depth confidence network and diagnostic method |
| CN105825230A (en) * | 2016-03-15 | 2016-08-03 | 中国农业大学 | Soil moisture status prediction method and system based on DBN (Deep Belief Network) model |
| CN106202946A (en) * | 2016-07-18 | 2016-12-07 | 燕山大学 | Clinker free calcium levels Forecasting Methodology based on degree of depth belief network model |
| CN106326634A (en) * | 2016-08-09 | 2017-01-11 | 浙江工业大学 | Hospital outpatient quantity predicting method based on deep belief network |
| CN106407690A (en) * | 2016-09-28 | 2017-02-15 | 厦门理工学院 | Outpatient number prediction method and system based on automatic deep belief network |
| CN106802977A (en) * | 2016-12-14 | 2017-06-06 | 同济大学 | One kind is used for sintering performance index prediction and Quality evaluation method |
| CN106769048A (en) * | 2017-01-17 | 2017-05-31 | 苏州大学 | Self adaptation depth confidence network Method for Bearing Fault Diagnosis based on Nesterov momentum methods |
Non-Patent Citations (1)
| Title |
|---|
| 刘月骁: "燃气调压器故障智能诊断技术的应用研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109596649A (en) * | 2018-11-29 | 2019-04-09 | 昆明理工大学 | A kind of method and device that host element concentration is influenced based on convolutional network coupling microalloy element |
| CN109784276A (en) * | 2019-01-16 | 2019-05-21 | 东南大学 | A kind of voltage dip feature extraction based on DBN and temporarily drop source discrimination method |
| CN110070217A (en) * | 2019-04-11 | 2019-07-30 | 武汉科技大学 | A kind of Forcasting Sinter Quality method of Kernel-based methods parameter |
| CN110218862A (en) * | 2019-05-09 | 2019-09-10 | 山东钢铁股份有限公司 | A kind of staflux is with addition of adjusting method |
| CN110533082A (en) * | 2019-08-09 | 2019-12-03 | 东北大学秦皇岛分校 | A kind of sintering, mixing and water adding control method based on dual model collaborative forecasting |
| CN110533082B (en) * | 2019-08-09 | 2022-02-01 | 东北大学秦皇岛分校 | Sintering mixed water adding control method based on dual-model collaborative prediction |
| US11221610B1 (en) | 2019-09-30 | 2022-01-11 | General Electric Company | Optimized support design for sintering parts with complex features |
| CN112070408A (en) * | 2020-09-14 | 2020-12-11 | 唐山学院 | Agglomerate composition forecasting model based on big data and deep learning |
| CN114697012A (en) * | 2020-12-30 | 2022-07-01 | 科大国盾量子技术股份有限公司 | Real-time wide area user node quantum key secret communication method |
| CN113517037A (en) * | 2021-03-25 | 2021-10-19 | 中南大学 | Method and system for predicting sintering ore FeO by fusing data and knowledge |
| CN113517037B (en) * | 2021-03-25 | 2022-11-15 | 中南大学 | Method and system for predicting sintering ore FeO by fusing data and knowledge |
| CN113095561A (en) * | 2021-04-06 | 2021-07-09 | 上海交通大学 | Method for predicting FeO content of sinter by multi-source information and long-and-short time memory network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108388762A (en) | Sinter chemical composition prediction technique based on depth confidence network | |
| CN109935280B (en) | Blast furnace molten iron quality prediction system and method based on ensemble learning | |
| CN105886680B (en) | A kind of blast furnace ironmaking process molten iron silicon content dynamic soft measuring system and method | |
| CN104809311B (en) | A kind of structural member method for predicting residual useful life based on multifactor fusion amendment | |
| CN106802977A (en) | One kind is used for sintering performance index prediction and Quality evaluation method | |
| CN104899425A (en) | Variable selection and forecast method of silicon content in molten iron of blast furnace | |
| CN111915080B (en) | Raw fuel cost optimal proportioning method based on molten iron quality constraint | |
| CN114969953B (en) | Optimized shield underpass tunnel design method and equipment based on Catboost-NSGA-III | |
| CN110070217A (en) | A kind of Forcasting Sinter Quality method of Kernel-based methods parameter | |
| CN108846526A (en) | A kind of CO2 emissions prediction technique | |
| CN104915518A (en) | Establishing method and application of two-dimensional prediction model of silicon content in hot metal in blast furnace | |
| CN107169205A (en) | A kind of classification model construction method of iron ore | |
| CN102601881B (en) | Method for monitoring on-line quality and updating prediction model of rubber hardness | |
| CN105740984A (en) | Product concept performance evaluation method based on performance prediction | |
| Li et al. | Dynamic time features expanding and extracting method for prediction model of sintering process quality index | |
| CN102842066B (en) | A kind of modeling method of biomass stove burning optimization | |
| CN106650102A (en) | Method for confirming parameters of prediction model for endurance quality of ocean concrete based on grey correlation | |
| CN107292029A (en) | A kind of determination method that sheet forming technological parameter is predicted based on forming defects | |
| Hu et al. | Multi-model ensemble prediction model for carbon efficiency with application to iron ore sintering process | |
| CN110472349A (en) | A kind of hot-rolled steel performance prediction method based on EEMD and depth convolutional network | |
| CN103778466B (en) | Converter steel-making temperature modeling forecasting method based on vector error and system thereof | |
| CN112069656A (en) | Durable concrete mix proportion multi-objective optimization method based on LSSVM-NSGAII | |
| CN102621953B (en) | Automatic online quality monitoring and prediction model updating method for rubber hardness | |
| CN110489844B (en) | Prediction method suitable for uneven large deformation grade of soft rock tunnel | |
| CN105574264A (en) | SVR soft measuring method for kiln tail decomposition rate of cement decomposing furnace |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180810 |
|
| RJ01 | Rejection of invention patent application after publication |