CN101339177A - Sintering mine FeO prediction system based on nerval net - Google Patents

Abstract

The invention relates to the technical field of intelligent forecast, in particular to a forecasting system of sintering ore FeO based on the neural network; the invention provides a forecasting system of sintering ore FeO based on the neural network, and the system can realize intelligent forecasting; the technical proposal thereof has the key points that the system comprises an automatic data collecting module, a training module, a forecasting module and a common module; the automatic data collecting module mainly collects and processes the parameters of sintering and batching automatically and transmits the data to the training module and the forecasting module; the training module trains the data sent by the automatic data collecting module and obtains weight and threshold; the forecasting module uses the weight and the threshold obtained by the training module to carry out forecasting on the data sent by the automatic data collecting module; the common module is used for declaration and storage of global variable and common variable. The invention can realize intelligent forecasting with high accuracy, and is particularly applicable to the forecasting control of sintering ore FeO of steel smelting.

Description

Sintering mine FeO prediction system based on neural network

Technical field

The present invention relates to the Technology of Intelligence Prediction field, concrete relate to a kind of sintering mine FeO prediction system based on neural network.

Background technology

FeO (iron protoxide) is the important component of sintering deposit, and its principal feature is the formation mechanism complexity, and it is big to fluctuate in sintering process, and influence factor is many, skewness in sintering deposit.Sintering process is the oxidation-reduction process of a complexity, even it is also different to join the FeO of the identical generation of charcoal amount.Therefore FeO is that stability is minimum in the sintering deposit all the components.Yet FeO again with the thing phase composition of sintering deposit, main performances such as mineral structure (macrostructure and micromechanism), intensity, size composition, reductibility, melting are closely related, and then influence blast furnace process.FeO also is the major influence factors of quality index such as sintering finished rate, throughput rate, sometimes even play a decisive role, and the fluctuation of sintering production process and deterioration, as the balance of returning mine is destroyed, generally all with FeO control imbalance very big relation is arranged.As can represent the relative furnace temperature of blast furnace with the silicone content in the pig iron, FeO can represent sintering temperature level relatively, is the concentrated expression of sintering atmosphere and heat levels.

In view of FeO control difficulty is big,, is difficult to its content of calculating and reaches forecast and the purpose of controlling with conventional method to sintering influence complexity.Field control at present partly is the empirical method that adopts artificial judgment, be exactly detect by an unaided eye to the viewport of sintering machine afterbody the at set intervals brightness and the thickness of the flourishing layer of sintering deposit of operator, what are estimated the FeO of sintering deposit the chances are, contrast with laboratory values again, just draw corresponding FeO value under certain flourishing layer thickness and brightness conditions, then fuel metering proportioning and operating parameter again through long-term accumulation, just are formed with the experience of usefulness.This determination methods can be received certain effect for veteran operative employee, but limitation also is conspicuous, at first is that operative employee's quality is different, and technical merit has height, and sense of responsibility also has difference; Next be raw material and fuel quality at any time all in fluctuation, the corresponding relation of flourishing layer thickness, brightness and FeO and imprecision mainly are that to influence the factor of FeO too many, this corresponding relation is an aspect.Therefore, adopt this mode to check the shortcoming that check lags behind, hit rate is low.

In addition, it is to adopt to detect automatically to analyze FeO content that part is at home and abroad arranged at present, and it mainly comprises two kinds of methods:

1, based on the sintering mine FeO content online test method of computer picture, the main thought of this method is with the CCD gamma camera tail cross section of sintering machine image to be carried out Computer Analysis, provide the FeO value, its shortcoming is that video camera is influenced by the tail high-temperature dust, image is easy to generate noise, and analysis result is very unstable.

2, magnetic permeability method.This method is that the magnetic conduction instrument is installed above sintering machine afterbody chassis, according to the sintering mine FeO difference different principle of magnetic then, can draw corresponding FeO by the magnetic permeability that detects sintering deposit, shortcoming is the FeO that only can detect sintering pallet top layer sintering deposit, can not represent whole sintering mine FeOs, and in fact the FeO content of chassis upper, middle and lower layer sintering deposit there is very big difference.

Summary of the invention

Technical matters to be solved by this invention is: at the deficiency of detection method in the prior art, propose a kind of sintering mine FeO prediction system based on neural network that can realize intelligent forecasting.

The technical scheme that the present invention solves the problems of the technologies described above employing is: based on the sintering mine FeO prediction system of neural network, it comprises automatic data collection module, training module, forecast module and public module; The main acquisition process sintering automatically of described automatic data collection module sends training module and forecast module to the batching parameter and with data; Described training module carries out training managing to the data that automatic acquisition module sends over, and obtains weights and threshold value; The utilization of described forecast module is forecast the data that the automatic acquisition module of data sends over by weights and threshold value that training module obtains; Described public module is used for the statement of global variable and public variable and deposits.

Described neural network is four layers of feedforward neural network.

Described four layers of feedforward neural network have two hidden layers and an input layer, an output layer, and its structure is

The stack of the nonlinear relationship that is input as 10 variablees of described output layer and the linear relationship of 1 variable.

The algorithm that described neural network adopts change study step-length and variation coefficient to combine is realized the neural network convergence.

The invention has the beneficial effects as follows: neural network has the self-study habit, self-organization, and adaptivity and robustness, extensive performance is strong, and the forecast precision height can be used for sintering process control and index forecast.Many for this influence factor of SINTERING PRODUCTION, the process mechanism complexity, the system that hysteresis quality is big adopts nerual network technique to set up forecast system, takes the adjustment measure in advance, reaches the purpose of optimal control, has important practical value.

Description of drawings

Fig. 1 is a system of the present invention main composition block diagram;

Fig. 2 is the structural representation of the neural network among the present invention;

Fig. 3 is neural metwork training process flow diagram flow chart among the present invention;

Fig. 4 is neural network forecasting process process flow diagram among the present invention.

Embodiment

The invention will be further described below in conjunction with drawings and Examples.

The present invention proposes a kind of sintering mine FeO prediction system, referring to Fig. 1, this system comprises automatic data collection module, training module, forecast module and public module; The main acquisition process sintering automatically of described automatic data collection module sends training module and forecast module to the batching parameter and with data; Described training module carries out training managing to the data that automatic acquisition module sends over, and obtains weights and threshold value; The utilization of described forecast module is forecast FeO by weights and threshold value that training module obtains to the data that the automatic acquisition module of data sends over; Described public module is used for the statement of global variable and public variable and deposits.

Neural network belongs to prior art, and according to neural network theory, the node number of input layer and output layer is that the number of the influence factor (parameter) of the target problem that solved by expectation decides.Each field facts have proved both at home and abroad, have the many neural networks of plural hidden layer and implicit unit number and have more strong functions than the neural network of having only a hidden layer, implicit unit number to lack, can handle large complicated problem, parallel processing capability is strong, neural network precision height, generalization ability is strong, strong robustness.Even small part neuronal necrosis wherein, neural network still can operate as normal.Therefore neural network structure of the present invention is designed to have four layers of feedforward neural network of two hidden layers.

The present invention chooses with the FeO relation and directly imports as neural network with the major parameter that is easy to detect, and comprises 11 variablees: the bed of material, machine speed, ignition temperature, person in charge's negative pressure, exhaust gas temperature, 2# bellows negative pressure, 2# box temperature, 16# box temperature, 21# box temperature, mixture moisture, fuel ratio; Choose sintering mine FeO content as output.According to neural network is tested, if with above-mentioned 11 variablees all from the 1st layer of input, the excitation function of hidden layer and output layer all uses nonlinear S function, the change fuel ratio, fix other variable, then neural network output variation is very little, does not reflect the true rule of fuel ratio to the FeO influence.And sintering theory shows with production practices both at home and abroad, fuel ratio has most important influence to the FeO of sintering deposit, the relation that almost is in line and rises, the every rising 1% of FeO, then fuel increases about 3.5kg/t, it is about 0.25% to be converted to proportioning, the every rising 1% of fuel ratio in other words, and FeO just should rise 4%.And the neural network test findings that we did to be fuel ratio risen 1.71%, FeO should rise 6.84% at least, and the output of actual neural network has only risen 1.14%, greatly differs from each other.If use the neural network of this structure to instruct production, only can cause to have a strong impact on to sinter quality, therefore must improve neural network structure.We directly are input to the 4th layer from extra play (with the 3rd layer of position equivalence) with fuel ratio as independent variable, with y _r=w _rX ₁₁Linear relationship act on output layer, fuel ratio is independent of other variable, its error back propagation directly arrives till the node of itself, not anti-pass is to the 2nd layer, output layer is exactly the stack of the linear relationship of the nonlinear relationship of 10 variablees and 1 variable like this, and this is the significant improvement to general neural network design.Improving its structure of back is

Wherein the bed of material, machine speed, ignition temperature, person in charge's negative pressure, exhaust gas temperature, 2# bellows negative pressure, 2# box temperature, 16# box temperature, 21# box temperature, these 10 variablees (node) of mixture moisture of input layer are represented in first " 10 "; " 40 " represent the node number of first hidden layer; The node number of second hidden layer of second " 10 " expression; This variable of fuel ratio (node) of " 1 " expression extra play input above second " 10 "; This variable of sintering mine FeO (node) of last " 1 " expression output layer.Its structural representation is referring to Fig. 2.

In the present invention, also the algorithm of neural network has been done improvement, the algorithm that has adopted change study step-length η and variation factor alpha to combine is realized network convergence, behind the every training bout of sample set, check E (n) (E (n) is a sample set error energy function summation) is as sample variance E _Av=just reduce step-length and momentum factor when (1/P) E (n) is near 2 ε, to guarantee network stabilization.Wherein: P is a total sample number, and ε is for being used for evaluating network convergent set-point, and setting the ε initialization value among the present invention is 0.02.The value that native system is set variance is carried out loop iteration before greater than 0.02 always and is calculated, up to less than 0.02.Adopt such algorithm not only to accelerate the speed of convergence of network, guaranteed network stabilization, also can overcome as adopting the BP algorithm and be absorbed in the defective of the local minimum point of curved surface easily.The specific algorithm design is as follows:

1, network output is calculated

It is the S function that the non-linear partial of network hidden layer and output layer adopts excitation function:

$f\left(\mathrm{net}\right)=\frac{1}{1+\exp (-\mathrm{net})}$

Extra play is output as linear function: y _r=w _rX ₁₁

(1) the 2nd layer of output:

$o_j^2=\frac{1}{1+\exp (-(\mathrm{\Σ}{x}_i{w}_{\mathrm{ij}}-{\mathrm{\θ}}_j))}$

x _iIt is the 2nd layer input; w _IjFor connect the 1st layer with the 2nd layer weights; θ _jIt is the 2nd layer threshold value; The span of i is 1 to 10, and the span of j is 1 to 40.

(2) the 3rd layers of output:

$o_k^3=\frac{1}{1+\exp (-(\mathrm{\Σ}{o}_j^2{v}_{\mathrm{jk}}-{\mathrm{\θ}}_k))}$

o _j ²It is the 3rd layer input; v _JkFor connect the 2nd layer with the 3rd layer weights; θ _kIt is the 3rd layer threshold value; The span of j is 1 to 40, and the span of k is 1 to 10.

(3) the 4th layers of output:

$o_m^4=w_r\·x_{11}+\frac{1}{1+\exp (-(\mathrm{\Σ}{o}_k^3{u}_{\mathrm{km}}-{\mathrm{\θ}}_m))}$

o _k ³It is the 4th layer input; u _KmFor connect the 3rd layer with the 4th layer weights; w _rWeights for fuel ratio; x ₁₁Be fuel ratio; θ _mIt is the 4th layer threshold value; The span of k is 1 to 10, m=1.

2, error back propagation and right value update

If adopt the BP algorithm, the error of the 4th layer of output and desired output (sample FeO) is along the network backpropagation, the modification refreshing weight.The evaluating network condition of convergence be all neuronic variances of output layer less than specified value ε, promptly think network convergence.

If the actual value of p training sample (expectation value) is O _p, the last one deck of network is output as o _m ⁴, output error: $e_p=o_p-o_m^4$

Owing to have only an output, then sample set error energy function summation is:

$E\left(n\right)=\frac{1}{2}\underset{p-1}{\overset{p}{\mathrm{\Σ}}}{e}_p^2\left(n\right),$ P is a total sample number, and n is an iterations, and obviously E (n) is a higher-dimension curved surface.

The deficiency of BP algorithm is that network convergence speed is slow, and there is local minimum in objective function.Adopt variable step η study for this reason and increase momentum term α,, can avoid learning to be absorbed in the local minimum point of curved surface again to accelerate speed of convergence and to guarantee network stabilization.Have for the n time iteration:

(1) backpropagation of calculating partial gradient

Any node for network has:

The 4th layer of output gradient: ${\mathrm{\δ}}_m^4\left(n\right)=o_m^4\left(n\right)(1-o_m^4\left(n\right))(o_p\left(n\right)-o_m^4\left(n\right)),$ m＝1；

The 3rd layer of output gradient: ${\mathrm{\δ}}_k^3\left(n\right)=o_k^3\left(n\right)(1-o_k^3\left(n\right))\underset{m=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\δ}}_m^4\left(n\right)u_{\mathrm{km}}\left(n\right),$ k＝1?to?10；

The 2nd layer of output gradient: ${\mathrm{\δ}}_j^2\left(n\right)=o_j^2\left(n\right)(1-o_j^2\left(n\right))\underset{k=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\δ}}_k^3\left(n\right)v_{\mathrm{jk}}\left(n\right),$ j＝1?to?40。

(2) calculate the weights increment

$\mathrm{\Δ}{u}_{\mathrm{km}}\left(n\right)=\mathrm{\η}{\mathrm{\δ}}_m^4\left(n\right)o_k^3\left(n\right);$ $\mathrm{\Δ}{v}_{\mathrm{jk}}\left(n\right)=\mathrm{\η}{\mathrm{\δ}}_k^3\left(n\right)o_j^2\left(n\right);$

$\mathrm{\Δ}{w}_{\mathrm{ij}}\left(n\right)=\mathrm{\η}{\mathrm{\δ}}_j^2\left(n\right)x_i\left(n\right);$ $\mathrm{\Δ}{w}_r\left(n\right)=\mathrm{\η}{\mathrm{\δ}}_m^4\left(n\right)x_{11}\left(n\right).$

$\mathrm{\Δ}{w}_r\left(n\right)=\mathrm{\η}{\mathrm{\δ}}_m^4\left(n\right)x_{11}\left(n\right)$ Be interpreted as: the weights increment of extra play equals to learn the input quantity that step-length multiply by the 4th layer of output gradient and extra play.

(3) right value update, and add the additional momentum item

u _km(n+1)＝u _km(n)+α·Δu _km(n-1)+Δu _km(n)；

v _jk(n+1)＝v _jk(n)+α·Δv _jk(n-1)+Δv _jk(n)；

w _ij(n+1)＝w _ij(n)+α·Δw _ij(n-1)+Δw _ij(n)；

w _f(n+1)＝w _f(n)+α·Δw _f(n-1)+Δw _f(n)。

η is the study step-length, and α is a momentum factor.

u _Km(n+1)=u _Km(n)+α Δ u _Km(n-1)+Δ u _Km(n) be interpreted as: next step weights of the 3rd node layer equal the weights increment that this step weights add this step, add the product of a momentum factor and previous step weights increment.

The training process flow process that the data that the automatic acquisition module of data is sent are carried out is referring to Fig. 3: at first use random function initialization weights and threshold value, and the object data of training is submitted to network; Carry out the computational grid second layer, the 3rd layer, the 4th layer input and output more successively; Follow the normalization errors of each layer of backwards calculation, and successively revise refreshing weight and threshold value; Then become study step-length and momentum factor, until the sample variance E of training _AvLess than setting value ε, training finishes.

Embodiment:

The sintering mine FeO prediction system based on neural network in this example adopts Visual Basic visual programming, human-computer interaction interface close friend.Land to data output totally 12 modules from system.4 important module are wherein arranged: automatic data collection module, training module, forecast module, public module.VB+SQL Sever2000 realization is adopted in system development, and data management adopts open data source ODBC configuration to call SQL Sever, and data connect employing ADODC, show to adopt Datagride control, program to realize adopting the VB code programming.It is big that the SQLSever database has memory space, and travelling speed is fast, characteristics easy to maintenance.Data in the historical data base of the IFIX 3.5 of the existing configuration of automatic data collection module collection, 5 seconds of frequency acquisition once.Operating system adopts WindowsXP, and forecast system is in the exploitation down of programming software Visual Basic6.0 the integration environment.Forecast system can move under the version more than the Windows98.

After system development is finished, on a certain sintering machine, implement.The ruuning situation of now randomly drawing a period of time is come the application of illustrative system.Training data be sometime the section sintering process image data, after the processing 90 groups of sample datas, submit to neural network and train, through 540 the step (6 bouts of sample set) iteration, network convergence.811 weights of neural network parameter, 51 threshold values are fixed, and are stored in the SQLSever database.This moment, the neural metwork training result was: sample set error energy function summation E (n)=1.35796160, variance E _Av(n)=and 0.01508846＜ε, its average error, maximum error, the absolute value of least error is respectively e _Av=0.135065100, e _Max=0.435301304, e _Min=0.000474930.

The weights and the threshold value that adopt training to obtain still adopt 90 groups of data of training sample to check neural network, the average error that obtains checking, and maximum error, the absolute value of least error is respectively: e _Av=0.135665, e _Max=0.657139, e _Min=0.001095.The permission fluctuation range of on-the-spot FeO coefficient of stabilization is ± 0.5%, hit even if now stipulate absolute error＜0.4% of training sample check, in 90 groups of samples that we detect, have only 2 groups miss, hit rate reaches 97.78%.

Calculate through check, the mean value that obtains training sample set check neural network output FeO is 7.53%, wherein the fuel ratio linear segment is output as 4.14%, the non-linear partial of other 10 input variables is output as 3.39%, the proportion of linear segment accounts for 54.98% as a result, the proportion 45.02% of non-linear partial, and this proves absolutely that the effect of each input variable is different in neural network, fuel ratio occupies absolute consequence, has also confirmed the correctness of neural network structure design.

Adopt training sample set sample data in addition to fan-in network, the extensive performance of supervising network, data are taken from 58 groups of data of other time period of sintering machine on the same stage, these data are that automatic acquisition module is passed to the forecast module, weights and the threshold value of utilizing training to obtain, the output of computational grid output layer, the result is prediction error e _Av=0.189226, e _Max=0.620053, e _Min=0.000084, hit rate is 91.14%.The forecast result satisfies production requirement fully.The flow process of this forecasting process is referring to Fig. 4: at first call in the weights and the threshold value that have trained, and go out to obtain input parameter from the automatic data collection module; The computational grid second layer, the 3rd layer, the 4th layer input and output are successively forecast subsequently.

The new parameter of input in forecasting process, deposit database in after, need not remove old data, make it constantly increase new breath, according to neural network theory, training sample is many more, and the network function that training is come out is powerful more, running into any sample can both handle, and generalization ability is better, but the training time increases.According to the neural metwork training rule, the training error and the verify error minimum of several groups of at last up-to-date samples that is to say, no matter sample set has much, forecast precision and nearest sample relation greatly, therefore continuous increase information will add network with FeO actual value new samples.Network will be trained certain excessively period, after sample increases the 50-100 bar, need training again, and one side forecasts that simultaneously with new data training network refreshing weight, they carry out, and are independent of each other in different modules.Adopt system of the present invention can realize intelligent forecasting, abandoned the disturbing factor of artificial judgement in the past, the accuracy height is specially adapted to the sintering mine FeO prediction control to smelting iron and steel.

Claims (5)

1. based on the sintering mine FeO prediction system of neural network, it is characterized in that: comprise automatic data collection module, training module, forecast module and public module; The main acquisition process sintering automatically of described automatic data collection module sends training module and forecast module to the batching parameter and with data; Described training module carries out training managing to the data that automatic acquisition module sends over, and obtains weights and threshold value; The utilization of described forecast module is forecast the data that the automatic acquisition module of data sends over by weights and threshold value that training module obtains; Described public module is used for the statement of global variable and public variable and deposits.

2. the sintering mine FeO prediction system based on neural network as claimed in claim 1 is characterized in that: described neural network is four layers of feedforward neural network.

3. the sintering mine FeO prediction system based on neural network as claimed in claim 2 is characterized in that: described four layers of feedforward neural network have two hidden layers and an input layer, an output layer, and its structure is

1

×1

10×40×10 。

4. the sintering mine FeO prediction system based on neural network as claimed in claim 3 is characterized in that: the stack of the nonlinear relationship that is input as 10 variablees of described output layer and the linear relationship of 1 variable.

5. as claim 1,2,3 or 4 described sintering mine FeO prediction systems based on neural network, it is characterized in that: the algorithm that described neural network adopts change study step-length and variation coefficient to combine is realized network convergence.

Images