CN101477374B - Continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network - Google Patents

Abstract

The invention relates to a combined diagnosis forecasting method for continuous casting steel leakage time-sequence space based on a fuzzy neural network, and belongs to the field of production and control of metallurgical continuous casting process. The method comprises the following steps: (1) acquiring data, namely, acquiring temperature data of continuous casting field thermocouples; (2) establishing a fuzzy ART neural network; (3) judging a single-row thermocouple sequential network, and inputting data of an upper row of thermocouples and a lower row of thermocouples; if the approaching degree of a single row of thermocouples is more than 0.95, warning; and if the approaching degree of the single row of thermocouples is between 0.85 and 0.95, entering step (4); and (4) judging a spatial network of a combined coupler; calculating the approaching degree of the thermocouples adjacent to the single row of thermocouples, if the approaching degree is between 0.85 and 0.95, calculating the threshold value P of the grouped coupler, and warning when the threshold value P is larger than 0.8. Through combined judgment of the single coupler and the combined coupler, the method improves identifying effect of a representative temperature mode and forecasting accuracy in a continuous casting adhesive steel leakage process through a network model.

Description

Continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network

Technical field

The invention belongs to the production and the control field of metallurgical continuous casting process, particularly a kind of continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network.

Background technology

In continuous casting production process, bleed-out is the industrial accident of tool harmfulness, in case take place to cause stopping production and having to change the equipment that is destroyed by the bleed-out accident, all there is bad influence in life-span to quality, personal safety and the equipment of the stability of operation, product, and then influencing the large-scale application of continuous casting steel machine, the loss that once typical bleed-out accident is caused may be near 200,000 dollars.Because sticker type breakout is the main cause that causes bleed-out, reduces sticker type breakout and become the key that reduces the continuous casting bleed-out rate.

In order to overcome the sticker type breakout accident; people have carried out a large amount of experimental studies; wherein except studying high performance covering slag and making various process conditions and parameter keeps the best of breed; utilize methods such as crystallizer heat interchange analysis, thermocouple temperature measurement, the wave analysis of shaking, friction measurement effectively continuous casting bleed-out to be forecast; wherein since thermocouple thermometry have response fast, economical and practical, be easy to characteristics such as maintenance, reliability height, thereby become domestic and international application bonding steel leakage forecasting procedure the most widely.

At present domestic and international nearly all breakout prediction system mainly on the basis of thermocouple thermometry, forecasts the continuous casting bonding steel leakage by logic determines and neural network.Wherein the neural network method has better self-adapting ability, robustness and fault-tolerant ability with respect to the logic determines method, and the growth in time of its performance improves constantly, and therefore becomes the development trend of bleed-out diagnosis forecast in recent years.The existing neural network breakout prediction method of utilizing, as (Zhao Qi, Zhu Miaoyong, based on the continuous casting bleed-out forecasting model [J] that improves fuzzy ART neural network, China is metallurgical, 2007,17 (10): 26-29,53), just predicted the continuous casting bleed-out forecast of single-row thermopair, this forecasting procedure has solved the warning to the bleed-out accident to a certain extent, but has more wrong report situation.

Summary of the invention

Low at existing breakout prediction precision, wrong report is frequent, the situation that the continuous casting breakout prediction method precision of single thermopair is not high and existence is reported by mistake.The invention provides a kind of continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network, by improvement to fuzzy ART neural network learning algorithm and network structure, and combination Fuzzy Pattern Recognition and fuzzy clustering algorithm, the novel state of conflict fuzzy neural network that this method will be set up is used for the forecasting process of bonding steel leakage, guarantee lower rate of failing to report and rate of false alarm, can discern in the bonding steel leakage process generation that accident is drawn in two kinds of typical temperature models and forecast effectively.

Technical scheme of the present invention is as follows: forecasting bonding steel leakage in the continuous casting production process, its essence identifies two kinds of typical temperature models (as shown in Figure 1) in the bonding steel leakage process exactly, in fact be exactly the problem of a dynamic waveform pattern-recognition, promptly to detect a large amount of temperature waveforms that obtain from thermopair, identify the waveform with bleed-out sign, to two kinds of typical temperature waveforms among the figure: temperature model 1 and temperature model 2 carry out Dynamic Recognition.

Continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network of the present invention comprises the steps:

(1) image data

Gather the temperature data of the on-the-spot thermopair of continuous casting, and temperature data is carried out pre-service.Pretreated process comprises bad data processing and data normalization, and wherein the bad data processing is with sample temperature t _iWith the previous moment temperature t _I-1Compare,, then the previous moment temperature is substituted the temperature of current time, i.e. t if both differ 30% _i=t _I-1Data normalization refers to: the data of gathering are carried out normalized, and formula is as follows:

${\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="H2009100101637E00031.png,H2009100101637E00032.png"\; state="\{\{state\}\}">t</figure-callout>}}_1^{*}=\left\{\begin{array}{c}\frac{t_i}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i^2}}-(\max T-\min T)<\mathrm{\&lambda;}\\ \frac{t_i-\min T}{\max T-\min T}-(\max T-\min T)>\mathrm{\&lambda;}\end{array}\right.$

In the formula: λ is the stable threshold of temperature survey sequence, generally gets 20 ° to 24 °.

(2) set up fuzzy ART neural network, and train.

Network model is made up of input layer F0, identification layer F1 and output layer F2, and the input X of input layer F0 is the temperature that step (1) is gathered, X=T={t ₁, t ₂..., t _n, t _i∈ [0,1]; Identification layer F1 activation value vector $\stackrel{\&RightArrow;}{S}=\left\{S_j\right\}$ Be current input

And the approach degree between the corresponding cluster centre of each node of F1 layer, approach degree adopts following formula to calculate:

$N(\tilde{A},\tilde{B})=\frac{2\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}\left(x_i\right)u_{\tilde{B}}\left(x_i\right)}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}^2\left(x_i\right)+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{B}}^2\left(x_i\right)}---\left(1\right)$

According to the competition details of the win, the activation value maximum is that the node J of corresponding approach degree maximum wins: S _J=max{S _j, j=1,2 ..., C}, the maximum approach value that it is corresponding $\mathrm{\&eta;}=S_J=N({\stackrel{\&RightArrow;}{X}}_k,{\stackrel{\&RightArrow;}{v}}_J).$

Be approach degree between vectorial A and B, subscript " ~ " represents that it is a vector.

Be certain element x among the vectorial A _iDegree of membership,

Be certain element x among the vectorial B _iDegree of membership, the computing formula that is subordinate to matrix and cluster centre adopts formula commonly used to calculate.

If X is the approach degree output of present networks Model Identification layer, V is the output layer weights, and Y is a desired output, E is the error of actual output and desired output, and then Y=VX+E utilizes quadrature least square method OLS that output layer weights V is carried out match, make network output reach expectation value, adopt C _ONDITIONALThe FCM algorithm is further adjusted cluster centre.

Simultaneously according to the following formula refreshing weight:

${\stackrel{\&RightArrow;}{W}}_J^{\mathrm{new}}={\stackrel{\&RightArrow;}{W}}_J^{\mathrm{old}}+\mathrm{\&beta;}\&times;({\stackrel{\&RightArrow;}{X}}_k-{\stackrel{\&RightArrow;}{W}}_J^{\mathrm{old}})$

In the formula

Be the network weight of newly determining,

(former) network weight for before, β is a learning coefficient, between the value [0,1].When the J node when being entrusted node, β=1.0, network is in quick learning state, network is set up a new classification pattern simultaneously; On the contrary, 0＜β＜1, network is in learning state at a slow speed, and network is attached to the new model information of current input sample in the corresponding cluster centre simultaneously;

Among the output layer F2: y _kBe sample X _kReal network output, f _kBe sample X _kThe output value of feedback, its size expression sample X _kTo the influence degree of cluster centre in the clustering algorithm, f _k=0 this sample of expression is to cluster centre not influence fully; f _k=1 this sample of expression has the greatest impact to cluster centre.

Set up good fuzzy ART neural network,, train the data input neural network of two kinds of typical temperature models (as shown in Figure 1) in the bonding steel leakage process.

(3) single-row thermopair sequential network is differentiated

For the fuzzy ART neural network that step (2) is set up, the data of heat extraction galvanic couple and following heat extraction galvanic couple in the input,

If single-row thermopair approach degree η 〉=ρ, ρ gets 0.95 warning;

If single-row thermopair approach degree η enters step (4) 0.85～0.95;

If single-row thermopair approach degree η is less than 0.85 then return step (1);

(4) the even spatial network of group is differentiated

Calculate this single-row thermopair adjacent column thermopair approach degree, if approach degree η, calculates the threshold value P=P of group idol this moment also in 0.85～0.95 scope _D+ P _D(1-P _Z), satisfy threshold value P greater than 0.8, report to the police;

Otherwise return step (1).

In the formula: P _DThe approach degree of single-row thermopair, P _ZThe approach degree of adjacent column thermopair.

Beneficial effect of the present invention: judge by combination to single even summation group idol, precision improves many, by fuzzy clustering algorithm and fuzzy ART neural network are organically combined, introduce Conditional FCM algorithm cluster centre is done further to adjust, thereby further improved network model the identification effect of representative temperature pattern and the forecast precision of model in the continuous casting bonding steel leakage process.

Description of drawings:

Fig. 1 is the representative temperature pattern diagram of continuous casting bonding steel leakage process of the present invention;

Fig. 2 is a spacing combined diagnosis synoptic diagram of the present invention;

Fig. 3 is the process flow diagram of bonding steel leakage diagnosis forecast among the present invention;

Fig. 4 is the structural drawing of fuzzy neural network among the present invention;

Fig. 5 differentiates synoptic diagram for single-row thermopair sequential network among the present invention;

Fig. 6 is network learning procedure process flow diagram among the present invention;

Fig. 7 differentiates synoptic diagram for the even spatial network of group among the present invention;

Fig. 8 is network time sequence spacing combined diagnosis process flow diagram among the present invention.

Embodiment:

Specific implementation process of the present invention is divided into two steps:

Adopt the bleed-out diagnosis forecast model of this method foundation (just to do individual statistics roughly here to training period with during using through dropping into nearly 1 year effect of on-line operation such as following table after the correction of nearly 5 months model training and model parameter, because training period and to use phase length different, so represent and represent) without number of times with ratio.

Table 1 model effect

(1) learning training of network

Utilize the historical data of continuous casting collection in worksite, therefrom choose have representative temperature pattern 1 and pattern 2 (as shown in Figure 1) feature and other pattern features in the continuous casting bonding steel leakage process the temperature survey sequence as training sample set, be used for the training of neural network model I and model II, wherein model I is used to discern the sample with temperature model 1 feature, and model II is used to discern the sample with temperature model 2 features.The training sample of two models, network structure and learning algorithm etc. are all identical, the desired output difference that is, and its desired output is respectively:

$y_k^{*}\left(I\right)=\left\{\begin{array}{cc}0.9& 1\&le;k\&le;23\\ 0.1& 24\&le;k\&le;34\end{array}\right.$ $y_k^{*}\left(\mathrm{II}\right)=\left\{\begin{array}{cc}0.1& 1\&le;k\&le;23\\ 0.9& 24\&le;k\&le;34\end{array}\right.$

The training process of network is divided into two stages to carry out, and determines cluster centre by phase one study is preliminary; Cluster centre is further adjusted and optimized by subordinate phase study, and definite output layer weights, idiographic flow is as shown in Figure 8.

(2) identification of network

After the learning training process of network model finished, the network that utilization trains carried out identification to the dynamic waveform of representative temperature pattern, and its detailed process is as follows:

A) carry out the data pre-service by the thermo-electric couple temperature data to the continuous casting field real-time acquisition, the input temp that obtains network is measured sequence

B) utilize the approach degree of the cluster centre computational grid identification layer that trains network to export:

$N(\tilde{A},\tilde{B})=\frac{2\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}\left(x_i\right)u_{\tilde{B}}\left(x_i\right)}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}^2\left(x_i\right)+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{B}}^2\left(x_i\right)}$

C) according to output layer weights v _jThe actual output of computational grid y:

$y=\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{v}_j{T}_j$

Single-row thermopair sequential network is differentiated: for the fuzzy ART neural network of setting up, the data of heat extraction galvanic couple and following heat extraction galvanic couple in the input, as shown in Figure 2, with all last heat extraction galvanic couples and following heat extraction galvanic couple input fuzzy ART neural network, single-row thermopair refers to heat extraction galvanic couple Ai and corresponding following heat extraction galvanic couple Bi;

If single-row thermopair approach degree η 〉=ρ, ρ gets 0.95 warning;

If entering the even spatial network of group 0.85～0.95, single-row thermopair approach degree η differentiates.

Organizing even spatial network differentiates: calculate this single-row thermopair adjacent column thermopair approach degree, if approach degree η, calculates the threshold value P=P of group idol this moment also in 0.85～0.95 scope _D+ P _D(1-P _Z), satisfy threshold value P greater than 0.8, report to the police;

In the formula: P _DThe approach degree of single-row thermopair, P _ZThe approach degree of adjacent column thermopair.

The bleed-out diagnosis forecast model that adopts this method to set up drops into the nearly 1 year effect of on-line operation after the correction of nearly 5 months model training and model parameter as shown in table 1, and rate of false alarm reduces greatly, and rate of failing to report is essentially 0.

Claims (1)

1. continuous casting bleed-out time sequence spacing combined diagnosis prediction method based on fuzzy neural network is characterized in that may further comprise the steps:

(1) image data; Gather the temperature data of the on-the-spot thermopair of continuous casting, and temperature data is carried out pre-service;

Pretreated process comprises bad data processing and data normalization, and wherein the bad data processing is with sample temperature t _iWith the previous moment temperature t _I-1Compare,, then the previous moment temperature is substituted the temperature of current time, i.e. t if both differ 30% _i=t _I-1Data normalization refers to: the data of gathering are carried out normalized, and formula is as follows:

$t_i^{*}=\left\{\begin{array}{c}\frac{t_i}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i^2}}-(\max T-\min T)<\mathrm{\&lambda;}\\ \frac{t_i-\min T}{\max T-\min T}-(\max T-\min T)>\mathrm{\&lambda;}\end{array}\right.$

In the formula: λ is the stable threshold of temperature survey sequence;

(2) set up fuzzy ART neural network, and train;

The ART neural network is made up of input layer F0, identification layer F1 and output layer F2, and the input X of input layer F0 is the temperature that step (1) is gathered, X=T={t ₁, t ₂..., t _n, t _i∈ [0,1]; N represents the number of times of collecting temperature; Identification layer F1 activation value vector

Be current input

And the approach degree between the corresponding cluster centre of each node of F1 layer, approach degree adopts following formula to calculate:

$N(\tilde{A},\tilde{B})=\frac{2\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}\left(x_i\right)u_{\tilde{B}}\left(x_i\right)}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{A}}^2\left(x_i\right)+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\tilde{B}}^2\left(x_i\right)}---\left(1\right)$

According to the competition details of the win, the activation value maximum is that the node J of corresponding approach degree maximum wins: S _J=max{S _j, j=1,2 ..., C}, C are the cluster centre number; The maximum approach value that it is corresponding

Be approach degree between vectorial A and B, it is vector for subscript～expression;

Be certain element X among the vectorial A _iDegree of membership,

Be certain element X among the vectorial B _iDegree of membership, simultaneously according to the following formula refreshing weight:

${\stackrel{\&RightArrow;}{W}}_J^{\mathrm{new}}={\stackrel{\&RightArrow;}{W}}_J^{\mathrm{old}}+\mathrm{\&beta;}\&times;({\stackrel{\&RightArrow;}{X}}_k-{\stackrel{\&RightArrow;}{W}}_J^{\mathrm{old}})$

In the formula

Be the network weight of newly determining, Be network weight before, β is a learning coefficient, between the value [0,1];

Among the output layer F2: y _kBe sample X _kReal network output, f _kBe sample X _kThe output value of feedback, its size expression sample X _kTo the influence degree of cluster centre in the clustering algorithm, f _k=0 this sample of expression is to cluster centre not influence fully; f _k=1 this sample of expression has the greatest impact to cluster centre;

(3) single-row thermopair sequential network is differentiated

For the fuzzy ART neural network that step (2) is set up, the data of heat extraction galvanic couple and following heat extraction galvanic couple in the input,

If single-row thermopair approach degree η 〉=ρ, ρ gets 0.95 warning;

If single-row thermopair approach degree η enters step (4) 0.85～0.95;

(4) the even spatial network of group is differentiated

Calculate this single-row thermopair adjacent column thermopair approach degree, if approach degree η, calculates the threshold value P=P of group idol this moment also in 0.85～0.95 scope _D+ P _D(1-P _Z), satisfy threshold value P greater than 0.8, report to the police;

In the formula: P _DThe approach degree of single-row thermopair, P _ZThe approach degree of adjacent column thermopair.

Images