CN102831269B - Method for determining technological parameters in flow industrial process - Google Patents

Abstract

The invention discloses a method for determining technological parameters in a flow industrial process. The method includes the steps of firstly, collecting production data, and using the technological parameters as input and product quality and output to build quality models; secondly, using rule extraction to extract corresponding rules between the input and the output from the production data so as to form a rule library; and thirdly, using optimization to find rules corresponding to given quality indexes according to the given quality indexes, initiating the technological parameters in the range of the rules, using the initiated technological parameters as input of the quality models to predict product quality, calculating errors between the predicted quality and the quality indexes, and correcting to find optimal parameters to enable the errors between the predicted quality and the quality indexes to be minimum. By the method, product yield is improved effectively and production cost is lowered.

Description

A kind of defining method of process flow industry process technological parameter

Technical field

The present invention relates to a kind of defining method of the process flow industry process technological parameter based on data-driven.

Background technology

Producing Process of Processing Industry generally has the features such as multivariate, non-linear, strong coupling, product quality changes with working condition change, as being subject to the interference of the multiple uncertain factors such as material composition, operating condition, equipment state, and operational process often has dynamic perfromance, be difficult to carry out accurate description with mechanism model.Meanwhile, because its production process is very complicated, the larger gap of product quality existence compared with same kind of products at abroad that domestic process industry is produced, main manifestations is unstable product quality, the life-span is low, rejection rate is high.

The control of product quality after all or the setting problem of technological parameter, namely according to the quality index backstepping technological parameter of product.

Current, in domestic each large Producing Process of Processing Industry, the adjustment of technological parameter is mainly by artificial experience, the method adopting exhaustive examination to gather carries out setting parameter, when working conditions change is frequent, only relies on artificial experience to be difficult to adjusting process parameter timely and accurately, cause product quality to can not get effective control, waste is serious, and easily occurs waste product.

Adopting a kind of reasonably method carry out optimizing process technological parameter and set, is the key issue place solving product quality conservative control.

Process industry can have accumulated a large amount of production datas in long-term production run, at present, the widespread use in the industrial production of a large amount of Novel meters, networked meters and sensing technology, obtain the real time data in a large amount of actual production process, these data have contained the various rules in production run, data driven technique can be utilized, from a large amount of data, obtain the quantitative relationship between technological parameter and product quality, and then realize the forecast of system, monitoring, the various desired function such as diagnosis and optimization.

Summary of the invention

The technical issues that need to address of the present invention are just the defect overcoming prior art, and provide a kind of defining method of the process flow industry process technological parameter based on data-driven, it can improve production product percent of pass effectively, reduce production cost.

For solving the problem, the present invention adopts following technical scheme:

The invention provides a kind of defining method of the process flow industry process technological parameter based on data-driven, described method comprises the following steps:

1) quality modeling: according to the production data gathered in actual production process, utilize neural net method, sets up with process parameter as input, product quality are for output quality model;

2) rule extraction: according to the production data gathered in actual production process, utilizes the method for rule extraction from the rule of correspondence between extracting data constrained input, formation rule storehouse;

3) process parameter optimizing: utilize optimization method according to given quality index, find the rule that given quality index is corresponding, and in regular scope initialization procedure technological parameter, using the input of initialized parameter as quality model, prediction product quality, calculate the error with quality index, according to error, optimization method is utilized to revise technological parameter, using rule as constraint condition in makeover process, ensure revised parameter still in regular scope, revise by optimizing, find the parameter of one group of optimum, make the error of the predicted value of model and quality index minimum.

The invention has the advantages that: the present invention adopts data driven technique, adopt optimization method, speed of searching optimization is fast, simple to operate; The quality index of product is made to be in optimum position in target range fast.

The result that each optimizing obtains can utilize quality model to carry out off-line prediction to product quality, avoids the artificial primary parameter that often adjusts and all needs online verification, so that cause a large amount of waste products; And, in optimizing process, utilize rule to retrain parameter, ensure that the parameter that optimizing obtains meets production reality.The present invention can weave into visual software with computerese, only need on software interface given quality index, just can obtain required process parameter value very soon, the optimization of process parameter be determined convenient and quick.

The present invention is applied widely, extend in the process industries such as Coal Chemical Industry, petrochemical industry, metallurgy, solve the empirical and tentative problem of processing parameter setting in production run, not only save energy and reduce the cost, and to stabilized product quality, enhance productivity significant.

Accompanying drawing explanation

Fig. 1 is the technology of the present invention route map.

Fig. 2 is neural network quality model predictive simulation result figure.

Fig. 3 is decision tree diagram.

Fig. 4 is BP neural network structure figure.

Fig. 5 is decision tree structure figure.

Fig. 6 is particle shifting principle figure.

Embodiment

Embodiment 1 is based on the defining method of the process flow industry process technological parameter of data-driven

As shown in Figure 1, concrete steps are Technology Roadmap:

1st step, set up quality model extracting method: first existing production data is normalized, eliminate the impact of different dimension, the data extracted in production run are utilized to set up quality model, and utilize the rule knowledge between the method Extraction technique of rule extraction and quality index, formation rule storehouse;

2nd step, setting quality target value, the initial value of algorithm parameter and the initial value of process parameter: technological parameter is when initialization, the rule of its correspondence is first found according to given quality target value, then by technological parameter initialization in the scope of rule, reduction optimizing space, accelerates speed of searching optimization;

3rd step, quality model: utilize quality model, predicts product quality according to the technological parameter after optimizing;

4th step, counting yield quality: according to the product quality of prediction, calculate the error with given quality target value, and in this, as the evaluation index optimized, error is less, proves with desired value more close;

5th step, judge whether product quality meets the demands: error amount and given threshold value are contrasted, if be less than threshold value, meet the demands, then stop optimizing, the procedure parameter value of its correspondence is optimal value, does not meet the demands, adopt optimization method optimizing: according to optimization method, renewal is optimized to process parameter; Obtain new process parameter value and then go to 3 steps;

Embodiment 2 emulation experiment

In order to verify the validity of the inventive method, emulation experiment is carried out to the optimization defining method of the technological parameter in the present invention.

This emulation experiment, for certain iron and steel enterprise's strip hot-dip galvanizing production run, according to given quality index, is optimized and is determined technological parameter.

Technological parameter comprises: air pressure p, air knife are to the distance d and the unit speed s that are with steel, and target product quality is zinc layer weight w.

In process of production, generally the zinc layer weight w above control cincture steel is carried out by these three technological parameters of adjustment p, d, s.

When the zinc layer weight that production one is specified, need to regulate this three parameters simultaneously, main by artificial experience operation at present, cause zinc layer weight to be controlled quickly and accurately, occur waste product.

The present embodiment is chosen 1000 groups of strip hot-dip galvanizing production datas and is utilized 800 groups of data to set up neural network quality model, and utilizes all the other 200 groups of data to carry out prediction checking.

Predicted root mean square error reaches 0.065, and precision reaches 94.5%,

Neural network quality model predictive simulation result figure as shown in Figure 2.

Embodiment 3 decision tree extracting rule

Equally, adopt identical data, utilize decision tree extracting rule, all the other 200 groups of data are verified model.

Here first to classify to zinc layer weight data, according to the different requirements to zinc layer weight, zinc layer weight is divided into 3 classifications, i.e. 115<w≤146 (1 class), 80≤w<115 (2 class), 48≤w<80 (3 class), and using the root node of these 3 classifications as decision tree, input attributes p, d, s are then as the leaf node of tree.

Fig. 3 is the decision tree obtained, and the rule of its correspondence is:

Rule 1: if 26 >=p, 15 >=d, then 120 >=w > 80

Rule 2: if 26 >=p, 15 < d, 658 >=s, then 120 >=w > 80

Rule 3: if 26 >=p, 15 < d, 65.7 < s, then 160 >=w > 120

Rule 4: if 26 < p, 15 >=d, then 80 >=w > 40

Rule 5: if 26 < p, 18 >=d > 15, then 120 >=w > 80

Rule 6: if 26 < p, 18 < d, then 160 >=w > 120

To plate 60g/m to belt steel surface ²during zinc layer weight, can find that this zinc layer weight corresponds to rule 4, namely constrain technological parameter span.

Embodiment 4 particle group optimizing method

Utilize the method for particle group optimizing, first determine that the number of population is 50, iterations is 150, the position X of initialization particle in regular scope _i=(p, d, s) (established technology parameter) and speed, assigned error is at ± 0.5g/m ²in scope, zinc layer weight is with 60g/m ²as set-point, utilize neural network quality model to predict the zinc layer weight that each particle is corresponding, and contrast with given carrying out, if wherein the error of certain particle is at ± 0.5g/m ²in scope, then namely the position of this particle is required technological parameter; If do not meet the demands, then particle carries out iteration, more the position of new particle and speed, and utilizes the position of rule constrain particle, until meet the demands or reach the precision of regulation.Optimized by successive ignition, the result obtained is: p=35.5, d=12.9, v=105.4, and in regular scope, zinc layer weight is 60g/m ², meet the demands completely.

Embodiment 5 neural network

As shown in Figure 4, neural network is generally divided into display in input layer, hidden layer 2, output layer 3, figure to only have a hidden layer to neural network structure.In order to without loss of generality, suppose there be h hidden layer, P training sample, i.e. P inputoutput pair (X here _k, Y _k), (k=1,2 ..., P).Wherein, X _kfor a kth sample input vector: " 1 " represents ground floor, and n is the dimension of input amendment, the neuron number namely in input layer; Y _kdesired output vector for a kth sample: Y _k=(y _k1, y _k2...., y _km), m is the dimension of output vector, the neuron number namely in output layer.

The activation function of output layer and hidden layer is here for S type activation function.

S type activation function: $f\left(x\right)=\frac{1}{1+e^{-x}}$

For a kth sample, i-th neuronic output valve of h layer network can be obtained by the forward-propagating process of working signal:

$x_{\mathrm{ki}}^h=f\left(u_{\mathrm{ki}}^h\right)---\left(1.1\right)$

$u_{\mathrm{ki}}^h=\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^h{x}_{\mathrm{kj}}^{h-1}+{\mathrm{\&theta;}}_i^h---\left(1.2\right)$

Wherein, be a h-1 layer jth neuronic output valve, be i-th neuron and a h-1 layer jth neuronic connection weights of h layer, be i-th neuronic threshold value of h layer.

From the input layer of network to hidden layer, then to output layer, each neuronic output valve can be obtained successively with formula (1.1), (1.2).

Provide the derivation of BP algorithm mathematics expression formula below.For convenience of explaining, below illustrate all for a kth sample.

The desired output of an output layer kth sample with the actual square error exported is:

$E=\frac{1}{2}\underset{l=1}{\overset{m}{\mathrm{\&Sigma;}}}{(x_l^H-y_l)}^2=\frac{1}{2}\underset{l=1}{\overset{m}{\mathrm{\&Sigma;}}}{\left(e_l\right)}^2---\left(1.3\right)$

At amendment w _ij, θ _itime, Δ w _ij, Δ θ _iand E _knegative gradient direction relevant, that is:

$\begin{array}{cc}\mathrm{\&Delta;}{w}_{\mathrm{ij}}=-\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}}& \mathrm{\&Delta;}{\mathrm{\&theta;}}_i=-\frac{\&PartialD;E}{\&PartialD;{\mathrm{\&theta;}}_i}\end{array}$

In order to improve the learning ability of neural network, in network training, add learning rate η, that is:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}}---\left(1.4\right)$

$\mathrm{\&Delta;}{\mathrm{\&theta;}}_i=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;{\mathrm{\&theta;}}_i}---\left(1.5\right)$

In the learning process of reality, learning rate η is very large on the impact of learning process.η is the step-length by gradient search.

Therefore, modified weight formula is:

w _ij(t+1)＝w _ij(t)+Δw _ij(t) (1.6)

Threshold value correction formula is:

θ _i(t+1)＝θ _i(t)+Δθ _i(t) (1.7)

Wherein, t is for revising number of times.

By formula (1.1), (1.2), (1.3) can obtain: $\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}}=\frac{\&PartialD;E}{\&PartialD;u_i^h}\frac{\&PartialD;u_i^h}{\&PartialD;w_{\mathrm{ij}}},\frac{\&PartialD;E}{\&PartialD;{\mathrm{\&theta;}}_i}=\frac{\&PartialD;E}{\&PartialD;u_i^h}\frac{\&PartialD;u_i^h}{\&PartialD;{\mathrm{\&theta;}}_i}$

By formula (1.1), (1.2), thus have: $\frac{\&PartialD;u_i^h}{\&PartialD;w_{\mathrm{ij}}}=\frac{\&PartialD;(\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^h{x}_{\mathrm{ij}}^{h-1}+{\mathrm{\&theta;}}_i^j)}{\&PartialD;w_{\mathrm{ij}}}=x_j^{h-1}$

$\frac{\&PartialD;u_i^h}{\&PartialD;{\mathrm{\&theta;}}_i}=\frac{\&PartialD;\left(\underset{j}{\mathrm{\&Sigma;}{w}_{\mathrm{ij}}^h{x}_{\mathrm{kj}}^{h-1}+{\mathrm{\&theta;}}_i^h}\right)}{\&PartialD;{\mathrm{\&theta;}}_i}=1$

Then

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}^h=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;w_{\mathrm{ij}}^h}=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;u_i^h}{x}_j^{h-1}---\left(1.8\right)$

$\mathrm{\&Delta;}{\mathrm{\&theta;}}_i=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;{\mathrm{\&theta;}}_i}=-\mathrm{\&eta;}\frac{\&PartialD;E}{\&PartialD;u_i^h}---\left(1.9\right)$

By formula (1.1), (1.2), (13) are known

$\frac{\&PartialD;E}{\&PartialD;u_i^h}=\frac{\&PartialD;E}{\&PartialD;x_i^h}\frac{\&PartialD;x_i^h}{\&PartialD;u_i^h}=\frac{\&PartialD;E}{\&PartialD;x_i^h}{x}_i^h(1-x_i^h)---\left(1.10\right)$

Existing layering is considered, obtains

(1) output layer

If h=H, then illustrate the output of output layer H.

By formula (1.3): wherein y _lexpectation value, i.e. constant, therefore have:

$\frac{\&PartialD;E}{\&PartialD;x_i^h}=\frac{\&PartialD;E}{\&PartialD;x_i^H}=\frac{\&PartialD;\left[\frac{1}{2}\underset{l=1}{\overset{m}{\mathrm{\&Sigma;}}}{(x_l^H-y_l)}^2\right]}{\&PartialD;x_i^H}=x_l^H-y_l---\left(1.11\right)$

According to formula (1.10) and formula (1.11), have

$\frac{\&PartialD;E}{\&PartialD;u_i^h}=\frac{\&PartialD;E}{\&PartialD;u_i^H}=(x_l^H-y_i)---\left(1.12\right)$

(2) hidden layer

If h<H, then this layer is hidden layer, at this moment should consider the effect of last layer to it.

$\frac{\&PartialD;E}{\&PartialD;x_i^h}=\underset{l}{\mathrm{\&Sigma;}}\frac{\&PartialD;E}{\&PartialD;u_i^{h+1}}\frac{\&PartialD;u_i^{h+1}}{\&PartialD;x_i^h}$

By formula (1.2), $u_{\mathrm{ki}}^h=\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^h{x}_{\mathrm{kj}}^{h-1}+{\mathrm{\&theta;}}_i^h,$ When can obtain, $u_l^{h+1}=\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{lj}}^{h+1}{x}_{\mathrm{kj}}^h+{\mathrm{\&theta;}}_i^{h+1},$ Thus have $\frac{\&PartialD;u_i^{h+1}}{\&PartialD;x_i^h}=\frac{\&PartialD;[\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{lj}}^{h+1}{x}_{\mathrm{kj}}^h+{\mathrm{\&theta;}}_l^{h+1}]}{\&PartialD;x_i^H}=w_{\mathrm{lj}}^{h+1}$

$\frac{\&PartialD;E}{\&PartialD;x_i^h}=\underset{l}{\mathrm{\&Sigma;}}\frac{\&PartialD;E}{\&PartialD;u_i^{h+1}}{w}_{\mathrm{lj}}^{h+1}$

Convolution (1.10) has: $\frac{\&PartialD;E}{\&PartialD;u_i^h}=\frac{\&PartialD;E}{\&PartialD;x_i^h}\frac{\&PartialD;x_i^h}{\&PartialD;u_i^h}=x_i^h(1-x_i^h)\underset{l}{\mathrm{\&Sigma;}}\frac{\&PartialD;E}{\&PartialD;u_{i9}^{h+1}}{w}_{\mathrm{lj}}^{h+1}$

For three-layer network (namely last one deck is for output layer, and only has one deck hidden layer)

Note $d_i^H=\frac{\&PartialD;E}{\&PartialD;u_i^h}=(x_i^H-y_i)x_i^H(1-x_i^H)$

The weights of output layer, threshold value correction formula:

$w_{\mathrm{ij}}(t+1)=w_{\mathrm{ij}}\left(t\right)+\mathrm{\&Delta;}{w}_{\mathrm{ij}}\left(t\right)=w_{\mathrm{ij}}\left(t\right)-\mathrm{\&eta;}{d}_i^H{x}_j^{H-1}---\left(1.13\right)$

${\mathrm{\&theta;}}_i(t+1)={\mathrm{\&theta;}}_i\left(t\right)+\mathrm{\&Delta;}{\mathrm{\&theta;}}_i\left(t\right)={\mathrm{\&theta;}}_i\left(t\right)-\mathrm{\&eta;}{d}_i^H---\left(1.14\right)$

Weights, the threshold value correction formula of middle hidden layer:

$w_{\mathrm{ij}}(t+1)=w_{\mathrm{ij}}\left(t\right)+\mathrm{\&Delta;}{w}_{\mathrm{ij}}\left(t\right)=w_{\mathrm{ij}}\left(t\right)-\mathrm{\&eta;}{x}_i^h(1-x_i^h)x_j^{h-1}\underset{l}{\mathrm{\&Sigma;}}{d}_l^H{w}_{\mathrm{li}}^H---\left(1.15\right)$

${\mathrm{\&theta;}}_i(t+1)={\mathrm{\&theta;}}_i\left(t\right)+\mathrm{\&Delta;}{\mathrm{\&theta;}}_i\left(t\right)={\mathrm{\&theta;}}_i\left(t\right)-\mathrm{\&eta;}{x}_i^h(1-x_i^h)\underset{l}{\mathrm{\&Sigma;}}{d}_l^H{w}_{\mathrm{li}}^H---\left(1.16\right)$

Four formula (1.13) more than obtained, (1.14), (1.15), (1.16) are four important formulas in neural network algorithm implementation procedure.

Embodiment 6 traditional decision-tree

Decision Tree algorithms utilizes information gain as the choice criteria of attribute, and the attribute selecting information gain maximum produces decision tree node, to make, when testing each non-leaf node, can obtain and record maximum classification information about tested.Branch is set up by the different values of this attribute, again subset recursive call the method for each branch is set up to the branch of decision tree node, until all collection only comprise other data of same class, finally obtain a decision tree, from decision tree, I just can obtain rule, and the form of decision tree as shown in Figure 5.

If X is the set of training data sample, n sample altogether, m classification, n training sample is divided into m class by the destination of study exactly, is designated as C={X ₁, X ₂..., X _m, if the training sample number of the i-th class is | X _i|=C _i, the probability that sample belongs to i class is expectation information then needed for this sample X classification is provided by following formula:

$I(X_1,X_2,...,X_m)=-\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}P\left(X_i\right){\log }_{2}P\left(X_i\right)---\left(2.1\right)$

If attribute A has v different value { a ₁, a ₂... a _v, at A=a _jthe example number belonging to the i-th class in situation is C _ij, i.e. P (X _i/ A=a _j) for the value of testing attribute A be a _jtime belong to the probability of the i-th class, note Y _jfor A=a _jtime example set, then its expectation information of training sample set pair attribute A is:

$E\left(Y_j\right)=\underset{j=1}{\overset{v}{\mathrm{\&Sigma;}}}P(X_i/A=a_j){\log }_{2}P(X_i/A=a_j)---\left(2.2\right)$

Each A=a that testing attribute A grows _jleaf node X _iinformation entropy for classified information is

$E\left(A\right)=\underset{j=1}{\overset{v}{\mathrm{\&Sigma;}}}P(A=a_j)H\left(Y_j\right)---\left(2.3\right)$

The information gain of acquisition is by attribute A top set:

Gain(X,A)＝I(X ₁,X ₂,...,X _m)-E(A) (2.4)

Gain (X, A) is larger, illustrates that the information selecting testing attribute A to provide for classifying is larger, therefore, according to the segmentation of the choice criteria as the testing attribute training sample of Gain (X, A), obtain into decision tree, finally decision tree is changed into rule.

Described process parameter optimizing adopts particle group optimizing (Particle Swarm Optimization, PSO) method, the rule that bond quality model and decision tree obtain realizes the optimization of technological parameter, other optimization methods go for this patent too, as ant group algorithm, fish-swarm algorithm, genetic algorithm etc.

Embodiment 7 particle group optimizing method

In PSO algorithm, the potential solution of each optimization problem is a bird in search volume, is referred to as " particle ", and namely the position of each particle is exactly a potential solution.Particle number is called population scale m, and i-th particle is X at the positional representation of d dimension space _i=(x _i1, x _i2..., x _id), (i=1,2 ..., m), speed V _i=(v _i1, v _i2..., v _id) determine the displacement of particle search mikey iterations.Calculate the p of each particle _f, fitness function is generally determined by function optimised in practical problems.According to the p of each particle _f, upgrade the p of each particle _band g _b.Here whether " optimum " is determined by concrete optimization problem: if this problem maximizing, then particle fitness is more optimum; Otherwise if this problem is minimized, then particle fitness is less is optimum.Particle is worth upgrades its speed and position by dynamically following the tracks of individual extreme value and the overall situation most.Particle upgrades its speed and position according to following formula:

v _ij(t+1)＝v _ij(t)+c1′rand()′(p _bj(t)-x _ij(t))+c2′rand()′(g _bj-x _ij(t)) 3.1)

x _ij(t+1)＝x _ij(t)+v _ij(t+1) (3.2)

In formula, i is i-th particle, j=1,2 ..., d.C1, c2 are Studying factors, are namely respectively regulated to the maximum step-length of the best particle of the overall situation and the flight of individuality best particle direction, if too little, then particle possibility wide region, if too large, then can cause to target area flying to suddenly, or fly over target area.T is iterations, and rand () is for being evenly distributed on the random number between (0-1).At no point in the update process, particle can not exceed the maximal rate v of algorithm setting in the speed that every one dimension flies _max, namely | v _ij(t+1) | <v _max, otherwise v _ij(t+1)=v _maxor=-v _max.Larger v is set _maxthe ability of searching optimum of particle populations can be ensured, v _maxthe local search ability of less then population is strengthened.Meanwhile, the coordinate of the every one dimension of particle is also limited in x in allowed band _max.

The shifting principle of particle as shown in Figure 6.

Utilize the concrete steps of particle group optimizing technological parameter as follows:

1. given quality index T and optimize precision E, using technological parameter air pressure p, air knife to the position of being with the distance d of steel and strip speed s as particle $X=\left[\begin{array}{c}{p}_{\mathrm{1,1}},p_{\mathrm{1,2}},s_{\mathrm{1,3}}\\ p_{\mathrm{2,1}},p_{\mathrm{2,2}},s_{\mathrm{2,3}}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ p_{m,1},d_{m,2},s_{m,3}\end{array}\right],$ The speed of initialization particle number m, each particle $V=\left[\begin{array}{c}{v}_{\mathrm{1,1}}{,v}_{\mathrm{1,2}},v_{\mathrm{1,3}}\\ v_{\mathrm{2,1}},v_{\mathrm{2,2}},v_{\mathrm{2,3}}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ v_{m,1},v_{m,2},v_{m,3}\end{array}\right],$ Position X, speed maximal value v _max, largest loop iterations t _max;

2. the individual extreme value place p of initialization _b=[p _p, d _p, s _p] and the individual extreme value p of correspondence _bf, global extremum position g _b=[p _g, d _g, s _g] and the global extremum g of correspondence _bf;

3. with the position X of particle for input, the quality model that utilization is set up above calculates quality Y corresponding to each particle, and contrasts with quality index T, calculates the fitness p of particle _f=| Y-T|;

4. p is drawn by comparing _b, p _bf, g _b, g _bf.To minimize:

If p _f<p _bf, p _bf=p _f, p _b=x _i=[p _{i, 1}, d _{i, 2}, s _{i, 3}]; Otherwise, p _bf, p _bconstant;

If p _f<g _bf, g _bf=p _f, g _b=x _i=[p _{i, 1}, d _{i, 2}, s _{i, 3}]; Otherwise, g _bf, g _bconstant;

5. position and the speed of each particle is upgraded;

The more speed of new particle and position:

V _ij(t+1)＝V _ij(t)+c1′rand()′(p _bj(t)-X _ij(t))+c2′rand()′(g _bj-X _ij(t))X _ij(t+1)＝X _ij(t)+V _ij(t+1)

Consider that the speed after upgrading is whether in the scope limited,

If V _ij(t+1) >v _max, then V _ij(t+1)=v _max;

If V _ij(t+1) <-v _max, then V _ij(t+1)=-v _max;

Otherwise V _ij(t+1) constant;

Meanwhile, in order to consider technological parameter in each searching process all in the scope of rule, first finding corresponding rule according to given quality index T, namely obtaining the scope of technological parameter, if the scope of a certain parameter is $B=\left[\begin{array}{c}{p}_{\min },d_{\min },s_{\min }\\ p_{\max },d_{\max },s_{\max }\end{array}\right],$ Then retrain as follows,

If X _i1(t+1) >p _max, then X _i1(t+1)=p _max;

If X _i1(t+1) <p _min, then X _i1(t+1)=p _min;

If X _i2(t+1) >d _max, then X _i2(t+1)=d _max;

If X _i2(t+1) <d _min, then X _i2(t+1)=d _min;

If X _i3(t+1) >s _max, then X _i3(t+1)=s _max;

If X _i3(t+1) <s _min, then X _i3(t+1)=s _min;

Otherwise X _ij(t+1) constant;

6. whether number of comparisons reaches maximum iteration time or default precision; Precision, i.e. p is preset if meet _f<E, algorithm convergence, the g of last iteration _b, g _bfbe exactly the optimal value position required by us and optimal value; Otherwise return step 3., algorithm continues iteration.

Last it is noted that obviously, above-described embodiment is only for example of the present invention is clearly described, and the restriction not to embodiment.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here exhaustive without the need to also giving all embodiments.And thus the apparent change of amplifying out or variation be still among protection scope of the present invention.

Claims (1)

1. a defining method for process flow industry process technological parameter, is characterized in that, described method comprises the following steps:

1) quality modeling: according to the production data gathered in actual production process, adopts the method for neural network, with process parameter be input, product quality sets up quality model for exporting;

2) rule extraction: according to the production data gathered in actual production process, the method utilizing Decision Tree Rule to extract from the rule of correspondence between extracting data constrained input, formation rule storehouse;

3) process parameter optimizing:

1. given quality index T and optimize precision E, using technological parameter air pressure p, air knife to the position of being with the distance d of steel and strip speed s as particle $X=\left[\begin{array}{c}{p}_{\mathrm{1,1}},d_{\mathrm{1,2}},s_{\mathrm{1,3}}\\ p_{\mathrm{2,1}},d_{\mathrm{2,2}},s_{\mathrm{2,3}}\\ ...\\ p_{m,1},d_{m,2},s_{m,3}\end{array}\right],$ The speed of initialization particle number m, each particle $V=\left[\begin{array}{c}{v}_{\mathrm{1,1}},v_{\mathrm{1,2}},v_{\mathrm{1,3}}\\ v_{\mathrm{2,1}},v_{\mathrm{2,2}},v_{\mathrm{2,3}}\\ ...\\ v_{m,1},v_{m,2},v_{m,3}\end{array}\right],$ Position X, speed maximal value v _max, largest loop iterations t _max;

2. the individual extreme value place p of initialization _b=[p _p, d _p, s _p] and the individual extreme value p of correspondence _bf, global extremum position g _b=[p _g, d _g, s _g] and the global extremum g of correspondence _bf;

3. with the position X of particle for input, the quality model that utilization is set up above calculates quality Y corresponding to each particle, and contrasts with quality index T, calculates the fitness p of particle _f=| Y-T|;

4. p is drawn by comparing _b, p _bf, g _b, g _bf;

If p _f<p _bf, p _bf=p _f, p _b=x _i=[p _{i, 1}, d _{i, 2}, s _{i, 3}]; Otherwise, p _bf, p _bconstant;

If p _f<g _bf, g _bf=p _f, g _b=x _i=[p _{i, 1}, d _{i, 2}, s _{i, 3}]; Otherwise, g _bf, g _bconstant;

5. position and the speed of each particle is upgraded;

The more speed of new particle and position:

V _ij(t+1)＝V _ij(t)+c1×rand()×(p _bj(t)-X _ij(t))+c2×rand()×(g _bj-X _ij(t))

X _ij(t+1)＝X _ij(t)+V _ij(t+1)

Consider that the speed after upgrading is whether in the scope limited,

If V _ij(t+1) >v _max, then V _ij(t+1)=v _max;

If V _ij(t+1) <-v _max, then V _ij(t+1)=-v _max;

Otherwise V _ij(t+1) constant;

Meanwhile, in order to consider technological parameter in each searching process all in the scope of rule, first finding corresponding rule according to given quality index T, namely obtaining the scope of technological parameter, if the scope of a certain parameter is $B=\left[\begin{array}{c}{p}_{\min },d_{\min },s_{\min }\\ p_{\max },d_{\max },s_{\max }\end{array}\right],$ Then retrain as follows,

If X _i1(t+1) >p _max, then X _i1(t+1)=p _max;

If X _i1(t+1) <p _min, then X _i1(t+1)=p _min;

If X _i2(t+1) >d _max, then X _i2(t+1)=d _max;

If X _i2(t+1) <d _min, then X _i2(t+1)=d _min;

If X _i3(t+1) >s _max, then X _i3(t+1)=s _max;

If X _i3(t+1) <s _min, then X _i3(t+1)=s _min;

Otherwise X _ij(t+1) constant;

6. whether number of comparisons reaches maximum iteration time or default precision; Precision, i.e. p is preset if meet _f<E, algorithm convergence, the g of last iteration _b, g _bfbe exactly the optimal value position required by us and optimal value; Otherwise return step 3., algorithm continues iteration.