CN104732067A - Industrial process modeling forecasting method oriented at flow object - Google Patents

Abstract

The invention discloses an industrial process modeling forecasting method oriented at a flow object. The method comprises the following steps: building FNT models, extracting an industrial flow object original data set S from a data warehouse which has already been generated by the flow object, creating an initial species group of the FNT models, and customizing the individual numbers of the species group as required, wherein each individual represents an FNT model; utilizing the PIPE algorithm for optimizing FNT model structures, and adopting mean square errors or root-mean-square errors for fitness functions; utilizing the particle swarm optimization (PSO) algorithm for optimizing FNT model parameters; utilizing the FNT models for conducting modeling forecast for a flow object production process. According to the method, based on the flexible neural tree, an equation of variation tendency among measuring point data of the flow object is obtained, the industrial production process is simulated, based on relevant parameters of a current production state, production states in a period of time in the future are forecast, so that an enterprise is assisted and instructed for adjusting the production process parameters, and the production is guided for drawing on advantages and avoiding disadvantages in a microcosmic sense.

Description

A kind of industrial process modeling Forecasting Methodology of Process-Oriented object

Technical field

The present invention relates to industrial flow production field, particularly relate to a kind of industrial process modeling Forecasting Methodology of Process-Oriented object.

Background technology

Along with the development of production technology, the control of procedure links is more and more stricter, and the interact relation therefore between manufacturing parameter becomes also more and more important; And along with the expansion of production data scale and for a long time accumulation, cause the generation of the production data of magnanimity, and this certainly will increase the relevance between the complicacy of procedure parameter and parameter further, sets up also to the model of flow object and brings larger difficulty.The correct effective parameter of selection is the prerequisite of effectively carrying out process model building, and the selection of flow object procedure parameter at present depends on the experience of workers with long time accumulation to a great extent, is theoretically unsound and science.

Summary of the invention

In order to overcome the deficiencies in the prior art, the invention provides a kind of industrial process modeling Forecasting Methodology of Process-Oriented object, the variation tendency formula of each measuring point data of flow object is obtained based on flexible Neural Tree, industrial processes are simulated, based on the production status of correlation parameter prediction following a period of time of current production status, thus auxiliary direction enterprise adjustment production procedure parameter, microcosmic guide production go after profits and advoid disadvantages.

For achieving the above object, the present invention takes following technical scheme:

An industrial process modeling Forecasting Methodology for Process-Oriented object, comprises the steps:

(1) flexible Neural Tree (flexible neural tree) FNT model is set up, industrial flow object raw data set S is extracted from the data warehouse that flow object has generated, create the initial population of FNT model, population at individual number customizes as required, and each individuality represents a FNT model;

(2) utilize PIPE algorithm optimization FNT model structure, adaptive value function adopts square error or root-mean-square error;

(3) Particle Swarm (PSO) algorithm optimization FNT model parameter is utilized;

(4) FNT model is utilized to carry out modeling and forecasting to flow object production run.

Further, in described step (1) industrial flow object raw data set S, each attribute or field represent the state of certain node in industrial flow object, its value can change along with the change of the change of time and other node states, the data of FNT model treatment need [0,1] between, so need to be normalized raw data set S, method for normalizing is as formula (1), wherein X is pending raw data, MAX and MIN is the minimum and maximum value of data attribute in raw data set S belonging to X respectively;

Y＝(X-MIN)/(MAX-MIN) (1)

Then raw data set S later for normalization is loaded into database, forms the data warehouse that can be directly used in data mining.

Further, the collection of functions F that uses of described FNT model and termination message collection T is described below:

S＝F∪T＝{+ ₂,+ ₃,...,+ _N}∪{x ₁,...,x _n} (2)

Wherein ,+ _irepresent non-leaf nodes information, i=2,3 ..., N, i representative function+ _ithe number of corresponding input variable; x ₁, x ₂... x _nfor leaf node information; The output of a non-leaf nodes is regarded as a flexible neuron calculate, namely+ _iit is the flexible neuron with i input; In the constructive process of Neural Tree, if nonterminal information+ _iselected, the actual value of i is random generation, its expression+ _iconnection weights between this node and his child's contact; Two adjustable parameter a of flexible actuation function _iand b _ialso be random generation; The excitation function of FNT; Be expressed as:

$f=(a_i,b_{i}x)=e^{-{(x-a_i/b_i)}^2---\left(3\right)}$

Flexible neuron+ _nbeing calculated as follows of input, node+ _nexcitation and be:

${\mathrm{net}}_n=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_j*x_j---\left(4\right)$

X _jrepresent the input of+n node, j=1,2 ... n.

Further, the search volume of described step (2) PIPE algorithm is the tree-like population produced according to the raw data set S preset, and individuality results from the probability vector space covering institute's likely individuality; Individual with the generation of probability prototype tree, expression is the tree construction of a n dimension, the maximum branch number that information in n representative function collection can produce, the non-leaf nodes of tree results from collection of functions F, leaf node results from termination message collection T, the number of the subtree of each node is decided by the function information of each node producible point of number, and the input of each branch has corresponding subtree to calculate, the analysis mode of tree be depth-first from left to right.

Further, described step (2) PIPE algorithm flow comprises:

(21) individual generation, produces individual with probability prototype tree, represent body one by one, wherein 0<j<=PS, PS represent the scale that per generation is individual;

(22) individual evaluation, each population at individual all to evaluate in given problem, and according to predefined adaptive value function formula, as shown in formula (5) and (6), calculate adaptive value the best individuality (individuality that adaptive value is minimum) of current population is marked as program performs till now, and best individuality is stored in in, Fit (i) represents i-th individual adaptive value, and p represents number of samples, with represent the actual sequence value of a jth sample and a jth sample final output valve through i-th individuality calculating respectively,

$\mathrm{Fit}\left(i\right)=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2---\left(5\right)$

$\mathrm{Fit}\left(i\right)=\sqrt{\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2}---\left(6\right)$

(23) individual study, in order to make the increase of the probability of current best individuality, need the probable value revising prototype tree, this process is called prototype tree and evolves, implementation procedure: the probability of first current best individuality value and produce this preferably individual all node N _jall relevant, computing formula is as follows:

$P\left(P_{{\mathrm{ROG}}_b}\right)\underset{j:N_j}{\mathrm{\&Pi;}}{P}_j\left(I_j\left(P_{\mathrm{RO}{G}_b}\right)\right)---\left(7\right)$

represent individual at a jth node place information, individual destination probability be calculated as follows:

$P_{\mathrm{TARGET}}=P\left(P_{\mathrm{RO}{G}_b}\right)+(1-P\left(P_{\mathrm{RO}{G}_b}\right))\&CenterDot;\mathrm{lr}\&CenterDot;\frac{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{\mathrm{ROG}}^{\mathrm{el}}\right)}{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{\mathrm{RO}{G}_B}\right)}---\left(8\right)$

Lr is a constant, represents learning rate; ε being a user-defined positive constant, according to P _tARGET, the probability of all single nodes all will repeatedly increase,

C ^lra constant, the number of times of impact circulation.C ^lrless, individual probability more close to destination probability P _tARGET, the number of times of circulation is more.Practice shows to work as c ^lrwhen=0.1, good balance can be obtained between degree of accuracy and speed.All adaptation vectors again standardized;

(24) variation of prototype tree, at node N _jplace has and produces current best individuality relevant vector dimension P _j(I) all with probability morph:

$P_{M_p}=\frac{P_M}{n\&CenterDot;\sqrt{\left|P_{{\mathrm{ROG}}_b}\right|}}---\left(9\right)$

P _mbe a parameter with definition, represent overall mutation probability, n represents the number of information in information set S, represent individual nodes.All made a variation according to formula below by the probability vector selected:

P _j(I)＝P _j(I)+mr·(1-P _j(I)) (10)

Mr is another constant user-defined, represents aberration rate;

(25) prototype hedge clipper branch, after the circulation of every generation terminates, prototype tree is all by beta pruning, and the probability vector of the node of subtree has at least a probable value to be greater than pruning threshold Tp just and be cut, and Tp is a larger decimal normally;

(26) termination condition, or circulation finds satisfied solution vector until the fixed value that reaches a program appraisal is a kind of.

Further, the mathematical description of described step (3) particle swarm optimization is as follows: set particle populations scale as N, and wherein the coordinate position vector representation of each particulate in D dimension space is velocity vector is expressed as particulate personal best particle, namely the optimal location that lives through of this particulate, is designated as colony's optimal location, the optimal location that namely in this Particle Swarm, any individual lives through, is designated as ${\stackrel{\&RightArrow;}{P}}_g=(p_{g1},p_{g2},...,p_{\mathrm{gd}},...,p_{\mathrm{gD}});$

The iterative formula of personal best particle is:

Colony's optimal location is position best in personal best particle, and speed and position iterative formula are respectively:

$v_{i,t+1}^d=v_{i,t}^d+c_1*\mathrm{rand}*(p_{i,t}^d-x_{i,t}^d)+c_2*\mathrm{Rand}*(p_{g,t}^d-x_{i,t}^d)---\left(12\right)$

$x_{i,t+1}^d=x_{i,t}^d+v_{i,t+1}^d---\left(13\right)$

Further, after described step (3) parameter optimization, FNT interior joint+ _noutput be calculated as follows:

${\mathrm{out}}_n=f(a_n,b_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2}---\left(14\right)$

And the result that whole FNT optimizes export can according to the principle of depth-first from left to right recursive calculation, as shown in formula (15):

Wherein, x ₁, x ₂..., x _nrepresent the input of the tree construction child nodes that FNT model is corresponding; w ₁, w ₂..., w _nfor the weight that limit is corresponding, for what optimize in FNT modeling process, by PSO algorithm optimization; Y is that formula (14) exports.

Further, flow object critical workflow and corresponding production input and output parameter are input in step (2) and the revised FNT model of step (3) by described step (4), obtain the output function of shape as formula 16, wherein A _nand B _nthe parameter after two groups of PSO optimize, net _nit is the tree structure after PIPE optimizes; This function is exactly the Changing Pattern of production procedure parameter, carries out detailed predicting for the change of producing future:

${\mathrm{out}}_n=f(A_n,B_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2}---\left(16\right)$

Beneficial effect: complicated production data and the strong problem of relevance in Process-Oriented object of the present invention, utilize the method for flexible Neural Tree FNT model mathematical modeling to simulate production run, completed the structure and parameter optimization of FNT respectively by probability enhanced program evolution algorithm PIPE and particle swarm optimization PSO.The automatic screening of realization flow industrial process parameter, thus find the important parameter affecting production run, for process optimization provides theoretical foundation, make flowchart process optimization and control to there is science and specific aim more, to obtain better optimal control effect; And find the Changing Pattern of manufacturing parameter, generate the mathematical function of Parameters variation, following variation tendency will be predicted in current production status parameters input function.The present invention have predict the outcome accurately, to production environment restriction less, the advantage of fast operation.

Accompanying drawing explanation

Fig. 1 is the model structure of the flexible Neural Tree of the present invention.

Fig. 2 is the modeling overall flow schematic diagram of FNT of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the present invention is further described.

The industrial process modeling Forecasting Methodology of a kind of Process-Oriented object provided by the invention, comprises the steps:

(1) FNT model is set up

FNT model modeling comprises data encasement and model initialization.

The data encasement of FNT model modeling, first by extracting from the data warehouse that flow object has generated, the data of extraction are selected according to pre-designed rule, guarantee the selected participation modeling process of the data attribute relevant to flow object modeling, delete useless data attribute, the correlativity of data attribute and flow object modeling can be judged by domain expert.Then process the data of some redundancies, ambiguity, the data layout of original isomery can be united, form the raw data set S of FNT model treatment.In industrial flow object raw data set S, each attribute or field represent the state of certain node in industrial flow object, and its value can change along with the change of the change of time and other node states.FNT model of the present invention in order to these different internodal interactional relations in modeling industrial flow object, derives a funtcional relationship just, is used for the situation of change representing that certain node affects by other nodes.

Because the data of FNT model treatment need [0,1] between, so need to be normalized raw data set S, method for normalizing is shown in formula (1), wherein X is pending raw data, MAX and MIN is the minimum and maximum value of data attribute in raw data set S belonging to X respectively.

Y＝(X-MIN)/(MAX-MIN) (1)

Then raw data set S later for normalization is loaded into database, forms the data warehouse that can be directly used in data mining.

FNT model initialization is the raw data set S according to aforementioned pre-defined process industry, and create the initial population of FNT model, in population, individual amount can customize as required, and each individuality represents a FNT model.In this process, the T in FNT model is termination message collection, each attribute that the raw data of corresponding process industry is concentrated.F is collection of functions, represents the process to termination message collection or collection of functions.The root node of FNT model represents the output of whole model, certain field that the raw data of corresponding process industry is concentrated, and this field is not in T.Assuming that comprise n attribute in current raw data set S, be respectively x ₁, x ₂..., x _n.If we attempt to create root node is x _imodel, seek other n-1 attribute to attribute x _iaffect situation, now x _ithe root node of corresponding FNT model, as the output of whole model, other n-1 attributes form termination message collection T, i.e. from removing x in the leaf node of FNT model _in-1 attribute in select.

For the FNT model in initial population, can by probability spanning tree stochastic generation.Afterwards, based on process industry raw data set pair its train.FNT model structure allows to select input variable, the connection between non-adjacent layers, and different nodes can apply different excitation functions, and definition elite individuality is sky, and its adaptive value is a very large real number value.The model structure of flexible Neural Tree as shown in Figure 1, model structure allows to select input variable, and the connection between non-adjacent layers and different nodes can apply different excitation functions.According to customizing messages, structure and the corresponding structural parameters of model can utilize optimized algorithm optimization :+i (i=2,3, N) non-leaf nodes information is represented, the output of a non-leaf nodes is regarded a flexible neuron as and is calculated, and information+i can be regarded as the flexible neuron with i input; X1, x2 ... xn is leaf node information, i.e. production status parameter.In the constructive process of Neural Tree, if a nonterminal information+i is selected, the actual value of i is random generation, and it represents the connection weights between this node of+i and his child's contact.

The collection of functions F that generation FNT model is used and termination message collection T is described below:

S＝F∪T＝{+ ₂,+ ₃,...,+ _N}∪{x ₁,...,x _n} (2)

+ _i(i=2,3 ..., N) and represent non-leaf nodes information, wherein i representative function+ _ithe number of corresponding input variable; x ₁, x ₂... x _nfor leaf node information; The output of a non-leaf nodes is regarded as a flexible neuron calculate; By this viewpoint, can information+ _iregard the flexible neuron with i input as.In the constructive process of Neural Tree, if nonterminal information+ _iselected, the actual value of i is random generation, its expression+ _iconnection weights between this node and his child's contact.Two adjustable parameter a of flexible actuation function _iand b _ialso be random generation.The excitation function of FNT can be expressed as:

$f=(a_i,b_{i}x)=e^{-{(x-a_i/b_i)}^2---\left(3\right)}$

Flexible neuron+ _nbeing calculated as follows of input, node+ _nexcitation and be:

${\mathrm{net}}_n=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j*x_j---\left(4\right)$

X _j(j=1,2 ... n) input of+n node is represented.After completing FNT initial setting up, need to be optimized its structure and parameter: the present invention adopts probability enhanced program evolution algorithm (PIPE algorithm) to be optimized the structure of FNT, and then utilizes the optimum configurations of particle cluster algorithm to FNT to be optimized.

(2) PIPE algorithm optimization FNT model structure is utilized

The learning method of PIPE algorithm combines with on behalf of the study (Generation-Based Learning, GBL) on basis and elite's study (Elitist Learning, EL) two kinds of learning methods, based on GBL learning method.Adaptive value function adopts square error or root-mean-square error.A kind of comprehensive algorithm of PIPE algorithm, the search volume of PIPE algorithm is the tree-like population produced according to the raw data set S preset, individuality result from can cover likely individual probability vector space, originally, the distribution of probability vector is random, in continuous evolutionary process, probability vector is constantly adjusted.Mainly comprise: (1) increases the probability of the preferably individual corresponding informance of every generation; (2) probability of the corresponding informance of best individuality till now also may be increased; (3) individual information probability variation.

Individual expression is the tree construction of a n dimension, the maximum branch number that the information in n representative function collection can produce.The non-leaf nodes of tree results from collection of functions F, and leaf node results from termination message collection T.The number of the subtree of each node is decided by the function information of each node producible point of number, and the input of each branch has corresponding subtree to calculate.Tree analysis mode be depth-first from left to right.Probability prototype tree (PPT) stores the information coming from individuality and and guides evolutionary search.A PPT normally complete n fork having countless multiple node sets, and n represents the maximum input number of non-leaf nodes.Each node N in probability prototype tree _j, as j>=0, all contain a variable probability vector with a constant Rj, be a z dimensional vector, z represents the information number that information set S comprises, every one dimension p of vector _j(I) all represent at node N _jthe probability of place's selection information I, I ∈ S, the probability of all information is added and equals 1.The probability of the function information of each node and the probability of termination message are equal separately when initialization.The initialized formula of probability is as follows:

$p_j\left(I\right)=\frac{p_T}{l}\&ForAll;I:\&Element;T---\left(5\right)$

$p_j\left(I\right)=\frac{1-p_T}{k}\&ForAll;I:I\&Element;F---\left(6\right)$

L represents the number of termination message collection, the number of k representative function information set.Individual node is corresponding with PPT node.The information of first such as individual node results from first node of probability prototype tree.When information is selected, first produce the random number r1 between a 0-1, then first probable value p of random number and the corresponding probability vector of PPT first node ₁(I) compare, if r1>p ₁(I), then reduce r1, make r1=r1-p ₁(I), then continue to compare successively, until r1<=p backward _j(I).If now j<=n, then a jth information will be selected, otherwise information n is selected; If I=R, R example Vj (R) will replace R as the nodal information value of this individuality.If p _j(R) >TR (arbitrary constant threshold value), then Vj (R)=Rj, otherwise the value of Vj (R) produces at random with regard to the character of Dependence Problem.After the prototype occurring meeting predetermined threshold value is set, PIPE algorithm stops, the FNT vibrational power flow after being optimized.

Algorithm flow is as follows:

1, individual generation, produces individual with probability prototype tree, represent body (0<j<=PS one by one; PS represents the scale that per generation is individual).

2, individual evaluation, each population at individual all to evaluate in given problem, and according to predefined adaptive value function formula, as shown in formula (7) and (8), calculate adaptive value the best individuality (individuality that adaptive value is minimum) of current population is marked as program performs till now, and best individuality is stored in in.Fit (i) represents i-th individual adaptive value, and p represents number of samples, with represent the actual sequence value of a jth sample and a jth sample final output valve through i-th individuality calculating respectively.

$\mathrm{Fit}\left(i\right)=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2---\left(7\right)$

$\mathrm{Fit}\left(i\right)=\sqrt{\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2}---\left(8\right)$

3, individual study.In order to make the increase of the probability of current best individuality, need the probable value revising prototype tree.This process is called prototype tree and evolves.Implementation procedure: the probability of first current best individuality value and produce this preferably individual all node N _jall relevant.Computing formula is as follows:

$P\left(P_{{\mathrm{ROG}}_b}\right)\underset{j:N_j}{\mathrm{\&Pi;}}{P}_j\left(I_j\left(P_{\mathrm{RO}{G}_b}\right)\right)---\left(9\right)$

represent individual at a jth node place information.Individual destination probability be calculated as follows:

$P_{\mathrm{TARGET}}=P\left(P_{\mathrm{RO}{G}_b}\right)+(1-P\left(P_{\mathrm{RO}{G}_b}\right))\&CenterDot;\mathrm{lr}\&CenterDot;\frac{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{\mathrm{ROG}}^{\mathrm{el}}\right)}{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{\mathrm{RO}{G}_B}\right)}---\left(10\right)$

Lr is a constant, represents learning rate; ε is being a user-defined positive constant.According to P _tARGET, the probability of all single nodes all will repeatedly increase.

C ^lra constant, the number of times of impact circulation.C ^lrless, individual probability more close to destination probability P _tARGET, the number of times of circulation is more.Practice shows to work as c ^lrwhen=0.1, good balance can be obtained between degree of accuracy and speed.All adaptation vectors again standardized.

4, the variation of prototype tree.At node N _jplace has and produces current best individuality relevant vector dimension P _j(I) all with probability morph:

$P_{M_p}=\frac{P_M}{n\&CenterDot;\sqrt{\left|P_{{\mathrm{ROG}}_b}\right|}}---\left(11\right)$

P _mbe a parameter with definition, represent overall mutation probability, n represents the number of information in information set S, represent individual nodes.All made a variation according to formula below by the probability vector selected:

P _j(I)＝P _j(I)+mr·(1-P _j(I)) (12)

Mr is another constant user-defined, represents aberration rate.

5, prototype hedge clipper branch, after the circulation of every generation terminates, prototype tree is all by beta pruning, and the probability vector of the node of subtree has at least a probable value to be greater than pruning threshold Tp just and be cut, and Tp is a larger decimal normally, such as Tp=0.93.

6, termination condition, or circulation finds satisfied solution vector until the fixed value that reaches a program appraisal is a kind of.

Elite's learning algorithm is mainly to search in the good search volume found on behalf of the learning algorithm on basis.PPT is evolved towards the direction of optimum individual.Implementation method uses optimum individual in above-mentioned 3rd step individuality study replace current optimum individual elite's learning algorithm is more effective in small-scale and free noise problem.

After the study of above-mentioned algorithm optimization, preferably individual structure is just fixed and saves.

(3) parameter of Particle Swarm (PSO) algorithm to FNT is optimized

Particle swarm optimization is a kind of search procedure based on population, wherein each individuality is called particulate, be defined as the potential solution tieing up problem to be optimized in search volume at D, preserve its history optimal location and the memory of fine-grained optimal location, and speed, develop generation at each, the information of particulate is combined the component of regulating the speed about on every one dimension, then be used to calculate new particles position, particulate is the continuous state changing them in multidimensional search space, until arrive balance or optimum state, or exceed till calculating restriction, contact unique between the different dimensions of problem space is introduced by objective function.In continuous space coordinate system, the mathematical description of particle swarm optimization is as follows: set particle populations scale as N, and wherein the coordinate position vector representation of each particulate in D dimension space is velocity vector is expressed as particulate personal best particle (i.e. this particulate live through optimal location) is designated as colony's optimal location (optimal location that namely in this Particle Swarm, any individual lives through) is designated as without loss of generality, for minimization problem, in the particle swarm optimization of initial release, the iterative formula of personal best particle is:

Colony's optimal location is position best in personal best particle, speed and position iterative formula respectively Ru shown in (14) and (15), wherein φ ₁and φ ₂two converging factors.

$v_{i,t+1}^d=\mathrm{\&chi;}(v_{i,t}^d+{\mathrm{\&phi;}}_1*\mathrm{rand}*(p_{i,t}^d-x_{i,t}^d)+{\mathrm{\&phi;}}_2*\mathrm{Rand}*(p_{g,t}^d-x_{i,t}^d))---\left(14\right)$

$x_{i,t+1}^d=x_{i,t}^d+v_{i,t+1}^d---\left(15\right)$

After successive ignition, after the difference value that formula (14) and (15) change between generations is less than given threshold value, stop PSO algorithmic procedure, the colony's result obtained is exactly the FNT parameter value after optimizing.

After optimizing, FNT interior joint+ _noutput be calculated as follows:

${\mathrm{out}}_n=f(a_n,b_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2}---\left(16\right)$

And the result that whole FNT optimizes export can according to the principle of depth-first from left to right recursive calculation, as shown in formula (17):

Wherein, x ₁, x ₂..., x _nrepresent the input of the tree construction child nodes that FNT model is corresponding; w ₁, w ₂..., w _nfor the weight that limit is corresponding, for what optimize in FNT modeling process, by PSO algorithm optimization; Y is that formula (16) exports.

(4) FNT model is utilized to carry out modeling and forecasting to production run

Flow process object data is analyzed, a large-scale procedure industrial object is decomposed, (namely in industrial flow object, the change of some link or node can cause the change of other links or node to find crucial flow process and corresponding production input and output parameter.To the link of other links or node state change or node be caused as input parameter, using being caused the link of change or node as output parameter), be entered in FNT, obtain the output function of shape as formula 18, wherein A _nand B _nthe parameter after two groups of PSO optimize, net _nit is the tree structure after PIPE optimizes; This function is exactly the Changing Pattern of production procedure parameter, carries out detailed predicting for the change of producing future.

${\mathrm{out}}_n=f(A_n,B_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2}---\left(18\right)$

Fig. 2 is the modeling overall flow embodiment of FNT of the present invention, and its particular content is as follows:

(1) initial value of the parameter used in PIPE and PSO algorithm is defined in.Definition elite individuality is sky, and its adaptive value is a very large real number value.Create initial population (flexible Neural Tree and its parameter).

(2) utilize the flexible Neural Tree structure of PIPE algorithm optimization, adaptive value function adopts square error or root-mean-square error.

(3) if a good tree construction is found, jump to the 4th step, otherwise go to the 2nd step (GBL).

(4) optimization of parameter is completed by algorithm of lowering one's standard or status.At this one-phase, the structure of tree is fixing, and best tree obtains by performing PIPE algorithm.The parameter vector that the parameter of optimal tree is formed is optimized by local search algorithm.

(5) if reach the circulation maximal value of particle cluster algorithm PSO, or within the significant time, (100 circulations) does not have good parameter vector to produce, and just jumps to the 6th step; Otherwise jump to the 4th step.

(6) if find satisfied solution vector just to terminate; Otherwise jump to the 2nd step.

The above is only the preferred embodiment of the present invention; be noted that for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (8)

1. an industrial process modeling Forecasting Methodology for Process-Oriented object, is characterized in that comprising the steps:

(1) FNT model is set up, and extracts industrial flow object raw data set S from the data warehouse that flow object has generated, and create the initial population of FNT model, population at individual number customizes as required, and each individuality represents a FNT model;

(2) utilize PIPE algorithm optimization FNT model structure, adaptive value function adopts square error or root-mean-square error;

(3) Particle Swarm (PSO) algorithm optimization FNT model parameter is utilized;

(4) FNT model is utilized to carry out modeling and forecasting to flow object production run.

2. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1, it is characterized in that: in described step (1) industrial flow object raw data set S, each attribute or field represent the state of certain node in industrial flow object, its value can change along with the change of the change of time and other node states, the data of FNT model treatment need [0, 1] between, so need to be normalized raw data set S, method for normalizing is as formula (1), wherein X is pending raw data, MAX and MIN is the minimum and maximum value of data attribute in raw data set S belonging to X respectively,

Y＝(X-MIN)/(MAX-MIN) (1)

Then raw data set S later for normalization is loaded into database, forms the data warehouse that can be directly used in data mining.

3. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1 and 2, is characterized in that: the collection of functions F that described FNT model is used and termination message collection T is described below:

S＝F∪T＝{+ ₂,+ ₃,...,+ _N}∪{x ₁,...,x _n} (2)

Wherein ,+ _irepresent non-leaf nodes information, i=2,3 ..., N, i representative function+ _ithe number of corresponding input variable; x ₁, x ₂... x _nfor leaf node information; The output of a non-leaf nodes is regarded as a flexible neuron calculate, namely+ _iit is the flexible neuron with i input; In the constructive process of Neural Tree, if nonterminal information+ _iselected, the actual value of i is random generation, its expression+ _iconnection weights between this node and his child's contact; Two adjustable parameter a of flexible actuation function _iand b _ialso be random generation; The excitation function of FNT; Be expressed as:

$f=(a_i,b_{i}x)=e^{-{(x-a_i/b_i)}^2}---\left(2\right)$

Flexible neuron+ _nbeing calculated as follows of input, node+ _nexcitation and be:

${\mathrm{net}}_n=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j*x_j---\left(4\right)$

X _jrepresent the input of+n node, j=1,2 ... n.

4. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1, it is characterized in that: the search volume of described step (2) PIPE algorithm is the tree-like population produced according to the raw data set S preset, individuality results from the probability vector space covering institute's likely individuality; Individual with the generation of probability prototype tree, expression is the tree construction of a n dimension, the maximum branch number that information in n representative function collection can produce, the non-leaf nodes of tree results from collection of functions F, leaf node results from termination message collection T, the number of the subtree of each node is decided by the function information of each node producible point of number, and the input of each branch has corresponding subtree to calculate, the analysis mode of tree be depth-first from left to right.

5. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1 or 4, is characterized in that: described step (2) PIPE algorithm flow comprises:

(21) individual generation, produces individual with probability prototype tree, represent body one by one, wherein 0<j<=PS, PS represent the scale that per generation is individual;

(22) individual evaluation, each population at individual all to evaluate in given problem, and according to predefined adaptive value function formula, as shown in formula (5) and (6), calculate adaptive value the best individuality (individuality that adaptive value is minimum) of current population is marked as program performs till now, and best individuality is stored in in, Fit (i) represents i-th individual adaptive value, and p represents number of samples, y ₁ ^jand y ₂ ^jrepresent the actual sequence value of a jth sample and a jth sample final output valve through i-th individuality calculating respectively,

$\mathrm{Fit}\left(i\right)=\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2---\left(5\right)$

$\mathrm{Fit}\left(i\right)=\sqrt{\frac{1}{P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}{({y_1}^j-{y_2}^j)}^2}---\left(6\right)$

(23) individual study, in order to make the increase of the probability of current best individuality, need the probable value revising prototype tree, this process is called prototype tree and evolves, implementation procedure: the probability of first current best individuality value and produce this preferably individual all node N _jall relevant, computing formula is as follows:

$P\left(P_{\mathrm{RO}{G}_b}\right)=\underset{j:N_j}{\mathrm{\&Pi;}}{P}_j\left(I_j\left(P_{{\mathrm{ROG}}_b}\right)\right)---\left(7\right)$

represent individual at a jth node place information, individual destination probability be calculated as follows:

$P_{\mathrm{TARGET}}=P\left(P_{{\mathrm{ROG}}_b}\right)+(1-P\left(P_{{\mathrm{ROG}}_b}\right))\&CenterDot;\mathrm{lr}\frac{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{\mathrm{ROG}}^{\mathrm{el}}\right)}{\mathrm{\&epsiv;}+\mathrm{FIT}\left(P_{{\mathrm{ROG}}_B}\right)}---\left(8\right)$

Lr is a constant, represents learning rate; ε being a user-defined positive constant, according to P _tARGET, the probability of all single nodes all will repeatedly increase,

C ^lra constant, the number of times of impact circulation.C ^lrless, individual probability more close to destination probability P _tARGET, the number of times of circulation is more.Practice shows to work as c ^lrwhen=0.1, good balance can be obtained between degree of accuracy and speed.All adaptation vectors again standardized;

(24) variation of prototype tree, at node N _jplace has and produces current best individuality relevant vector dimension P _j(I) all with probability morph:

$P_{m_P}=\frac{P_M}{n\&CenterDot;\sqrt{|P_{\mathrm{RO}{G}_b}}}---\left(9\right)$

P _mbe a parameter with definition, represent overall mutation probability, n represents the number of information in information set S, represent individual nodes.All made a variation according to formula below by the probability vector selected:

P _j(I)＝P _j(I)+mr·(1-P _j(I)) (10)

Mr is another constant user-defined, represents aberration rate;

(25) prototype hedge clipper branch, after the circulation of every generation terminates, prototype tree is all by beta pruning, and the probability vector of the node of subtree has at least a probable value to be greater than pruning threshold Tp just and be cut, and Tp is a larger decimal normally;

(26) termination condition, or circulation finds satisfied solution vector until the fixed value that reaches a program appraisal is a kind of.

6. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1, it is characterized in that: the mathematical description of described step (3) particle swarm optimization is as follows: set particle populations scale as N, wherein the coordinate position vector representation of each particulate in D dimension space is velocity vector is expressed as particulate personal best particle, namely the optimal location that lives through of this particulate, is designated as colony's optimal location, the optimal location that namely in this Particle Swarm, any individual lives through, is designated as ${\stackrel{\&RightArrow;}{P}}_g=(p_{g1},p_{g2},\&CenterDot;\&CenterDot;\&CenterDot;,p_{\mathrm{gd}},\&CenterDot;\&CenterDot;\&CenterDot;,p_{\mathrm{gD}});$

The iterative formula of personal best particle is:

Colony's optimal location is position best in personal best particle, and speed and position iterative formula are respectively:

$v_{i,t+1}^d=v_{i,t}^d+c_1*\mathrm{rand}*(p_{i,t}^d-x_{i,t}^d)+c_2*\mathrm{Rand}*(p_{g,t}^d-x_{i,t}^d)---\left(12\right)$

$x_{i,t+1}^d=x_{i,t}^d+v_{i,t+1}^d---\left(13\right)$

7. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 6, is characterized in that: after described step (3) parameter optimization, FNT interior joint+ _noutput be calculated as follows:

${\mathrm{out}}_n=f(a_n,b_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2---\left(14\right)}$

And the result that whole FNT optimizes export can according to the principle of depth-first from left to right recursive calculation, as shown in formula (15):

Wherein, x ₁, x ₂..., x _nrepresent the input of the tree construction child nodes that FNT model is corresponding; w ₁, w ₂..., w _nfor the weight that limit is corresponding, for what optimize in FNT modeling process, by PSO algorithm optimization; Y is that formula (14) exports.

8. the industrial process modeling Forecasting Methodology of a kind of Process-Oriented object according to claim 1, it is characterized in that: flow object critical workflow and corresponding production input and output parameter are input in step (2) and the revised FNT model of step (3) by described step (4), obtain the output function of shape as formula 16, wherein A _nand B _nthe parameter after two groups of PSO optimize, net _nit is the tree structure after PIPE optimizes; This function is exactly the Changing Pattern of production procedure parameter, carries out detailed predicting for the change of producing future:

${\mathrm{out}}_n=f(A_n,B_n,{\mathrm{net}}_n)=e^{-{({\mathrm{net}}_n-a_n/b_n)}^2---\left(16\right)}$