CN100402708C - Real-time control method in metal electrodeposition process based on neuron networks - Google Patents

Abstract

The present invention relates to a real-time control method, particularly to a real-time control method in a metal electrodeposition process based on neuron networks. The present invention calculates each parameter quantity required to be regulated in the current depositing tank by collecting each parameter in a metal deposit tank to a system in real time by a background calculation module, and then each parameter is regulated by an output module. Consequently, the present invention comprehensively holds the multiple non-linearity relation of the concentration, the temperature, the pH value, the mixing speed and the current density of a solution in the metal deposit tank and the metal deposit quality and the various replenished solutes. In addition, the neuron network algorithm used by the present invention can randomly approaches nonlinearity, namely that as long as the number of the neuron is sufficient, the random ideal outcome can be obtained.

Description

The method of controlling in real time based on neural network in the metal electrodeposition process

Technical field

The present invention relates to a kind of method of real-time control, is the method for controlling in real time based on neural network in metal electrodeposition process specifically.

Technical background

Existing owing to parameters such as mixing solutions concentration, temperature, pH value, stirring velocity, current density in the galvanic deposition cell in the electrodeposition process constantly change, between structure that it is inner and the structure and these structures and exist quite complicated singular nonlinear relation between himself, do not exist in general or very difficultly find explicit funtcional relationship, even found funtcional relationship reluctantly, that also can be the very complicated nonlinear partial differential equation of a class, quite is difficult to hold.In fact, the Nonlinear Singular system all can not describe (they are to belong to some low-dimensional submanifolds) with the subspace, and it is the comparison difficulty that the low-dimensional submanifold directly is discussed.Existing description to this complex nonlinear relation is based on the method for linearity or non-linear regression, and this method itself just exists coarse problem.

Summary of the invention

The present invention has overcome above-mentioned shortcoming, and the real-time control method of a kind of accuracy height, stable working state is provided.

The present invention solves the technical scheme that its technical problem takes: may further comprise the steps:

(1) parameter input module is used for importing the parameters of solution in the metal electrodeposition process sedimentation tank, comprises concentration, temperature, pH value, stirring velocity, current density;

(2) interface engine module is used for to backstage computing module pass data;

(3) backstage computing module forces into by neural network model non-linear that method calculates the parameters value, each neural network input be its cum rights input value is long-pending partially with it, and each neuronic output is the RBF that network is imported;

(4) output module, the output calculated result, and operating device changes operation result into quantized control.

Described backstage computing module can adopt Matlab as the backstage computation engine.

In the computing module of described backstage, before the neural network computing, also can comprise course of standardization process.

Described interface engine module can comprise a calling module of opening up data buffer module, a data format converting module and an operation file, and above-mentioned three module order in no particular order.

Described parameter input module can comprise a parameter verification and control module, and described parameter verification comprises parameter-definition territory control module, parameter step length control module, data layout control module, and above-mentioned three sequence of modules in no particular order.

Can comprise control information and data message in the described module load module, and be responsible for the parameters of input is proofreaded.

Also can comprise the inquiry process of an input mode before the described parameter input module,, enter different load modules, comprise at least three input models in the described input mode by judging different input modes.

The present invention is by with real-time the collecting in the system of the parameters in the metal deposit tank, calculates the amount that needs the parameters of regulating in the current sedimentation tank by the backstage computing module again, by output module parameters regulated again.Therefore, the present invention's complex nonlinear relation of having held the quality of the concentration of solution, temperature, pH value, stirring velocity, current density and its metal deposition in the metal deposit tank all sidedly and having replenished various solutes.In addition, the neural network algorithm that the present invention adopts can be forced into non-linearly arbitrarily, promptly as long as nerve cell layer is abundant, just can reach the result of any desired.Present method can be widely used in fields such as forest product industry, wood-based plate, metal processing, control science, for example can adopt present method being used for making in the three-dimensional pattern wood-based plate template procedure.

Description of drawings

Fig. 1 is a control principle synoptic diagram of the present invention

Fig. 2 is a control flow chart of the present invention

Fig. 3 is the synoptic diagram of parameter input module among the present invention

Fig. 4 is the schema of interface engine module among the present invention

Fig. 5 is the synoptic diagram of course of standardization process among the present invention

Fig. 6 is the synoptic diagram of calculating process in the course of standardization process among the present invention

Fig. 7 is the synoptic diagram of neural network calculating process among the present invention

Fig. 8 is the linear regression graph of operation result Performance Evaluation of the present invention

Embodiment

As shown in Figure 1, 2, the present invention mainly is made of inquiry module, parameter input module, parameter verification module, interface engine module, standardization module, backstage computing module and the output module of input mode, at first with the parameters of solution in the metal electrodeposition groove that collects, comprise that data messages such as concentration, temperature, pH value, stirring velocity, current density are through after the inquiry of input mode, activate the relevant parameters load module, export to interface engine through totalizer through parameter identification and check and correction.In interface engine, at first open up a data buffer zone, in this district, realize the format conversion of above-mentioned parameter data, after the calling of neural network model file, just can enter the backstage computing module, promptly use Matlab as the backstage computation engine, calculate the current value that needs each parameter of adjustment with neural network model, give the first layer of neural network through the data transfer of preliminary treatment, need standardization at this one deck, calculate Euclidean distance after determining through neural network is preliminary again, by the weights function of generation standard after the correction of self learning model, this function will be determined strength of solution at last, temperature, the pH value, stirring velocity, the weights of the influence of current density.The second layer at neural network, carry out the linear function that normalization process produces weights after the first layer weights function is by the correction of self learning neural networks layer weights module with receiving, reach the ideal output of various solute requirements at last, therefore, as long as nerve cell layer is abundant, just can reach the result of arbitrary accuracy.By control corresponding equipment the operation result that obtains is converted into quantified controlling to parameters in the galvanic deposition cell again, thereby finishes real-time monitoring and control whole metal electrodeposition process.For example, when solution temperature reduces, collect current temperature value by Temperature sampler, after process aforementioned calculation process draws the temperature value that needs to adjust, by temperature regulator solution is heated, when the proportioning of certain solute in the solution reduces, collect the concentration of current this kind solute by the concentration collector, draw the value that needs this additional solute through the aforementioned calculation process, finish replenishing by liquid replenishment control device at last this kind solution.Because the process of this detection and control is carried out in real time, thereby guarantee the continual and steady of parameters in the electric depositing solution, also just guaranteed the quality of metal electrodeposition.

The inquiry module of described input mode, as shown in Figure 2, by inquiry and judgement to input type parameter i, judge different input modes, enter the load module of three kinds of different models again, be model I load module, model II load module and model III load module,, then enter parameter input module through after this inquiry module and after activating corresponding load module.

Described parameter input module, as shown in Figure 3, be control information, the data message that data collection station is collected, the image data that comprises transmitters such as the interior concentration of sedimentation tank, temperature, current density, pH value, by this module and interactive interfacing data message and control information associated with the data, promptly carry out the control of parameter-definition territory, parameter step length control, data layout control.Wherein the control of parameter-definition territory is mainly used in the correction parameter input, error handling processing, and the initial setting up of each field of definition is determined by the target sample data, but its parameters needed interval of custom-made before the generating unit work; Parameter step length control is used to change the step-length of parameter trim button, and is easy to use; Data layout control is to select several position effective digitals and which kind of science method of counting when the input data.Control information is transparent with data message communicating by letter of load module inside, can freely select and distinguish input control information and data message, and the two keeps synchronously forever.

Described interface engine module, as shown in Figure 4, because what adopt in the present embodiment is the special mathematical operation software Matlab instrument as its backstage computing and analysis, therefore at first will be provided with one is used for the module that front end is communicated by letter with backstage Matlab, i.e. Matlab interface engine.Particularly, the interface of parameter information by Matlab and interface development program with input is sent to the Matlab working space exactly, participates in calculating, and can return operation result with the suitable data form after the computing end, simultaneously, the chart that provides a series of network performances and training result to analyze.The interface engine module comprises a calling module of opening up data buffer module, a data format converting module and an operation file, and above-mentioned three module order in no particular order.The main flow process of Matlab interface engine comprises: open Matlab engine, supplemental characteristic format conversion, parameter by user interface opening up to the transmission of backstage Matlab and suitable big or small buffer zone.

Described backstage computing module also comprises standardization module and neural network computing module, promptly before entering real operational network, the input sampled data will be carried out in advance, standardization afterwards, processing mode as shown in Figure 5, standardization is carried out in codomain interval for the input sampled data, is 1 standard vector by after the standardization input vector and target output vector being quantified as zero-mean and deviation.Be implementation procedure below by zero-mean and deviation type function.

[pn，meanp，stdp，tn，meant，stdt]＝prestd(p，t)

Parameter meaning: p network input vector

T target output vector

Input vector after pn quantizes

The average of meanp input vector

The deviation of stdp input vector

Target output vector after tn quantizes

The average of meant target output

The deviation of stdt target output

Concrete standardisation process as shown in Figure 6.Provide the operation link of standardization device below:

$\mathrm{mean}\left(P^{\&prime;}\right)=\mathrm{mean}\left(\left[\begin{array}{cccc}{p}_{11}& p_{21}& ...& p_{R1}\\ p_{12}& p_{22}& ...& {\mathrm{<figure-callout\; id="2"\; label="p\; R"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">p</figure-callout>}}_R\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ p_{1Q}& p_{2Q}& ...& p_{\mathrm{RQ}}\end{array}\right]\right)$

$={\left[\begin{array}{cccc}\frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{p}_{1i}}{Q}& \frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{p}_{2i}}{Q}& ...& \frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{p}_{\mathrm{Ri}}}{\mathrm{<figure-callout\; id="1"\; label="Q"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">Q</figure-callout>}}\end{array}\right]}_{1\&times;R}$

$\mathrm{std}\left(P^{\&prime;}\right)=\mathrm{std}\left(\left[\begin{array}{cccc}{p}_{11}& p_{21}& ...& p_{R1}\\ p_{12}& p_{22}& ...& {\mathrm{<figure-callout\; id="2"\; label="p\; R"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">p</figure-callout>}}_R\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ p_{1Q}& p_{2Q}& ...& p_{\mathrm{RQ}}\end{array}\right]\right)$

$={[{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(p_{1i}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{p}_{1i})}^2\right)}^{\frac{1}{2}}{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(p_{2i}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{p}_{2i})}^2\right)}^{\frac{1}{2}}...{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(p_{\mathrm{Ri}}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{p}_{\mathrm{Ri}})}^2\right)}^{\frac{1}{2}}]}_{1\&times;R}$

$P_n=(P-\mathrm{meanp}\&times;\mathrm{oneQ})\&CenterDot;/(\mathrm{stdp}\&times;\mathrm{oneQ})$

$\mathrm{mean}\left(T^{\&prime;}\right)=\mathrm{mean}\left(\left[\begin{array}{cccc}{t}_{11}& t_{21}& ...& t_{S1}\\ t_{12}& t_{22}& <figure-callout\; id="2"\; label="...\; t\; S"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">.</figure-callout>..& t_S{2}_{}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ t_{1Q}& t_{2Q}& ...& t_{\mathrm{SQ}}\end{array}\right]\right)$

$={\left[\begin{array}{cccc}\frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{t}_{1i}}{Q}& \frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{t}_{2i}}{Q}& ...& \frac{\underset{i=1}{\overset{i=Q}{\mathrm{\&Sigma;}}}{t}_{\mathrm{Si}}}{\mathrm{<figure-callout\; id="1"\; label="Q"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">Q</figure-callout>}}\end{array}\right]}_{1\&times;S}.$

$\mathrm{std}\left(T^{\&prime;}\right)=\mathrm{std}\left(\left[\begin{array}{cccc}{t}_{11}& t_{21}& ...& t_{S1}\\ t_{12}& t_{22}& <figure-callout\; id="2"\; label="...\; t\; S"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">.</figure-callout>..& t_S{2}_{}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ t_{1Q}& t_{2Q}& ...& t_{\mathrm{SQ}}\end{array}\right]\right)$

$={[{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(t_{1i}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{t}_{1i})}^2\right)}^{\frac{1}{2}}{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(t_{2i}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{t}_{2i})}^2\right)}^{\frac{1}{2}}...{\left(\frac{1}{Q-1}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{(t_{\mathrm{Si}}-\frac{1}{Q}\underset{i=1}{\overset{Q}{\mathrm{\&Sigma;}}}{t}_{\mathrm{Si}})}^2\right)}^{\frac{1}{2}}]}_{1\&times;S}$

$T_n=(T-\mathrm{meant}\&times;\mathrm{oneQ})\&CenterDot;/(\mathrm{stdt}\&times;\mathrm{oneQ})$

Behind the standardization device, enter formal neural network computing module, as shown in Figure 7:

With reference to Fig. 7 neural network computing module:

Q node arranged in this neural network module, and R ties up input, and S ties up output.The neural network input vector is P, and the weights of neural network hidden layer are made as P ', and the cum rights input of each neuron node of this layer is the distance between input vector and the weight vector, i.e. Euclidean distance ‖ dist ‖.‖ dist ‖ is the Euclidean distance weight function, and weight function is added to weight on the input matrix to obtain the cum rights input matrix.

For dist (W, P), W is S * R weight matrix, P is a Q dimension input column vector matrix, (W P) returns S * Q dimensional vector distance matrix to dist.

In this model, weight matrix W is defined as P ', then

$\mathrm{dist}({P^{\&prime;}}_{Q\&times;R},P_{R\&times;Q})$

$=\mathrm{dist}(\left[\begin{array}{cccc}{p}_{11}& p_{12}& ...& p_{1R}\\ p_{21}& p_{22}& ...& p_{2R}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ p_{Q1}& {\mathrm{<figure-callout\; id="2"\; label="p\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">p</figure-callout>}}_Q& ...& p_{\mathrm{QR}}\end{array}\right],\left[\begin{array}{cccc}{p}_{11}& p_{21}& ...& p_{Q1}\\ p_{12}& p_{22}& ...& {\mathrm{<figure-callout\; id="2"\; label="p\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">p</figure-callout>}}_Q\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ p_{1R}& p_{2R}& ...& p_{\mathrm{QR}}\end{array}\right])$

$={\left[\begin{array}{ccccc}0& d_{12}& d_{13}& ...& d_{1Q}\\ d_{21}& 0& d_{23}& ...& d_{2Q}\\ d_{31}& d_{32}& 0& ...& d_{3Q}\\ .& .& .& & .\\ .& .& .& & .\\ .& .& .& & .\\ d_{Q1}& d_{Q2}& {\mathrm{<figure-callout\; id="3"\; label="d\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">d</figure-callout>}}_Q& ...& 0\end{array}\right]}_{Q\&times;Q}$

D in the formula _IjDistance between i row vector of representing matrix P ' and j column vector of matrix P, so the element on the diagonal lines is 0.

And then with dist (P ', P) and b ¹Make dot product, promptly

$\mathrm{dist}(P^{\&prime;},P)\&CenterDot;*{\mathrm{<figure-callout\; id="1"\; label="b"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">b</figure-callout>}}_1=\left[\begin{array}{ccccc}0& d_{12}& d_{13}& ...& d_{1Q}\\ d_{21}& 0& d_{23}& ...& d_{2Q}\\ d_{31}& d_{32}& 0& ...& d_{3Q}\\ .& .& .& & .\\ .& .& .& & .\\ .& .& .& & .\\ d_{Q1}& d_{Q2}& {\mathrm{<figure-callout\; id="3"\; label="d\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">d</figure-callout>}}_Q& ...& 0\end{array}\right]\&CenterDot;*\left[\begin{array}{cccc}{b}_{11}& b_{12}& ...& b_{1Q}\\ b_{21}& b_{22}& ...& b_{2Q}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ b_{Q1}& {\mathrm{<figure-callout\; id="2"\; label="b\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">b</figure-callout>}}_Q& ...& b_{\mathrm{QQ}}\end{array}\right]$

$=\left[\begin{array}{ccccc}0& b_{12}*d_{12}& b_{13}*d_{13}& ...& b_{1Q}*d_{1Q}\\ b_{21}*d_{21}& 0& b_{23}*d_{23}& ...& b_{2Q}*d_{2Q}\\ b_{31}*d_{31}& b_{32}*d_{32}& 0& ...& b_{3Q}*d_{3Q}\\ .& .& .& .& \\ .& .& .& .& \\ .& .& .& .& \\ {\mathrm{<figure-callout\; id="1"\; label="b\; Q"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">b</figure-callout>}}_Q*d_{Q1}& {\mathrm{<figure-callout\; id="2"\; label="b\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">b</figure-callout>}}_Q*d_{Q2}& {\mathrm{<figure-callout\; id="3"\; label="b\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">b</figure-callout>}}_Q*{\mathrm{<figure-callout\; id="3"\; label="d\; Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">d</figure-callout>}}_Q& ...& 0\end{array}\right]$

In each the neural network input of network hidden layer is that its cum rights input is long-pending with its inclined to one side value, and shown in following formula, and each neuronic output is the RBF of network input.

Utilize gaussian kernel function (Gaussian kernel function), to be shown below as the form of basis funciton:

$u_j=\exp [-\frac{{(X-C_j)}^T(X-C_j)}{2{{\mathrm{\&delta;}}_j}^2}],j=\mathrm{1,2},...,N_h$

Wherein, u _jBe the output of j hidden node, X=(x ₁, x ₂..., x _n) ^TBe the input sample, C _jBe the central value of Gaussian function, δ _jBe generalized constant, N _hIt is the number of hidden nodes.Action function in its hidden layer node (kernel function) will produce response in the part to input signal, that is to say, when input signal during near the central range of kernel function, hidden node will produce bigger output, thus, this neural network has local approximation capability, so radial primary function network also becomes local perception field network.By following formula as can be known, the output area of node is between 0 and 1, if a neuronic weight vector equates (transposition) with its input vector, its cum rights input will be 0, when its network is input as 0, then be output as 1, and more near the center of node, output valve is bigger for the input sample.

Adopt the gaussian basis function, possess following advantage:

1, representation is simple, even input does not increase too many complicacy yet for multiparameter;

2, radial symmetry;

3, slipperiness is good, and order derivative exists arbitrarily;

4,, thereby be convenient to carry out theoretical analysis because this basis funciton is represented simple and analyticity is good

Through the basis funciton producer, neuron node is output as a ¹, begin to enter the network linear layer.At first will be in the network linear layer through a normalization process device, and then enter common linear neuron.In the normalization process device, adopt the normprod function to come the output vector n of computational grid ²Normprod is a weight function, and weight function is added to weight and obtains the cum rights matrix on the input matrix.For normprod (W, P), W is S * R weight matrix, P is a Q dimension input column vector matrix, (W P) returns S * Q and ties up regular dot product normprod.

In this network, the weight matrix of network linear layer is made as the target output T of network _{S * Q}, promptly

$\mathrm{normprod}(T,a^1)$

$=\mathrm{normprod}\left(\left[\begin{array}{cccc}{t}_{11}& t_{12}& ...& t_{1Q}\\ t_{21}& t_{22}& ...& t_{2Q}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ t_{S1}& {\mathrm{<figure-callout\; id="2"\; label="t\; S"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">t</figure-callout>}}_S& ...& t_{\mathrm{SQ}}\end{array}\right]\left[\begin{array}{cccc}{a}_{11}& a_{12}& ...& a_{1Q}\\ a_{21}& a_{22}& ...& a_{2Q}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{Q1}& a_{\mathrm{<figure-callout\; id="2"\; label="Q"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">Q</figure-callout>}2}& ...& a_{\mathrm{QQ}}\end{array}\right]\right)$

$=\left[\begin{array}{cccc}\frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{1j}*a_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{1j}*a_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}& \mathrm{\&Lambda;}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{1j}*a_{\mathrm{jQ}}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{jQ}}}\\ \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{2j}*a_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{2j}*a_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}& \mathrm{\&Lambda;}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{2j}*a_{\mathrm{jQ}}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{jQ}}}\\ M& M& & M\\ \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{\mathrm{Sj}}*a_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="1"\; label="j"\; filenames="C20051006649500201.png,C20051006649500241.png"\; state="\{\{state\}\}">j</figure-callout>}1}}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{\mathrm{Sj}}*a_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051006649500201.png"\; state="\{\{state\}\}">j</figure-callout>}2}}& \mathrm{\&Lambda;}& \frac{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{t}_{\mathrm{Sj}}*a_{\mathrm{jQ}}}{\underset{j=1}{\overset{j=Q}{\mathrm{\&Sigma;}}}{a}_{\mathrm{jQ}}}\end{array}\right]$

Neural network computing of the present invention is one and approaches device, as long as hidden unit is abundant, it just can approach the first continuous function of any M and to the nonlinear function of arbitrary the unknown, always exist one group of weights make network to this function to approach effect best.The network second layer also has and the network input neuron node number identical with object vector, and the weight matrix with the second layer is made as the object vector matrix T here.

Described network output module, output module as shown in fig. 1: after network training finished, the output that comes the emulation neural network with the sim function compared thereby export with target, checks the performance of neural network.The relation that neural network is exported and target is exported that function postreg has utilized linear regressive methods analyst, i.e. neural network output changes the velocity of variation with respect to target output variation, thereby has assessed the training result of neural network.

a＝sim(net，p)

[m，b，r]＝postreg(a，t)

Function postreg has returned 3 values, and m and b represent optimum regression collinear slope and y y-intercept respectively, and when m equals 1, when b equaled 0, neural network output and target output were identical, and the neural network of this moment has the performance of optimum.R represents the relation conefficient that network output and target are exported, and it approaches 1 more, and the output of expression network is approaching more with target output, and the neural network performance is good more.In the figure that function postreg shows, X-coordinate is target output, and ordinate zou is network output, "." the expression data, ideal regression straight line (straight line when neural network output equals target output) is represented that by solid line the optimum regression straight line is illustrated by the broken lines.

In the neural network output module, what provide each output wood property index respectively is used for weighing neural network performance linear regression graph shape, output pattern as shown in Figure 8, its precision has reached 98.999% as can be seen from analogous diagram.

Result with every computing passes through control corresponding equipment more at last, as temperature regulator, liquid make-up controller etc., solution in the galvanic deposition cell is carried out control corresponding and adjusting, thereby make the parameters of the solution in the metal deposit tank keep stable, guarantee the quality of metal electrodeposition.

Claims (7)

1. the method for controlling in real time based on neural network in the metal electrodeposition process is characterized in that: may further comprise the steps:

(1) parameter input module is used for importing the parameters of solution in the metal electrodeposition process sedimentation tank, comprises concentration, temperature, pH value, stirring velocity, current density;

(2) interface engine module is used for to backstage computing module pass data;

(3) backstage computing module forces into by neural network model non-linear that method calculates the parameters value, each neural network input be its cum rights input value is long-pending partially with it, and each neuronic output is the RBF that network is imported;

(4) output module, the output calculated result, and operating device changes operation result into quantized control.

2. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 is characterized in that: described backstage computing module adopts Matlab as the backstage computation engine.

3. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 and 2 is characterized in that: in the computing module of described backstage, also comprised course of standardization process before the neural network computing.

4. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 and 2, it is characterized in that: described interface engine module comprises a calling module of opening up data buffer module, a data format converting module and an operation file, and above-mentioned three module order in no particular order.

5. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 and 2, it is characterized in that: described parameter input module comprises a parameter verification and control module, described parameter verification comprises parameter-definition territory control module, parameter step length control module, data layout control module, and above-mentioned three sequence of modules in no particular order.

6. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 and 2 is characterized in that: comprise control information and data message in the described parameter input module, and be responsible for the parameters of input is proofreaded.

7. the method for controlling in real time based on neural network in the metal electrodeposition process according to claim 1 and 2, it is characterized in that: the inquiry process that also comprises an input mode before the described parameter input module, by different input modes, enter different load modules, comprise at least three input models in the described input mode.

Images