CN105290122A - AGC system thickness measurement device supported by RBF network - Google Patents

Abstract

The invention provides an AGC system thickness measurement device supported by a RBF network. The device is characterized by consisting of a RBF module, an AGC system module, a TDC module, a thickness meter and a learning single-frame mill. The RBF module consists of three neutral network modules for respectively accommodating three layers of artificial neutral networks; the three neutral network modules consist of a neutral module A, a neutral module B and a neutral module C; and the AGC system comprises a RBF interaction module, a control interaction module and a calculation module.

Description

A kind of AGC system measurer for thickness of RBF network support

Technical field

The present invention relates generally to the thickness measure of single stand cold mill steel, cold rolling aluminium, cold rolling Zn system, be specifically related in AGC system thickness measure, utilize RBF network to carry out the application supported.

Background technology

The precision of thickness is the most important technical indicators of all metal rolled industry products, and the control of thickness and precision is the technical problem of the key in rolling field.Affect thickness because have the reason of three aspects: 1. the factor of rolled piece, it comprises the thickness and hardness etc. of supplied materials, and this disturbance is random, and it depends on the precision of a procedure.2. the factor of milling train, it comprises roll eccentricities, degree of regulation etc., and the adjustment of this factor to thickness and precision serves conclusive effect.

3. the factor of technique, it comprises roll gap and the lubrication wet goods of roll welding, and the concrete quantitative analysis of this disturbance is the difficult point of research always, does not also set up Mathematical Modeling accurately so far, can only ensure the perfect performance of technique in production process by rule of thumb.

At present, five kinds of THICKNESS CONTROL modes are mainly contained to the adjustment of thickness.

1) feed forward type THICKNESS CONTROL.It precomputes supplied materials thickness deviation by last frame or entrance calibrator, and be fed to next frame forward, adjusts screw-down in advance within the predetermined time, to ensure the precision of target thickness.Because feed-forward control belongs to opened loop control, be not generally used alone in actual applications, be all combine with follow-up THICKNESS CONTROL mode together with use.

2) monitoring THICKNESS CONTROL.After band steel, aluminium strip, zinc band shut out from milling train, measure the actual of rolled piece by the calibrator being arranged on milling train outlet side and shut out thickness, and the thickness recorded and set-point are compared, obtain thickness deviation, through computing, thickness signal is converted to roll gap adjustment signal again, exports to hydraulic mechanism to reach the object eliminating thick difference.In actual use, the damage of calibrator is caused owing to considering the restriction of rolling mill structure, the not ease for maintenance of calibrator and preventing strip from rupturing, generally thickness gauge is contained in from the distant position of roll gap, cause the hysteresis quality of thickness measuring like this, for dead-time system, present technical field does not provide reasonable solution yet.

3) tension force THICKNESS CONTROL.It is the relation intercoupled that tension force controls with THICKNESS CONTROL, and the change of tension force can change roll-force significantly, thus can change the thickness shut out.It is the thickness deviation measured according to rack outlet side calibrator, finely tunes forward pull and backward pull, eliminates thickness deviation whereby.Tension force fine setting has two kinds of modes, and one regulates gap values between rollers, and a kind of is the speed regulating milling train, and the latter is conventional mode.But in order to ensure the stability of the operation of rolling, general tension force THICKNESS CONTROL, only for regulating little thickness deviation, uses as accurate adjustment.

4) flow THICKNESS CONTROL.It is according to the equal principle of metal mass flow, utilizes the velocity variations of each frame to eliminate thickness deviation, because this technology only relates to single chassis THICKNESS CONTROL, therefore with less than flow THICKNESS CONTROL.

5) thickness gauge type THICKNESS CONTROL.Here " thickness gauge " is a kind of thickness gauge of simulation, and it is by recording roll-force and unloaded roll gap, and utilize spring equation to calculate the actual of this moment and shut out one-tenth-value thickness 1/10, now whole frame achieves the function of thickness gauge.It can overcome delayed deficiency detection time of calibrator THICKNESS CONTROL, can detect that roll-force changes the strip varied in thickness caused, to realize quick adjustment immediately.This control mode is the automatic thickness control mode of existing core.

The present invention is on the basis of above-mentioned Lung biopsy, and the data utilizing thickness gauge AGC to calculate introduce the algorithm of RBF neural, establishes the RBF artificial neural network of three layers, and dynamic self-adapting in addition, finally realize fast on hardware TDC.

The intelligent method of RBF neural and dynamic self-adapting algorithm combines with gaugemeter AGC and determines accurate thickness deviation by the present invention first, and realizes on TDC.The Mathematical Modeling of THICKNESS CONTROL can be optimized by this technology further, improve the precision of THICKNESS CONTROL, have than the better simplicity of simple use neural network algorithm and study property more simultaneously.

About the existing techniques and methods of strip THICKNESS CONTROL can with reference to following document:

[1]YoshikazuMToshioI.ModemizationofgaugecontrolsystematSumitomoWakayama5standcoldmill.IronandSteelEngineer.1999

[2]PortmannNF.LindhoffD.Applicationofneuralnetworksinrollingmillautomation.1995(02)

[3]LuisEZarate.NeuralNetworksandFuzzyRulesBasedControlForColdRollingProcessviaSensitivityFactors.2001

[4] Wang Li, Ge Ping, Sun Yikang. based on the cold continuous rolling strip flatness and gauge multivariable Control of fuzzy RBF neural network. University of Science & Technology, Beijing's journal.2002(5)

[5]Chernshenglin，Yunlonglay，Chienwaho，AlbertChinyhhlin，Ninchunehang.Anovelexperimentaldevicewithmodifiedlasershadowspotandopticalstraingaugeset-up.Measurement.2005，37：9-19.

Summary of the invention

In order to overcome above-mentioned defect, the invention provides a kind of RBF network support AGC system measurer for thickness, it is characterized in that:

It is by RBF module, AGC system module, TDC module, thickness gauge, and study singlestandmill forms.Three neural network modules of described RBF module three-layer artificial neural network by splendid attire respectively form, and are respectively neural modules A, neural module B, neural module C; Described AGC system comprises RBF interactive module, controls interactive module, computing module.

Aforementioned RBF module, AGC system, TDC module, thickness gauge are all electrically connected with computer.Aforementioned TDC module comprises, TDCRACK-UR5213, CPU module-CPU551, input/output module-SM500, MPI communication module-CP50M0; Study singlestandmill has several detectors, these several detectors are all electrically connected with RBF module.

Described thickness gauge comprises, clamper, ultrasonic transmitter, ultrasonic receiver, center module, computing module, support.Described thickness gauge appearance has shell coated, and aforementioned ultrasonic wave launcher, ultrasonic receiver side are set up in parallel, and Displacement Meters of exerting pressure is positioned at the side of pressure applicator displacement, and two distance between plates measurement mechanisms are positioned at the side of sample.

Computer is furnished with display unit and display floater, input equipment; Power module, it is for giving aforementioned RBF module, and AGC system module, TDC module, thickness gauge, study singlestandmill is all electrically connected with supply of electrical energy.Aforementioned thicknesses meter can seal setting, and be wherein furnished with thermostat and make its inner maintenance constant temperature, thickness gauge inside has insulation lining.

Utilize aforesaid measurer for thickness to carry out the method for thickness measure, it is characterized in that: the model determining thickness gauge AGC system in AGC system module, and emulate with simulink; Detection obtains the thickness data of the input and output of study singlestandmill; RBF neural is established, utilize RBF neural thickness data that the theory of study commonly uses the input and output of singlestandmill in RBF module; Utilize above-mentioned data, training adjustment RBF neural; The AGC system model of the type that is improved in AGC system module, and the adjustment of dynamic self-adapting has been carried out to relevant data; Carried out the realization of thickness measure scheme by TDC module, above carry out quick realization.In AGC system module, determine the model of thickness gauge AGC system, its concrete operations are as follows: set up the Mathematical Modeling comprising fluid-percussion model of isolated, rolling force model, mill spring model milling train.

Set up the step of RBF neural, corresponding neural modules A, neural module B, neural module C set up the RBF neural of three layers respectively, use Gaussian function as RBF, ground floor neuron is input neuron, comprise roll gap deviation, forward pull deviation, backward pull deviation, second layer neuron is hidden neuron, comprises at least 4 hidden neurons, third layer neuron is output neuron, is rolling force deviation.

The step of described training adjustment RBF neural comprises: input roll gap deviation, forward pull deviation, backward pull deviation, provide rolling force deviation, simulation result according to rolling force deviation will as inputoutput data as training data, feedback adjustment RBF neural.The study of described RBF neural is based on closest clustering algorithm, does not need the number determining hidden layer unit in advance.

Described TDC module includes TDC frame-UR5213, CPU module-CPU551, input/output module-SM500, MPI communication module-CP50M0.Described ultrasonic transmitter is high-frequency band pass ultrasonic wave transmitting probe.

Determine the model of traditional thickness gauge AGC system, and emulate with simulink; Empirical equation is utilized to obtain thickness, tension force, the gap values between rollers data of the input and output of thickness AGC system singlestandmill.Establish RBF neural, utilize RBF neural to learn the input and output of legacy system.Set up the RBF neural of three layers, use Gaussian function as RBF, ground floor neuron is input neuron, comprise roll gap deviation, forward pull deviation, backward pull deviation, second layer neuron is hidden neuron, comprise at least 4 hidden neurons, third layer neuron is output neuron, is rolling force deviation.

According to the inputoutput data that formula obtains, training adjustment RBF neural.

The step of training adjustment RBF neural comprises: input roll gap deviation, forward pull deviation, backward pull deviation, provide rolling force deviation, the simulation result according to rolling force deviation will as inputoutput data as training data, feedback adjustment RBF neural.

The AGC system model of the type that is improved, and the adjustment of dynamic self-adapting has been carried out to relevant data.

Dynamic self-adapting algorithm is based on closest clustering algorithm, and this algorithm is a kind of online adaptive cluster learning algorithm, does not need the number determining hidden layer unit in advance.

According to a further aspect in the invention, provide the quick realization of RBF neural thickness gauge AGC system on Siemens TDC, can comprise: the hardware configuration of milling train: TDCRACK (frame)-UR5213, CPU module-CPU551, defeated people's output module-SM500, MPI communication module-CP50M0; The realization of RBF neural control algolithm on TDC: inline diagnosis module, RBF module and normalization module

The thickness gauge AGC system that technology according to the present invention is determined can reflect the actual conditions of real single chassis strip, aluminium strip, zinc band, can solve the problem that exit thickness precision is not high, improves study property again of data, can realize fast on hardware.

Accompanying drawing explanation

Accompanying drawing describes embodiments of the present invention, wherein:

Fig. 1 is the block diagram that the dynamic self-adapting of RBF neural AGC system realizes;

Fig. 2 is the Mathematical Modeling of the roll-force that Simulink sets up;

Fig. 3 is RBF neural exemplary plot;

Thickness fluctuation figure when Fig. 4 is RBF THICKNESS CONTROL;

Fig. 5 is the RBF network neural unit in CFC.

Detailed description of the invention

Below with reference to the accompanying drawings example according to the embodiment concrete for the thickness control system improved by RBF neural of the present invention is described.

Fig. 1 shows a kind of AGC system based on RBF neural dynamic self-adapting and realizes block diagram.

As shown in Figure 1, can comprise the steps: according to the RBF neural self-adaptative adjustment AGC system of this embodiment of the invention the model determining traditional thickness gauge AGC system, and emulate with simulink; Empirical equation is utilized to obtain the thickness data of the input and output of thickness AGC system singlestandmill; Establish RBF neural, utilize RBF neural to learn the input and output of legacy system; According to the inputoutput data that formula obtains, training adjustment RBF neural; The AGC system model of the type that is improved, and the adjustment of dynamic self-adapting has been carried out to relevant data; Hardware TDC has carried out quick realization.

Below each step and the details that wherein relates to are described in detail.

(1) model of traditional thickness gauge AGC system, is determined.Gaugemeter AGC be present most enterprises THICKNESS CONTROL mode, be also topmost mode.

Elastoplasticity curve when the most basic principle of gauge automatic control is rolling, band steel actual shuts out relation between thickness h, actual roll gap S and actual roll-force P as shown in spring equation:

$h=S+\frac{P-P_0}{M}$

H is the exit thickness of milling train, and unit is mm; S is actual mill roll-gap value, and unit is mm; P is roll-force, P ₀for unloaded roll-force, unit is KN; M is mill stiffness, and unit is KN/mm.

When roll-force by P change to P ' time, the converted quantity δ P=P-P ' of roll-force, from spring equation, exit thickness variable quantity δ h=δ P/M.In order to eliminate this thick poor δ h, then adjustment amount of roll gap δ S should be:

$\mathrm{\δS}=\frac{M+Q}{M}\mathrm{\δh}=\frac{M+Q}{M^2}\mathrm{\δP}$

Q is the plastic coefficient of rolled piece.

Thickness gauge AGC utilizes roll gap and roll-force increment signal, according to mill spring equation estimated thickness deviation, then considers rolling mill screwdown efficiency, finally carries out corresponding adjustment to eliminate thickness deviation to roll gap.When thickness gauge AGC starts to drop into, roll gap S and roll-force P is sampled some time averaging, obtain locking roll gap S _lwith locking roll-force P _l, thus locking thickness h can be calculated _l.

Rolling force model.SIMS (Sims) formula based on Ao Luowan (Orowan) sex change field forces balance theory is best suited for the theoretical formula of hot strip rolling rolling force model, and its citation form is as follows:

P＝Bl′ _cQ _pKK _T

P is roll-force, unit K N;

B be rolled piece rolling forward backward averaging width (spread minimum during general rolling strip, negligible, directly can dedicated bandwidth replace), unit m;

L ' _cfor considering the floor projection of the roll after flattening and rolled piece contact arc, unit mm;

Q _pfor considering external friction (stress state) influence coefficient on contact arc;

K is the physical condition being decided by metal material chemical and distortion--the flow of metal resistance of deformation temperature, deformation velocity and deformation extent, and K=β σ, wherein β generally gets 1.15, units MPa;

K _tfor front and back tensile stress is to the influence coefficient of roll-force.

Wherein bandwidth B is multiplied by contact arc floor projection length l ' _cfor contact area, be the geometrical factor determining roll-force, influence coefficient Q _pand K _tfor determining the mechanics factor of roll-force, K is then the physics and chemistry factor of influence of rolled power.

Contact arc floor projection length l ' after flattening _c.Substantially can not change for single chassis bandwidth, therefore determine that contact area mainly determines l _cwith l ' _c.

Contact arc floor projection length l _cfor:

$l_c=\sqrt{\mathrm{R\Δh}}=\sqrt{R(h_0-h_1)}$

When roll is subject to very large roll-force, roll will be crushed, if the roller radius after flattening is R ', contact arc floor projection length will become l ' _c, then

${l\′}_c=\sqrt{R\′\mathrm{\Δh}}$

The influence coefficient Q of contact arc external friction _p.Calculate under different rolling conditions Sims formula, the data of gained carry out regression analysis, wherein adopt Q _p=f (l _c/ h _m, ε) structure best, the coefficient correlation of recurrence is all higher

Q _pcomputing formula

The present invention adopts the form of the correlation analysis formula 2 of Sims formula that is:

$Q_p=0.8049+0.2488\frac{l_c}{h_m}+0.0393\mathrm{\ϵ}\frac{l_c}{h_m}-0.3393\mathrm{\ϵ}+0.0732{\mathrm{\ϵ}}^2\frac{l_c}{h_m}$

In addition, when rolling force model is used for concrete milling train, can be revised by inlet coefficient K.Survey rolling force value in a large number according to scene and the milling train that statistical computation makes the result energy load reality of tube rolling simulation is carried out to equation coefficients K.

Hot-rolling metal plastic deformation resistance σ.Deformation drag is a very important physical parameter in tube rolling simulation formula, up to now, all adopts following functional form: σ=f (T, u, ε) in deformation drag research

Along with computer controls and the development of Mathematical Modeling, occurred a collection of employing metal plastic deformation resistance formula, its main version has: σ=exp (a+bT) u ^(c+dT) e ⁿwith σ=exp (a+bT) u ^(c+dT)k _e.

T=(t+273)/1000, T represents deformation temperature, unit K; T is also deformation temperature, unit DEG C; U is deformation velocity, unit s ^-1; E is deformation extent; k _efor deformation extent influence coefficient; A, b, c, d, n are regression coefficient, and variety classes respectively has a set of coefficient.Fig. 2 gives the simulation model of roll-force simulink.

(2) utilize empirical equation to obtain the thickness data of the input and output of thickness AGC system singlestandmill, forward pull deviation, backward pull deviation, roll gap deviation and rolling force deviation can be obtained.Selected breast roller radius is 400mm; Milling train equivalent stiffness is 4700kN/mm; The wide 875mm of supplied materials parameter plate; Roll linear velocity is 10m/s; Supplied materials setting thickness is 7.0mm; Outlet setting thickness is 4.2mm.Can obtain as next group data:

Forward pull deviation	Backward pull deviation	Roll gap deviation	Rolling force deviation
				N/mm	N/mm	mm	N
-0.164872	0.0247795	0.0057	860.46251
				-0.425712	0.0488237	0.0051314	1869.45455
-0.558822	0.0825803	0.004231	3328.72259
				-0.83812	0.1216657	0.0031029	5084.40081
0.5052429	-0.261044	-0.005516	-2141.9685
				0.7659498	-0.298451	-0.003623	-3798.0227
1.0806408	-0.342001	-0.001301	-5804.9919

(3) establish RBF neural, utilize RBF neural to learn the input and output of legacy system.The rule of thumb inputoutput data of simulation that obtains of formula, for obtaining more accurate model and can realizing on hardware, the present invention utilizes radial basis RBF neural to have learned this traditional AGC system.

Artificial neural network is by the basis to the long research of human brain, manually realizes the network structure of its basic 26S Proteasome Structure and Function.Basic RBF neural is three layers of feedforward network with single hidden layer, and every one deck of network has diverse effect, and ground floor is input layer, and be made up of some perception unit, network and external environment couple together by they; The second layer is an only hidden layer in network, its effect be from the input space to hidden layer space carry out nonlinear transformation; Third layer is output layer, and it is linear, for the enable mode acting on input layer provides response.

For three layers of RBF neural, the inventive method is described below: set up 3 layers of RBF neural.The formation of RBF radial primary function network comprises three layers, and wherein every one deck has diverse effect.Input layer is made up of some source points (perception unit), and network and external environment couple together by they, and the second layer is an only hidden layer in network, and its effect carries out nonlinear transformation between the input space to concealed space; In most of the cases concealed space has higher dimension.Output layer is linear, and it is that the enable mode (signal) acting on input layer provides response.Fig. 3 is RBF neural exemplary plot, and the RBF neural that the present invention uses is described.

The feature of neutral net is as follows: a parallel and distributed Information Processing Network structure, this network structure is generally made up of multiple neuron, each neuron is by a single output, it can be connected to other neuron a lot, its input has multiple connecting path, the corresponding link weight coefficients of each connecting path.Neutral net is the one simulation of biological neural network and is similar to.

Model has been set up, and first uses test data test network to reach the precision of needs, using 800 groups of test datas as input, observes the exit thickness of mill model output and the exit thickness of network calculations.Fig. 4 gives thickness fluctuation scope during RBF THICKNESS CONTROL.

(4) adjustment of dynamic self-adapting has been carried out to relevant data, set up dynamic self-adapting RBF neural AGC system.

This network needs to realize on TDC, and present invention employs a kind of dynamic self-adapting RBF network model, this model is based on closest clustering algorithm.This algorithm is a kind of online adaptive cluster learning algorithm, does not need the number determining hidden layer unit in advance.It is optimum for completing the RBF network that cluster obtains, and this algorithm can on-line study.

(1) the Gaussian function width r that suitable is selected, define a vectorial A (1) for depositing the output vector sum belonging to all kinds of, define a counter B (1) and belong to all kinds of number of samples for adding up, wherein 1 is classification number.

(2) from the 1st data to (x ₁, y ₁) start, at x ₁on set up a cluster centre, make c ₁=x ₁, A (1)=y ₁, B (1)=1.The RBF network of such foundation, only has a hidden unit, and the center of this hidden unit is c ₁, this hidden unit is W to the weight vector of output layer ₁=A (1)/B (1).

(3) consider that the 2nd sample data is to (x ₂, y ₂), obtain x ₂to c ₁distance d (the x of this cluster centre ₂-c ₁).If d is (x ₂-c ₁) < r, then c ₁for x ₂nearest neighbor classifier, and make A (1)=y ₁+ y ₂, B (1)=2, W ₁=A (1)/B (1); If d is (x ₂-c ₁) > r, then using x2 as a new cluster centre, and make c ₂=x ₂, A (2)=y ₂, B (2)=1.In the RBF network of above-mentioned foundation, add a hidden unit again, this hidden unit is W to the weight vector of output layer ₂=A (2)/B (2).

(4) suppose that we consider that a kth sample data is to (x _k, y _k) (k=3,4,, N) time, there is M cluster centre, its central point is respectively c ₁, c ₂,, c _m, existing M hidden unit in the RBF network of above-mentioned foundation.Obtain x respectively again _kto the distance d (x of this M cluster centre _k-c _i), i=1,2,, M, if d is (x _k-c _i) be the minimum range in these distances, i.e. c _ifor x _knearest neighbor classifier, then: if d (x _k-c _i) > r, then by x _kas a new cluster centre.Make c _m+1=x _k, M=M+1, A (M)=y _k, B (M)=1, and keep the value of A (i), B (i) constant, i=1,2, M-1, add M hidden unit again in the RBF network of above-mentioned foundation, and this hidden unit is W to the weight vector of output layer _m=A (M)/B (M).If d is (x _k-c _i) < r, do to calculate as follows: A (j)=A (j)+y _k, B (j)=B (j)+1.As i ≠ j, i=1,2,, M, keeps the value of A (i), B (i) constant.Hidden unit to the weight vector of output layer is:

w _i＝A(i)/B(i)

(5) its output of RBF network of setting up according to above-mentioned rule should be:

$f\left(x^k\right)=\frac{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_i{e}^{(-||x^k-{\mathrm{ci}\left|\right|}^2/r^2)}}{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{e}^{(-||x^k-{\mathrm{ci}\left|\right|}^2/r^2)}}$

The size of radius r determines the complexity of dynamic self-adapting RBF network.R is less, and the clusters number obtained is more, and amount of calculation is larger, and precision is also higher; R is larger, and the clusters number obtained is fewer, and amount of calculation is less, but precision is also lower.Because r is an one dimension parameter, usually can find the r that suitable by experiment with control information, this norm more applicable more simultaneously than the number and one of determining hidden unit facilitates many.Due to each input, export data to a new cluster may be produced, therefore this dynamic self-adapting RBF network, be actually simultaneously at the self-adaptative adjustment carrying out parameter and structure two processes.

(5) on hardware TDC, quick realization has been carried out.The application of TDC is very extensive, and all directions that almost can be applied to Industry Control get on, and excellent performance may be used for transmission and the storage of large-scale breathing crack system and signal.

Hardware configuration instrument is that the hardware of automation projects carries out configuration and setting parameter, and the demand according to milling train of the present invention carries out hardware configuration to it, and TDC system architecture is put identical with the installation site of reality.Can comprise: the hardware configuration of milling train: TDCRACK (frame)-UR5213, CPU module-CPU551, defeated people's output module-SM500, MPI communication module-CP50M0.

In TDC frame, allow the slot of interpolation shown in green in software, in general, except CP5100 ethernet communication template, the template of other TDC does not have strict installation site restriction.CPU551 is inserted the 1st slot of frame herein, SM500 inserts the 5th slot of frame, and CP50M0 inserts the 15th slot of frame.

In hardware catalogue, select CPUModules subdirectory, select CPU551 by its drag and drop to the corresponding slot in frame, select Slotl in this example, system-based cycle period is created by CPU in this example, and cycle period is defined as 1.0ms.Need definition CPU Interruption in this example, interrupt interval selects 2ms.Configuration communication module CP50M0, communication module is as optional template in this example, and the inline diagnosis of user program can use the diagnosis connection interface of CPU template, CP50M0 to contain two communication interface X1/X2, and user can define configuration respectively to it.Because communication module is in this example as the passage of inline diagnosis.So need to be defined as MPI agreement.After completing configuration, compiling hardware configuration also preserves the configuration just completing whole hardware.

Program of the present invention is the realization of the control algolithm RBF to (four), and the neuronic quantity of RBF algorithm automatically generates, and after the data training of off-line, just can obtain the basic composition parameter of network, comprise each center vector c and weight w.Use these two parameters on the basis of collocation network calculations, just can build the model of control algolithm.Program is made up of three parts, is inline diagnosis module, RBF module and normalization module respectively.

The foundation of RBF module, the present invention trains the hidden neuron number obtained to be 31, center vector has 31, so the quantity calculated is very large, but be all the calculating of repeatability, this has wherein used the FRM module of CFC, FRM module is a certain moduli block, be different from general mathematical computations module, input has four, but the symbol calculated is not fixing, can by determining four calculated relationship inputted to arranging of FRM port, operable compute sign comprises elementary arithmetic computing addition, subtraction, multiplication and division, power, absolute value, SIN function, delivery, be not less than the smallest positive integral etc. of x, so this is purposes module very widely.RBF module in the present invention, because amount of calculation is huge, repeatability is high, so select based on this module, forms whole network by the switching of different calculating.

Be used as a neuron with three FRM modules, include the calculating of center vector and weights and input signal in a neuron, combining is exactly a complete neutral net.Fig. 5 illustrates the RBF network neural unit in CFC.

The real single chassis operation of rolling can be reflected according to system of the present invention, optimize operation of rolling thickness and precision, improve the production efficiency of rolling.Above-mentioned detailed description of the invention, does not form limiting the scope of the invention.It is to be understood that depend on designing requirement and other factors, various amendment, combination, sub-portfolio can be there is and substitute in those skilled in the art.Any amendment within the spirit and principles in the present invention, equivalently to replace and improvement etc., all should be included within scope.

Claims (4)

1. the measurer for thickness that the AGC system of being furnished with RBF network is supported, is characterized in that:

It is by RBF module, AGC system module, TDC module, thickness gauge, and study singlestandmill forms;

Three neural network modules of described RBF module three-layer artificial neural network by splendid attire respectively form, and are respectively neural modules A, neural module B, neural module C;

Described AGC system comprises RBF interactive module, controls interactive module, computing module;

Aforementioned RBF module, AGC system, TDC module, thickness gauge are all electrically connected with computer;

Aforementioned TDC module comprises, TDCRACK-UR5213, CPU module-CPU551, input/output module-SMS00, MPI communication module-CP50MO;

Study singlestandmill has several detectors, these several detectors are all electrically connected with RBF module;

Described thickness gauge comprises, clamper, ultrasonic transmitter, ultrasonic receiver, center module, computing module, support;

Described thickness gauge appearance has shell coated, and aforementioned ultrasonic wave launcher, ultrasonic receiver side are set up in parallel, and Displacement Meters of exerting pressure is positioned at the side of pressure applicator displacement, and two distance between plates measurement mechanisms are positioned at the side of sample;

Computer is furnished with display unit and display floater, input equipment;

Power module, it is for giving aforementioned RBF module, AGC system module, TDC module, thickness gauge, and study singlestandmill is all electrically connected with supply of electrical energy;

Aforementioned thicknesses meter can seal setting, and be wherein furnished with thermostat and make its inner maintenance constant temperature, thickness gauge inside has insulation lining.

2. utilize the measurer for thickness described in claim 1 to carry out a method for thickness measure, it is characterized in that:

In AGC system module, determine the model of thickness gauge AGC system, and emulate with simulink;

Detection obtains the thickness data of the input and output of study singlestandmill;

RBF neural is established, utilize RBF neural thickness data that the theory of study commonly uses the input and output of singlestandmill in RBF module; Utilize above-mentioned data, training adjustment RBF neural;

The AGC system model of the type that is improved in AGC system module, and the adjustment of dynamic self-adapting has been carried out to relevant data; Carried out the realization of thickness measure scheme by TDC module, above carry out quick realization.

3. the method for thickness measure according to claim 2, is characterized in that:

In AGC system module, determine the model of thickness gauge AGC system, its concrete operations are as follows: set up the Mathematical Modeling comprising fluid-percussion model of isolated, rolling force model, mill spring model milling train;

Set up the step of RBF neural, corresponding neural modules A, neural module B, neural module C set up the RBF neural of three layers respectively, use Gaussian function as RBF, ground floor neuron is input neuron, comprise roll gap deviation, forward pull deviation, backward pull deviation, second layer neuron is hidden neuron, comprises at least 4 hidden neurons, third layer neuron is output neuron, is rolling force deviation;

The step of described training adjustment RBF neural comprises: input roll gap deviation, forward pull deviation, backward pull deviation, provide rolling force deviation, simulation result according to rolling force deviation will as inputoutput data as training data, feedback adjustment RBF neural;

The study of described RBF neural is based on closest clustering algorithm, does not need the number determining hidden layer unit in advance.

Described TDC module includes TDC frame-UR5213, CPU module-CPU551, input/output module-SM500, MPI communication module-CP50MO.

4. the ultrasonic transmitter that uses of measurer for thickness as claimed in claim 1, is characterized in that:

Described ultrasonic transmitter is high-frequency band pass ultrasonic wave transmitting probe.