CN204595644U - Based on the aluminum-bar heating furnace temperature of combustion automaton of neural network - Google Patents

Abstract

The utility model is based on the aluminum-bar heating furnace temperature of combustion automaton of neural network, containing heating furnace combustion system and temperature automatic control electrical system, temperature automatic control electrical system comprises ANN (Artificial Neural Network) Control module, and ANN (Artificial Neural Network) Control module is connected with PID controller.Neural network is divided into forward direction and backward learning two parts, and the forward direction of neural network is for exporting the scale parameter K in PID control _p, integral parameter K _iwith differential coefficient K _d, the backward learning of neural network is used for adaptive adjustment own net weighting coefficient, makes to export the pid control parameter corresponding under certain optimal control law.The method not only can make aluminum-bar heating furnace temperature of combustion control to reach technological requirement, reduces heating-up temperature and changes the rejection rate caused, improve the work efficiency of aluminium bar pressurization; The regulating time of temperature arrival needed for stable state can also be reduced, reduce energy loss, reach the object of energy-saving and emission-reduction.

Description

Based on the aluminum-bar heating furnace temperature of combustion automaton of neural network

Technical field

The utility model relates to a kind of aluminum-bar heating furnace temperature of combustion automatic control technology field.

Background technology

Aluminium section bar manufacturing enterprise belongs to high energy consumption, maximum discharge manufacturing enterprise, and extruding production cost controls to be one of Important Problems of enterprises pay attention.Aluminum-bar heating furnace is the equipment coordinating extruder to use, and uses heated by natural gas to extrusion process design temperature on aluminium bar ingot casting, is conducive to extruded.Aluminum-bar heating furnace is the link that in extrusion line, energy consumption is maximum, therefore reduces the energy consumption of heating furnace, can ensure extruding administration measure, run economically, realize the object of energy-saving and emission-reduction.

Traditional temperature controls adoption rate, integration, differential control, be called for short PID control method, PID control method structure is simple, good stability, method of adjustment, it is a kind of linear control method, according to ratio, the integration and differentiation of the deviation of temperature given value and actual temperature value, form controlled quentity controlled variable by the form of linear combination, the rock gas input quantity of heating furnace is controlled.But due to production on-site environment complexity, be many-sided to the influence factor of temperature, result in temperature is a nonlinear time varying parameter, and there is longer time-lag effect in combustion process.Therefore PID method is used to control to be difficult to obtain good control effects to temperature.

Neural network is a kind of nonlinear network system of simulating human human thinking, there is distributed storage and concurrent collaborative processing power, the network weights coefficient of adaptive adjustment self can be required for different complex environments and multi objective control, with arbitrary accuracy Approximation of Arbitrary Nonlinear Function, may be used for that there is non-linear, hysteresis quality, high-precision control object.

Utility model content

The utility model object is, a kind of aluminum-bar heating furnace temperature of combustion automatic control system based on neural network is proposed, system is according to current heating furnace temperature of combustion and the difference of giving fixed temperature, dynamically adjust the output parameter (PID input parameter) of neural network, realize temperature automatic control; Simultaneously by the back-propagation method of neural network, the constantly network weights coefficient of adjustment neural network, makes it export corresponding to the pid control parameter under certain optimal control law.

The technical solution adopted in the utility model is: a kind of aluminum-bar heating furnace temperature of combustion automaton based on neural network, and its improvement is: include heating furnace combustion system and temperature automatic control electrical system;

Wherein, described heating furnace combustion system comprises natural gas system, air system, electric actuator and combination firelock and forms, and described natural gas system and air system are connected with combination firelock by electric actuator;

Wherein, described temperature automatic control electrical system comprises temperature sensor and PID controller; Described PID controller is connected with electric actuator by D/A module, and described temperature sensor is connected with PID controller by A/D module.

Wherein, described natural gas system comprises gas pipeline and is installed on gas filter, reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve and the big fire solenoid valve on gas pipeline; Described reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve are connected with electric actuator with big fire solenoid valve;

Described air system comprises air duct, the butterfly valve be installed on air duct; Described butterfly valve is connected with electric actuator.

Wherein, described temperature automatic control electrical system comprises ANN (Artificial Neural Network) Control module, and described ANN (Artificial Neural Network) Control module is connected with PID controller.

Wherein, PID controller comprises temperature-measuring module, temperature comparison module and output control module;

Wherein, described temperature-measuring module is connected with temperature sensor and obtains the in-furnace temperature of heating furnace, and in-furnace temperature is passed to temperature comparison module; Described temperature comparison module compares in-furnace temperature (c _t) and design temperature (r _t,) draw temperature gap (e _t=r _t-c _t), and by temperature gap (e _t=r _t-c _t) be passed to output control module;

Wherein, described output control module comprises the first calculation control module and the second calculation control module;

Wherein, the first calculation control module receives temperature gap, and compares with threshold value, when temperature gap is greater than threshold value, carries out heating with full power;

Wherein, the second calculation control module receives temperature gap, and compares with threshold value, when temperature gap is less than threshold value: by ANN (Artificial Neural Network) Control module, PID controller is adjusted.

For reaching this object, first the utility model provides a kind of PID control method, comprising: the linear combination being divided into ratio, integration and differentiation three part, forms PID governing equation:

$P=K_p[e\left(t\right)+\frac{1}{T_i}\∫e\left(t\right)\mathrm{dt}+T_d\frac{\mathrm{de}\left(t\right)}{\mathrm{dt}}]---\left(1\right)$

K _pfor proportional gain, T _ifor integral time, T _dfor derivative time, P is controlled quentity controlled variable, and e (t) is measured value y (t) and the deviation of set-point r (t).In order to by above-mentioned continuous print PID governing equation discretize, the setting sampling period is T, and sampling sequence number is n, then integration and differentiation is discrete turns to:

${\∫}_0^{t}e\left(\mathrm{\τ}\right)\mathrm{d\τ}\≈\underset{j=0}{\overset{k}{\mathrm{\Σ}}}e\left(j\right)T---\left(2\right)$

$\frac{\mathrm{de}\left(t\right)}{\mathrm{dt}}\≈\frac{e\left(n\right)-e(n-1)}{T}---\left(3\right)$

E (n) and e (n-1) is respectively the deviation of n-th time and (n-1)th time sampling, then the governing equation of PID can discretely turn to:

$\begin{array}{c}P\left(n\right)=K_p\{e\left(n\right)+\frac{T}{T_i}\underset{j=0}{\overset{n}{\mathrm{\Σ}}}e\left(j\right)+\frac{T_d}{T}[e\left(n\right)-e(n-1)\left]\right\}\\ =K_{p}e\left(n\right)+K_i\underset{j=0}{\overset{n}{\mathrm{\Σ}}}e\left(j\right)+K_d[e\left(n\right)-e(n-1)]\end{array}---\left(4\right)$

K _pfor proportional gain, K _ifor integral coefficient, K _dfor differential coefficient.According to the e (n) obtained under present sample, obtain controlled quentity controlled variable for controlling the natural gas input quantity of heating furnace according to formula (4).According to the difference of profile material, aluminium bar extruding working temperature is between 400 ° to 450 °, and ultimate temperature is 550 °.Therefore three intervals can be divided into according to temperature variation: heating zone [normal temperature-400 °], steady-state zone [400 °-500 °], inhibition zone [500 °-550 °], need to aluminium bar fast heating to reach the temperature of extruding at heating zone, now should heighten proportional gain, increase sampling period, minimizing integral coefficient and differential coefficient to realize rapid temperature-raising; Steady-state zone needs temperature stabilization in this interval, now should reduce proportional gain, minimizing sampling period, adjustment integral coefficient and differential coefficient to control to realize accurate temperature, inhibition zone is then temperature will be suppressed to arrive this region liquefy to avoid aluminium bar, now should heighten proportional gain to realize suppressing temperature to rise fast.Therefore control system needs dynamically to change control coefrficient to different phase residing for temperature.

Neural network just has above-mentioned dynamic parameters function, neural network has input layer, hidden layer and output layer, each node is referred to as neuron, not contact between the neuron of same layer, the neuron of different layers is connected to form a network between two, contact is each other called weighting coefficient, and initialization weighting coefficient is the random value between (-1,1).Neural network is divided into forward direction and error back propagation two parts, and neural network and PID control to combine, and forms the control system with dynamic parameters.Scale parameter K during the corresponding PID of its output layer parameter difference controls _p, integral parameter K _iwith differential coefficient K _d, its input layer parameter is temperature deviation e (n), the controlled quentity controlled variable P (n) of present sample.The parameter that PID controls exports and is realized by forward direction part, according to the neuron number of hidden layer, successively calculates neuronic output valve from input layer-hidden layer-output layer, and the neuron of every one deck exports and only has an impact to the neuronic input of lower one deck.Neuronic output activation function adopts the Sigmoid function of non-negative symmetry:

$f\left(x\right)=\frac{e^x}{e^x+e^{-x}}---\left(5\right)$

Given one by n neuronic input layer, p neuronic hidden layer, the network structure of q neuronic output layer composition, defines its variable as follows:

The output vector of input layer: x=(x ₁, x ₂..., x _n)

The input variable of hidden layer: hi=(hi ₁, hi ₂..., hi _p)

The output variable of hidden layer: ho=(ho ₁, ho ₂..., ho _p)

The input variable of output layer: yi=(yi ₁, yi ₂..., yi _q)

The output variable of output layer: yo=(yo ₁, yo ₂..., yo _q)

Given desired output variable: d=(d ₁, d ₂..., d _q)

The forward direction of neural network is calculated as follows: in the t time sampling, error function is $e\left(t\right)=\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2,$ Being input as of hidden layer ${\mathrm{hi}}_h\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ih}}{x}_i\left(t\right),h=\mathrm{1,2},...p,$ The output of hidden layer is ho _h(t)=f (hi _h(t)), being input as of output layer the output of output layer is yo _o(t)=f (yi _o(t)), the scale parameter K of the corresponding PID control system of output of output layer _p, integral parameter K _iwith differential coefficient K _d.

The neuron weighting coefficient adjustment of neural network is realized by error back propagation part, the error back propagation method adaptive adjustment own net weighting coefficient of neural network, makes it export corresponding to the pid control parameter under certain optimal control law:

1. first error of calculation function e (t) to each neuronic partial derivative of output layer

$\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ho}}}=\frac{\∂e\left(t\right)}{{\∂\mathrm{yi}}_o\left(t\right)}\frac{{\∂\mathrm{yi}}_o\left(t\right)}{{\∂w}_{\mathrm{ho}}}$

$\frac{{\∂\mathrm{yi}}_o\left(t\right)}{{\∂w}_{\mathrm{ho}}}=\frac{\∂\left(\underset{h=1}{\overset{p}{\mathrm{\Σ}}}{w}_{\mathrm{ho}}{\mathrm{ho}}_h\left(t\right)\right)}{{\∂w}_{\mathrm{ho}}}={\mathrm{ho}}_h\left(t\right)$

$\frac{\∂e\left(t\right)}{{\∂\mathrm{yi}}_o\left(t\right)}=\frac{\∂\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2\right)}{{\∂\mathrm{yi}}_o\left(t\right)}=-(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))f^{\′}\left({\mathrm{yi}}_o\left(t\right)\right)=-{\mathrm{\δ}}_o\left(t\right)$

2. with the neuronic partial derivative of output layer revise the neuronic weighting coefficient of output layer, μ represents learning rate:

$w_{\mathrm{ho}}(t+1)=w_{\mathrm{ho}}\left(t\right)+\mathrm{\μ}\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ho}}}=w_{\mathrm{ho}}\left(t\right)-\mathrm{\μ}{\mathrm{\δ}}_o\left(t\right){\mathrm{ho}}_h\left(t\right)$

3. utilize the weighting coefficient of hidden layer to output layer, the partial derivative of output layer with the output error of calculation function of hidden layer to each neuronic partial derivative of hidden layer

$\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ih}}}=\frac{\∂e\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}\frac{{\∂\mathrm{hi}}_h\left(t\right)}{{\∂w}_{\mathrm{ih}}}$

$\frac{{\∂\mathrm{hi}}_h\left(t\right)}{{\∂w}_{\mathrm{ih}}}=\frac{\∂\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ih}}{x}_i\left(t\right)\right)}{{\∂w}_{\mathrm{ih}}}=x_i\left(t\right)$

$\frac{\∂e\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}=\frac{\∂\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2\right)}{{\∂\mathrm{ho}}_h\left(t\right)}\frac{{\∂\mathrm{ho}}_h\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}=-\left(\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{\mathrm{\δ}}_o\left(t\right)w_{\mathrm{ho}}\right)f^{\′}\left({\mathrm{hi}}_h\left(t\right)\right)=-{\mathrm{\δ}}_h\left(t\right)$

4. with the partial derivative of hidden layer neuron revise the weighting coefficient of hidden layer neuron

$w_{\mathrm{ih}}(t+1)=w_{\mathrm{ih}}\left(t\right)+\mathrm{\μ}\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ih}}}=w_{\mathrm{ho}}\left(t\right)-\mathrm{\μ}{\mathrm{\δ}}_h\left(t\right)x_i\left(t\right)$

5. calculate error function now as error arrives appointment

Precision, then stop iteration, otherwise continue backpropagation with the weighting coefficient of roll-off network.

The beneficial effects of the utility model are:

For aluminum-bar heating furnace temperature of combustion nonlinear time-varying, feature that time-lag effect is strong, control on basis at traditional PID, add Neural Network Control Algorithm and implementation, control for temperature self-adaptation, neural network is divided into forward direction and backward learning two parts, and the forward direction of neural network is for exporting the scale parameter K in PID control _p, integral parameter K _iwith differential coefficient K _d, the backward learning of neural network is used for adaptive adjustment own net weighting coefficient, makes to export the pid control parameter corresponding under certain optimal control law.The overshoot of nerve network control system step response curve is less, and the settling time of system shortens greatly, and the number of oscillation reduces, and system table reveals good stability.

PID control method combines with neural net method, forms the control system with dynamic parameters, the scale parameter K in the corresponding PID control of output layer parameter difference of neural network _p, integral parameter K _iwith differential coefficient K _d, its input layer parameter is temperature deviation e (n), the controlled quentity controlled variable P (n) of present sample.The parameter that PID controls exports and is realized by forward direction part, successively calculates neuronic output valve from input layer-hidden layer-output layer, and the neuron of every one deck exports and only has an impact to the neuronic input of lower one deck.

Be applied to the temperature controlled neural network of aluminum-bar heating furnace and there is backward learning function, the backward learning method adaptive adjustment own net weighting coefficient of neural network, it is exported corresponding to the pid control parameter under certain optimal control law.First error of calculation function is to each neuronic partial derivative of output layer, then the neuronic weighting coefficient of output layer neuronic partial derivative correction output layer is used, recycling hidden layer to the weighting coefficient of output layer, the partial derivative of output layer and the output error of calculation function of hidden layer to each neuronic partial derivative of hidden layer, finally use the weighting coefficient of the partial derivative correction hidden layer neuron of hidden layer neuron, by repeatedly learning the adaptive control realizing neural network, heating furnace temperature of combustion is prevented to fluctuate.

The utility model temperature of combustion automatic control system is in conjunction with traditional PID control and neural net method, fixing ratio, integration, differential parameter is adopted to regulate temperature of combustion different from traditional PID control, neural network dynamically can revise the pid control parameter in the different temperatures stage according to current temperature signal and last controlled quentity controlled variable of sampling, and utilizes error back propagation method with roll-off network weighting coefficient simultaneously.The method not only efficiently avoid the complex parameters adjustment that data hysteresis quality causes, and also effectively reduces energy loss.

Accompanying drawing explanation

Fig. 1 is the utility model aluminium section bar heating furnace temperature of combustion automatic control system structural drawing.

Fig. 2 is neural network adaptive algorithm system construction drawing of the present utility model.

Fig. 3 is the ANN (Artificial Neural Network) Control that provides of an embodiment of the present utility model and PID controlling curve comparison diagram.

Embodiment

Below in conjunction with drawings and the specific embodiments, the utility model is carried out more in detail and complete explanation.Be understandable that, specific embodiment described herein only for explaining the utility model, but not to restriction of the present utility model.It also should be noted that, for convenience of description, illustrate only the part relevant to the utility model in accompanying drawing but not full content.

Aluminum-bar heating furnace temperature of combustion automaton based on neural network of the present utility model, include heating furnace combustion system A and temperature automatic control electrical system B two ingredients, as shown in Figure 1, described heating furnace combustion system A comprises natural gas system 1, air system 2, electric actuator 3 and combination firelock 4 and forms, and described natural gas system 1 and air system 2 are connected with combination firelock 4 by electric actuator 3.Wherein, described temperature automatic control electrical system B comprises temperature sensor 5 and PID controller 6; Described PID controller 6 is connected with electric actuator 4 by D/A module, and described temperature sensor 5 is connected with PID controller 6 by A/D module.Wherein, described natural gas system comprises gas pipeline and is installed on gas filter, reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve and the big fire solenoid valve on gas pipeline; Described reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve are connected with electric actuator with big fire solenoid valve; Described air system comprises air duct, the butterfly valve be installed on air duct; Described butterfly valve is connected with electric actuator.

Described temperature automatic control electrical system B also comprises ANN (Artificial Neural Network) Control module 7, and described ANN (Artificial Neural Network) Control module 7 is connected with PID controller 6.Wherein, PID controller 6 comprises temperature-measuring module, temperature comparison module and output control module; Wherein, described temperature-measuring module is connected with temperature sensor and obtains the in-furnace temperature of heating furnace, and in-furnace temperature is passed to temperature comparison module; Described temperature comparison module compares in-furnace temperature (c _t) and design temperature (r _t,) draw temperature gap (e _t=r _t-c _t), and by temperature gap (e _t=r _t-c _t) be passed to output control module; Wherein, described output control module comprises the first calculation control module and the second calculation control module; Wherein, the first calculation control module receives temperature gap, and compares with threshold value, when temperature gap is greater than threshold value, carries out heating with full power; Wherein, the second calculation control module receives temperature gap, and compares with threshold value, when temperature gap is less than threshold value, is adjusted by ANN (Artificial Neural Network) Control module 7 pairs of PID controller 6, namely adopts modified integral algorithm of PID to carry out adjusting of controling parameters.

Above-mentioned PID control method, comprising: the linear combination being divided into ratio, integration and differentiation three part, forms PID governing equation:

$P=K_p[e\left(t\right)+\frac{1}{T_i}\∫e\left(t\right)\mathrm{dt}+T_d\frac{\mathrm{de}\left(t\right)}{\mathrm{dt}}]---\left(1\right)$

K _pfor proportional gain, T _ifor integral time, T _dfor derivative time, P is controlled quentity controlled variable, and e (t) is measured value y (t) and the deviation of set-point r (t).In order to by above-mentioned continuous print PID governing equation discretize, the setting sampling period is T, and sampling sequence number is n, then integration and differentiation is discrete turns to:

${\∫}_0^{t}e\left(\mathrm{\τ}\right)\mathrm{d\τ}\≈\underset{j=0}{\overset{k}{\mathrm{\Σ}}}e\left(j\right)T---\left(2\right)$

$\frac{\mathrm{de}\left(t\right)}{\mathrm{dt}}\≈\frac{e\left(n\right)-e(n-1)}{T}---\left(3\right)$

E (n) and e (n-1) is respectively the deviation of n-th time and (n-1)th time sampling, then the governing equation of PID can discretely turn to:

$\begin{array}{c}P\left(n\right)=K_p\{e\left(n\right)+\frac{T}{T_i}\underset{j=0}{\overset{n}{\mathrm{\Σ}}}e\left(j\right)+\frac{T_d}{T}[e\left(n\right)-e(n-1)\left]\right\}\\ =K_{p}e\left(n\right)+K_i\underset{j=0}{\overset{n}{\mathrm{\Σ}}}e\left(j\right)+K_d[e\left(n\right)-e(n-1)]\end{array}---\left(4\right)$

K _pfor proportional gain, K _ifor integral coefficient, K _dfor differential coefficient.According to the e (n) obtained under present sample, obtain controlled quentity controlled variable for controlling the natural gas input quantity of heating furnace according to formula (4).According to the difference of profile material, aluminium bar extruding working temperature is between 400 ° to 450 °, and ultimate temperature is 550 °.Therefore three intervals can be divided into according to temperature variation: heating zone [normal temperature-400 °], steady-state zone [400 °-500 °], inhibition zone [500 °-550 °], need to aluminium bar fast heating to reach the temperature of extruding at heating zone, now should heighten proportional gain, increase sampling period, minimizing integral coefficient and differential coefficient to realize rapid temperature-raising; Steady-state zone needs temperature stabilization in this interval, now should reduce proportional gain, minimizing sampling period, adjustment integral coefficient and differential coefficient to control to realize accurate temperature, inhibition zone is then temperature will be suppressed to arrive this region liquefy to avoid aluminium bar, now should heighten proportional gain to realize suppressing temperature to rise fast.Therefore control system needs dynamically to change control coefrficient to different phase residing for temperature.

Neural network just has above-mentioned dynamic parameters function, neural network has input layer, hidden layer and output layer, each node is referred to as neuron, not contact between the neuron of same layer, the neuron of different layers is connected to form a network between two, contact is each other called weighting coefficient, and initialization weighting coefficient is the random value between (-1,1).Neural network is divided into forward direction and error back propagation two parts, and neural network and PID control to combine, and forms the control system with dynamic parameters.Scale parameter K during the corresponding PID of its output layer parameter difference controls _p, integral parameter K _iwith differential coefficient K _d, its input layer parameter is temperature deviation e (n), the controlled quentity controlled variable P (n) of present sample.The parameter that PID controls exports and is realized by forward direction part, according to the neuron number of hidden layer, successively calculates neuronic output valve from input layer-hidden layer-output layer, and the neuron of every one deck exports and only has an impact to the neuronic input of lower one deck.Neuronic output activation function adopts the Sigmoid function of non-negative symmetry:

$f\left(x\right)=\frac{e^x}{e^x+e^{-x}}---\left(5\right)$

Given one by n neuronic input layer, p neuronic hidden layer, the network structure of q neuronic output layer composition, defines its variable as follows:

The output vector of input layer: x=(x ₁, x ₂..., x _n)

The input variable of hidden layer: hi=(hi ₁, hi ₂..., hi _p)

The output variable of hidden layer: ho=(ho ₁, ho ₂..., ho _p)

The input variable of output layer: yi=(yi ₁, yi ₂..., yi _q)

The output variable of output layer: yo=(yo ₁, yo ₂..., yo _q)

Given desired output variable: d=(d ₁, d ₂..., d _q)

The forward direction of neural network is calculated as follows: in the t time sampling, error function is $e\left(t\right)=\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2,$ Being input as of hidden layer ${\mathrm{hi}}_h\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ih}}{x}_i\left(t\right),h=\mathrm{1,2},...p,$ The output of hidden layer is ho _h(t)=f (hi _h(t)), being input as of output layer the output of output layer is yo _o(t)=f (yi _o(t)), the scale parameter K of the corresponding PID control system of output of output layer _p, integral parameter K _iwith differential coefficient K _d.

The neuron weighting coefficient adjustment of neural network is realized by error back propagation part, the error back propagation method adaptive adjustment own net weighting coefficient of neural network, makes it export corresponding to the pid control parameter under certain optimal control law:

1. first error of calculation function e (t) to each neuronic partial derivative of output layer

$\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ho}}}=\frac{\∂e\left(t\right)}{{\∂\mathrm{yi}}_o\left(t\right)}\frac{{\∂\mathrm{yi}}_o\left(t\right)}{{\∂w}_{\mathrm{ho}}}$

$\frac{{\∂\mathrm{yi}}_o\left(t\right)}{{\∂w}_{\mathrm{ho}}}=\frac{\∂\left(\underset{h=1}{\overset{p}{\mathrm{\Σ}}}{w}_{\mathrm{ho}}{\mathrm{ho}}_h\left(t\right)\right)}{{\∂w}_{\mathrm{ho}}}={\mathrm{ho}}_h\left(t\right)$

$\frac{\∂e\left(t\right)}{{\∂\mathrm{yi}}_o\left(t\right)}=\frac{\∂\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2\right)}{{\∂\mathrm{yi}}_o\left(t\right)}=-(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))f^{\′}\left({\mathrm{yi}}_o\left(t\right)\right)=-{\mathrm{\δ}}_o\left(t\right)$

2. with the neuronic partial derivative of output layer revise the neuronic weighting coefficient of output layer, μ represents learning rate:

$w_{\mathrm{ho}}(t+1)=w_{\mathrm{ho}}\left(t\right)+\mathrm{\μ}\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ho}}}=w_{\mathrm{ho}}\left(t\right)-\mathrm{\μ}{\mathrm{\δ}}_o\left(t\right){\mathrm{ho}}_h\left(t\right)$

3. utilize the weighting coefficient of hidden layer to output layer, the partial derivative of output layer with the output error of calculation function of hidden layer to each neuronic partial derivative of hidden layer

$\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ih}}}=\frac{\∂e\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}\frac{{\∂\mathrm{hi}}_h\left(t\right)}{{\∂w}_{\mathrm{ih}}}$

$\frac{{\∂\mathrm{hi}}_h\left(t\right)}{{\∂w}_{\mathrm{ih}}}=\frac{\∂\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ih}}{x}_i\left(t\right)\right)}{{\∂w}_{\mathrm{ih}}}=x_i\left(t\right)$

$\frac{\∂e\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}=\frac{\∂\left(\frac{1}{2}\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{(d_o\left(t\right)-{\mathrm{yo}}_o\left(t\right))}^2\right)}{{\∂\mathrm{ho}}_h\left(t\right)}\frac{{\∂\mathrm{ho}}_h\left(t\right)}{{\∂\mathrm{hi}}_h\left(t\right)}=-\left(\underset{o=1}{\overset{q}{\mathrm{\Σ}}}{\mathrm{\δ}}_o\left(t\right)w_{\mathrm{ho}}\right)f^{\′}\left({\mathrm{hi}}_h\left(t\right)\right)=-{\mathrm{\δ}}_h\left(t\right)$

4. with the partial derivative of hidden layer neuron revise the weighting coefficient of hidden layer neuron

$w_{\mathrm{ih}}(t+1)=w_{\mathrm{ih}}\left(t\right)+\mathrm{\μ}\frac{\∂e\left(t\right)}{{\∂w}_{\mathrm{ih}}}=w_{\mathrm{ho}}\left(t\right)-\mathrm{\μ}{\mathrm{\δ}}_h\left(t\right)x_i\left(t\right)$

Calculate error function now as error arrives designated precision, then stop iteration, otherwise continue backpropagation with the weighting coefficient of roll-off network.

The automatic control realization process of furnace temp is: first temperature sensor is measured in real time firelock temperature and the temperature data collected is converted into voltage signal, and voltage signal is converted to the discernible digital signal of host computer through A/D module I CP-7017; Then host computer is by current temperature signal input neural network, obtains the scale parameter K for calculating PID control system _p, integral parameter K _iwith differential coefficient K _dsimultaneously neural network according to current temperature signal with last controlled quentity controlled variable error originated from input back-propagation method of sampling with roll-off network weighting coefficient, the dynamic self-adapting parameter realizing PID control system regulates, neural network structure figure as shown in Figure 2, finally obtain control signal according to PID control system, control signal is converted to the current signal controlling electric actuator through D/A module I CP-7024, realize the control to furnace temp.

PID Initial parameter sets is as follows, upper temperature limit 530 degree, lower limit 200 degree, upper deviation 550 degree, the lower limit of variation 180 degree, output current 4-20mA, and gain is set to 3,2 seconds integral time, 0.5 second derivative time.When temperature is close to setting value 530 degree, electric current exports as 4mA, and now electric actuator aperture is minimum, and firelock exports minimum, and at this time system can constantly be contrasted by feedback quantity and setting value, and control temperature is at about 530 degree.

ANN (Artificial Neural Network) Control and the PID controlling curve comparison diagram of the utility model embodiment shown in Fig. 3, horizontal ordinate is the sampling time, ordinate is the temperature value after A/D conversion, can find out in figure, PID controlling curve of comparing, neural network realizes online adaptive adjustment to three of PID controling parameters according to temperature variation, the overshoot of control system step response curve is less, the settling time of system shortens greatly, and the number of oscillation reduces, and system table reveals good stability.

Be only preferred embodiment of the present utility model described in upper, be not limited to the utility model, to those skilled in the art, the utility model can have various change and change.All do within spirit of the present utility model and principle any amendment, equivalent replacement, improvement etc., all should be included within protection domain of the present utility model.

Claims (4)

1. based on an aluminum-bar heating furnace temperature of combustion automaton for neural network, it is characterized in that: include heating furnace combustion system and temperature automatic control electrical system;

Wherein, described heating furnace combustion system comprises natural gas system, air system, electric actuator and combination firelock and forms, and described natural gas system and air system are connected with combination firelock by electric actuator;

Wherein, described temperature automatic control electrical system comprises temperature sensor and PID controller; Described PID controller is connected with electric actuator by D/A module, and described temperature sensor is connected with PID controller by A/D module.

2., as claimed in claim 1 based on the aluminum-bar heating furnace temperature of combustion automaton of neural network, it is characterized in that:

Described natural gas system comprises gas pipeline and is installed on gas filter, reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve and the big fire solenoid valve on gas pipeline; Described reduction valve, the total solenoid valve of combustion gas, igniter solenoid valve are connected with electric actuator with big fire solenoid valve;

Described air system comprises air duct, the butterfly valve be installed on air duct; Described butterfly valve is connected with electric actuator.

3. as claimed in claim 1 or 2 based on the aluminum-bar heating furnace temperature of combustion automaton of neural network, it is characterized in that: described temperature automatic control electrical system comprises ANN (Artificial Neural Network) Control module, described ANN (Artificial Neural Network) Control module is connected with PID controller.

4., as claimed in claim 3 based on the aluminum-bar heating furnace temperature of combustion automaton of neural network, it is characterized in that:

Wherein, PID controller comprises temperature-measuring module, temperature comparison module and output control module;

Wherein, described temperature-measuring module is connected with temperature sensor and obtains the in-furnace temperature of heating furnace, and in-furnace temperature is passed to temperature comparison module; Described temperature comparison module compares in-furnace temperature (c _t) and design temperature (r _t,) draw temperature gap (e _t=r _t-c _t), and by temperature gap (e _t=r _t-c _t) be passed to output control module;

Wherein, described output control module comprises the first calculation control module and the second calculation control module;

Wherein, the first calculation control module receives temperature gap, and compares with threshold value, when temperature gap is greater than threshold value, carries out heating with full power;

Wherein, the second calculation control module receives temperature gap, and compares with threshold value, when temperature gap is less than threshold value: adjusted PID controller by ANN (Artificial Neural Network) Control module.