CN102129259B - Neural network proportion integration (PI)-based intelligent temperature control system and method for sand dust environment test wind tunnel - Google Patents

Abstract

The invention discloses a neural network proportion integration (PI)-based intelligent temperature control system for a sand dust environment test wind tunnel, which is characterized in that: (1) a neural network system structure is established; (2) the blended learning and training on neural network parameters is performed; (3) PI control is performed; and (4) amplitude limiting processing is performed. Control variable coordinated control factors are obtained by combining a PI controller and a neural network and using the self-adaption of the neural network and capabilities of off-line learning and on-line learning so as to effectively determine main control equipment and auxiliary control equipment. Coordinated effective control on controlled object temperature is performed through the PI controller; and after the PI controller outputs a control variable, amplitude limiting processing is performed on the controlled variable by using an S function, so that the control variable is optimized. The system overcomes the influence of poor coordination on control in the prior art, improves the reliability and the coordination of the control, broadens the application range, and can also be used for coordinated control of output signals of other sensors.

Description

For the intelligent temperature control system based on neuron network PI controller and the method for sand dust environment test tunnel

Technical field

The present invention relates to a kind of intelligent temperature control system and method based on neuron network PI controller (proportional integral) that is applied to sand dust environment test tunnel, for high precision and the High-reliability Control of sand dust environment test tunnel temperature.

Background technology

Sand dust environment is an important environmental factor that causes many engineerings and/or Weapon Equipment Failure, and its main types of damage has: erosion, wearing and tearing, corrosion and infiltration etc.Sand dust environmental test is to analyze, and evaluates all kinds equipment and the serviceability of instrument under desert or arid area wind sand environment, reliability, the important means of stability.No matter be at national military standard, or all the temperature standard of sand dust environmental test made to strict regulation in other kinds standard.

Due to the needs of modern scientific research and actual production, increasing place all needs sand dust environmental test, again because the variation range of thermal load in the test wind and wind-tunnel is very large and sudden, this has just more and more highlighted the temperature controlled importance of sand dust environmental test simultaneously.The domestic temperature control at large-scale backwash formula sand dust environment test tunnel is at present mainly to adopt table of equipment cooler and two kinds of distinct temperature control devices of electric heater in air-conditioning bypass.And due in the time carrying out low speed drifting dust and the test of high speed drifting dust, what take is heating and two kinds of diverse Temperature Control Measures of refrigeration, and difference variation scope is increased, and controls difficulty and increases.Simultaneously because the temperature control in sand dust environmental test has pure stickiness, the features such as some uncertainties of non-linear and parameter, be easy to produce larger vibration or error, therefore traditional control strategy is often difficult to meet the index request of performance, is also feature and the difficult point that the control of sand dust environment test tunnel temperature is different from ordinary temp control object simultaneously.

In circulation air path, rotation speed of fan changes, secondary air flow and temperature variation, the variation of environment temperature all can cause the thermal perturbation in air channel, and by regulating electric heater heating power or cooling water flow, can make thermal perturbation controlled, but thermal load difference in air channel is very large, need to adopt different coordination control measure.How to eliminate the disturbance that these interference sources cause, and overcome the impact that various nonlinear factors are brought temperature control system, the gordian technique that high precision of temperature and high harmony control method are whole domain of control temperature.

At present, the temperature control solving under sand dust environment test tunnel has two kinds of thinkings, and one is on hardware configuration, to adopt temperature control device, and temperature can in normal scope, be regulated and controled.Although this scheme can well improve the control of temperature, but increase the complicacy of hardware, improve cost, and in reliability, had very large restriction; The 2nd, software control mode.Which is by adopting modeling and control algorithm, approach the ability of Nonlinear Mapping arbitrarily by control strategy and learning ability, temperature is carried out to real-time control, and the error causing due to non-linear factor is effectively compensated, affect the precision of its compensation, guarantee its accuracy and harmony.

Summary of the invention

Technology of the present invention is dealt with problems: overcome prior art in high precision, the deficiency that high stability and coordination control aspect exist, intelligent temperature control system and the method for a kind of PI based on neural network in sand dust environment test tunnel have been proposed, realize temperature controlled stability and harmony, in the temperature control of whole sand dust environment test tunnel, effectively improved temperature controlled coordination and accuracy; Expanded its range of application simultaneously, and simple, be easy to realize.

According to an aspect of the present invention, provide a kind of PI (proportional integral) intelligent temperature control system of environmental simulation system, described environmental simulation system comprises multiple control gears, it is characterized in that comprising:

A nerve network controller, for according to the measured air speed value obtaining of wind speed inductor that is arranged on described environmental simulation system, produces its output;

PI controller, for according to being arranged on the measured temperature value obtaining of temperature inductor of described environmental simulation system and the output of described nerve network controller, obtains corresponding rough controlled quentity controlled variable.

According to a further aspect of the present invention, a kind of intelligent temperature control method of environmental simulation system is provided, described environmental simulation system comprises multiple control gears, it is characterized in that comprising:

According to the measured air speed value obtaining of wind speed inductor being arranged in described environmental simulation system, produce the output of a nerve network controller;

Utilize PI controller, according to being arranged on the measured temperature value obtaining of temperature inductor in described environmental simulation system and the output of described nerve network controller, obtain corresponding rough controlled quentity controlled variable.

Accompanying drawing explanation

Fig. 1 is the overview flow chart of sand dust environment test tunnel of the present invention.

Fig. 2 is the control flow chart of sand dust environment test tunnel temperature intelligent of the present invention control.

Fig. 3 is the principle assumption diagram of sand dust environment test tunnel temperature intelligent of the present invention control based on neuron network PI controller control.

Fig. 4 is the structural drawing of the nerve network controller of sand dust environment test tunnel temperature intelligent of the present invention control.

Fig. 5 is the neuron network PI controller control structure figure based on blended learning training algorithm of sand dust environment test tunnel temperature intelligent of the present invention control.

Fig. 6 is that sand dust environment test tunnel of the present invention is based on neuron network PI controller intelligent temperature control system learning training algorithm overview flow chart.

Fig. 7 is the process flow diagram of the improvement BP learning algorithm of sand dust environment test tunnel temperature intelligent of the present invention control.

Reference numeral:

E (t): the accurate amount e of temperature deviation _c: the accurate amount of error rate

R (t): temperature input value y (t): controlled Temperature Quantity output valve

W: the connection weights Z between every layer _cv: coordinate controlling elements

V: air speed value

101: temperature inductor 102:PI controller

103: circulating cooling water regulating valve 105: electric heater power regulating eqiupment

106: wind speed inductor 107: nerve network controller

S601: judge whether difference exceedes definite value S602:BP off-line learning

S603: hereditary on-line study

Embodiment

Technical solution of the present invention provides a kind of PI intelligent temperature control system based on neural network for the test of sand dust environmental wind tunnel, and this temperature control system has at least two temperature control devices, and its feature is:

-first, by the wind speed inductor being contained in circulation air path, the air speed value measuring is passed to nerve network controller, nerve network controller show that according to this air speed value is coordinated controlling elements, and being defined as major control mechanism according to these coordination controlling elements one in described at least two temperature control devices, all the other control gears are confirmed as auxiliary controls;

-afterwards, the temperature value that this is coordinated to controlling elements and measures with the temperature inductor in circulation air path, passes to traditional PI controller, carries out a series of map functions in PI controller, finally obtains corresponding rough controlled quentity controlled variable;

-then, utilize S function to carry out corresponding amplitude limiting processing the rough control variable obtaining, draw final accurate controlled quentity controlled variable, at least two temperature control devices described in then passing to;

-carried out the corresponding operation of controlled quentity controlled variable by control mechanism, adjust the heating power of electric heater and regulate circulating cooling water flow, now controlled device temperature has obtained corresponding control and change;

Temperature inductor in-wind-tunnel sends measured value temperature to PI controller again, whether completes desired control, as be desired temperature value to check, and keeps stable, otherwise continues preceding step, until reach requirement.

As Figure 1-3 according in embodiments of the invention, described two temperature control devices comprise an electric heater power regulating eqiupment and a circulating cooling water regulating valve.

Say further, for the requirement of above-mentioned intelligent temperature control, from essence, the Based Intelligent Control of the temperature of the environmental simulation system of sand dust environment test tunnel be a kind of be subject to non-linear and time the change factor dynamic model that affects, the model of complexity can adopt nonlinear feedback linearization technique to design control law like this.The present invention combines traditional PI system and forms the PI Temperature Intelligent Control System based on neural network with neural network, any non-linear expression's ability of utilizing neural network to have, by neural network to being carried out effective coordination by control mechanism, determine which control gear is that main control mechanism and which control gear are auxiliary controls, show that is coordinated controlling elements, thereby realize the coordination control of temperature variation in wind-tunnel, to reaching the requirement of system performance index, as shown in Figure 3.Traditional PI control system has simplicity of design, is easy to realization, high reliability, is particularly useful for setting up the deterministic control system of mathematical models, and the parameter of its controller can be adjusted according to the linear mathematical model of controlled device.But when plant characteristic changes or when operating condition changes, controlling quality will decline, and even affects the normal operation of control system.Therefore, the present invention proposes, the mode that application neural network and conventional PI control device combine, utilizes neural network coordinating to realize above-mentioned combination on the basis of controlling, utilize neural network to realize and determine coordination control, thereby the reliability of PI controller and accuracy are guaranteed.

According to one embodiment of present invention, described PI controller comprises:

The one PI controller 102, for according to described temperature value and described coordination controlling elements, produces the first rough controlled quentity controlled variable,

The 2nd PI controller 109, for according to described temperature value and described coordination controlling elements, produces the second rough controlled quentity controlled variable,

Described intelligent temperature control system further comprises:

The first limiter 108, for the described first rough controlled quentity controlled variable is carried out to amplitude limit and optimization process, thereby produces the meticulous controlled quentity controlled variable of optimization for described electric heater power regulating eqiupment control gear 105,

The second limiter 110, for the described second rough controlled quentity controlled variable is carried out to amplitude limit and optimization process, thereby produces the meticulous controlled quentity controlled variable of optimization for described chilled water modulation valve control device 103.

According to one embodiment of present invention, the first limiter 108 and the second limiter 110 utilize S function to carry out described amplitude limit and optimization process.

As shown in Figure 4, according to one embodiment of present invention, the neural network that is applied to the coordination controlling elements of coordinating control adopts the Three Tiered Network Architecture of single hidden layer.Wherein, ground floor is input layer, and the second layer is hidden layer, and the 3rd layer is output layer.The transforming function transformation function of hidden layer unit is radial basis function form, and it is a kind of non-negative nonlinear function to the decay of central point radial symmetry of local distribution.This neural network is nonlinear by input layer space to the mapping in hidden layer space, and hidden layer space is linear to the mapping of output region, and the learning training speed after so greatly having accelerated has been avoided local minimum problem.

According to a preferred embodiment of the present invention, for this neural network, first the weighting parameter that utilizes improvement BP learning algorithm afterwards to adjust this neural network carries out the global optimization of off-line learning process, then utilize the weight coefficient of genetic algorithm on-line optimization neural network, make temperature controlled coordination controlling elements there is good degree of stability and degree of accuracy in parameter variation and external disturbance situation.

Intelligent temperature control system according to an embodiment of the invention comprises a mixed learning algorithm module (501), and it comprises:

Judge module, for judging described temperature error (e _c) whether exceed a predetermined value,

An off-line dynamic temperature model learning module, at described temperature error (e _c) carry out off-line dynamic temperature model learning while having exceeded described predetermined value, thus utilize BP algorithm to realize the off-line global optimizing of weights, the margin of error is dwindled, and shorten the required time of on-line study,

On-line study training module, at described temperature error (e _c) carry out on-line study training while being no more than described predetermined value, by genetic algorithm, produce multiple starting points and start search simultaneously higher-dimension solution space is random, carry out guidance search direction by fitness function, finally obtain optimum weighting parameter.

Fig. 6 has shown the flow process of the preferred embodiment, wherein, after initial value, judges whether the variation difference of input quantity exceedes predetermined value (step S601), judges that whether temperature gap is excessive, whether exceedes a boundary value; If value beyond the mark, will carry out the BP off-line learning (step S602) of model; On the contrary, if temperature gap value not beyond the mark (being that in wind-tunnel, sum of any pressure drop variations amplitude is not very large) is carried out on-line study training (step S603).By above-mentioned flow process, can obtain preferably temperature controlled coordination controlling elements in wind-tunnel, make PI controller can effectively coordinate to control.

Shown in Fig. 7, be the algorithm flow chart of BP off-line learning training algorithm according to an embodiment of the invention.

Below, describe according to an embodiment of the invention, for the concrete layout of one of the PI intelligent temperature control system based on neural network of temperature environment simulation system (as test air tunnel):

(1) nerve network system structure

Adopt three layers of feedforward network of single hidden layer to form according to the neural network of a specific embodiment of the present invention.The transforming function transformation function of hidden layer unit adopts the Sigmoid function of Symmetrical, and it is a kind of non-negative nonlinear function to the decay of central point radial symmetry of local distribution.This neural network is nonlinear by input layer space to the mapping in hidden layer space, and be linear from hidden layer space to insinuating of output layer space, such neural network has realized accelerates the feature of pace of learning, and has avoided certain oscillatory and local minimum problem.As shown in figure (4).

1) input layer

Input layer has adopted 3 special inputs, corresponding input v (air speed value) respectively, error e _c(error of temperature), and constant 1, constant plays an interference here, and input mode vector is x=[v, e _c, 1], compared with the structure x=[v of the input vectors of 2, e _c] more realistic working environment.

The input-output function of input layer is:

O _i＝x(i) (1)

In formula, the number that i is input layer, i=1,2,3

2) the neuronic of hidden layer hidden layer is input as:

${\mathrm{net}}_j\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}{O}_i---\left(2\right)$

In formula, w _jifor input layer is to the weights of hidden layer

Hidden layer neuron is output as:

O _j(k)＝f(net _j(k)) (3)

In formula, the number of j hidden layer neuron, the activation function that f is hidden layer

The activation function of hidden layer is got the Sigmoid function of Symmetrical:

$f\left(x\right)=\tanh \left(x\right)=\frac{\exp \left(x\right)-\exp (-x)}{\exp \left(x\right)+\exp (-x)}---\left(4\right)$

3) the neuronic of output layer output layer is input as:

${\mathrm{net}}_l\left(k\right)=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{O}_j\left(k\right)---\left(5\right)$

In formula, w _ijfor hidden layer is to the weights of output layer

Output layer is neuronic to be output as:

O _l(k)＝g(net _l(k)) (6)

In formula,

$g\left(x\right)=\frac{1}{2}(1+\tanh \left(x\right))=\frac{\exp \left(x\right)}{\exp \left(x\right)+\exp (-x)}---\left(7\right)$

The neuronic output of output layer corresponding coordination controlling elements Z _cv

(2) blended learning of neural network parameter training

On the basis of above-mentioned neural network model, the mode that adopts online training and off-line training to combine is carried out, and through after prejudging, decision is to carry out BP off-line training, still carries out the online training based on genetic algorithm.

In hybrid learning algorithm, the BP algorithm after improving is by the weights and bias between continuous off-line roll-off network neuron, makes the model of the continuous approaching to reality of network.Online genetic learning training algorithm, some limitation in order to overcome the training of BP off-line learning, because genetic learning training algorithm is apply the higher-dimension solution space multiple starting points of random generation and start search in searching process simultaneously, carry out guidance search direction by fitness function, thereby region of search is wide, search efficiency is high.And, in the time controlling in real time, before adjusting weighting parameter online at every turn, to first judge that whether temperature gap is excessive, if exceed fixed ceiling, still will first carry out the model learning training of off-line, until difference in allowed band after, then carry out online learning training.Can reduce like this vibration, and shorten the control time, as shown in Figure 6.

Because error back propagation (Back Propagation is called for short BP) algorithm has the advantages such as simple, easy to learn, speed of convergence is very fast, thereby is widely used in the adjustment of network weight.But BP algorithm is in the time being applied to real-time control, and its pace of learning is slow, is easily absorbed in the shortcoming of local minimum, sometimes also may can not get global optimum.This is due in the time that network changes another training sample into by a training sample.Thereby the toning that easily causes weight coefficient due to larger initial value error lengthens the adjustment time.Therefore, according to one embodiment of present invention, introduce a factor of momentum a, to reduce overtravel, be more conducive to the acceleration study of off-line.As shown in Figure 7, after dynamic model initialization, process layer by layer through Hidden unit from input layer, until output layer, it is the function (S702) of input and weights, obtain the output (S703) of hidden layer and the each unit of output layer, obtain desired output and the actual deviation (S704) of exporting, deviate is judged to (S705), if exist deviation to enter backpropagation between actual output and desired output, oppositely return by former forward-propagating approach, calculate Hidden unit error (S706), and solve (S707) by the negative gradient direction of error function, afterwards each layer of neuronic weight coefficient revised to (S708), finally make the error function trend minimum of expecting.Factor of momentum a is one and determines the weighting factor of results of learning in the past, common 0 < a < 1, the scope of corresponding another variation factor η is 0 < η < 0.5, and effect is now comparatively desirable.Meanwhile, after carrying out iteration, these two factors all can constantly be adjusted at every turn, revise.Consider the coupling between each variable, get performance index function and be:

$J=E\left(k\right)=(Z_{\mathrm{cv}}^{\&prime;}-Z_{\mathrm{cv}}^{\&prime;\&prime;})+\frac{1}{2}\mathrm{\&Sigma;}[r\left(k\right)-y\left(k\right){]}^2---\left(8\right)$

The modified computing formulae that finally obtains weights is:

$\mathrm{\&Delta;}{W}_{\mathrm{ij},(j-i)p}^k=-\mathrm{\&eta;}\left(k\right)\&CenterDot;{\mathrm{\&delta;}}_{\mathrm{ij}}\left(k\right)u_{(j-1)p}^k+a\left(k\right)\left[\mathrm{\&Delta;}{W}_{\mathrm{ij},(j-1)p}^k\right]---\left(9\right)$

${\mathrm{\&delta;}}_{\mathrm{ij}}\left(k\right)=u_{\mathrm{ij}}\left(k\right)[1-u_{\mathrm{ij}}\left(k\right)]\underset{p=1}{\overset{N(i-1)}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_{(i-1)p}\&CenterDot;W_{\mathrm{ij},(i-1)p}^k--\left(10\right)$

After having adjusted weights by above-mentioned off-line learning, unavoidable and actual parameter has a certain distance, for this reason, utilizes subsequently the online weighting parameter of revising in real time the network that obtains of genetic algorithm, search procedure, take raw parameter as basis, is carried out in the less field of raw parameter.

Genetic algorithm is a kind of searching algorithm based on natural selection and population genetic mechanism, and it has simulated breeding, mating and variation phenomenon in natural selection and natural genetic process.It regards each possible solution as body one by one in colony, and each individuality is encoded into the form of character string, according to predetermined objective function, each individuality is evaluated, and provides a fitness value.These individual fitness values utilize genetic operator to carry out genetic manipulation to these individualities, retain excellent individuality, eliminate poor individuality, make final weights towards a good state development.At this, we apply traditional binary coding mode, in the time that the scale of neural network is slightly large, the length of chromosome individuality will be very large, thereby affect the efficiency of genetic algorithm, for this this intelligent system adopts floating-point encoding, what floating-point encoding method was used is the actual value of decision variable, and individual code length equals the number of its decision variable.In native system, the weights variable of neural network has 24, selects the inverse of error sum of squares as fitness function.For genetic manipulation, the selection algorithm in genetic manipulation adopts the method for how much sequences of standardization.This sort method sorts to individuality by the value of fitting, and distributes and selects probability according to individual position.How much individual selection new probability formula of sequence definition of standardization are:

$P_i=\frac{q}{1-{(1-q)}^n}{(1-q)}^{r-1}---\left(11\right)$

Wherein: the selection probability that q is optimized individual, r is individual sequence number, n is Population Size

For floating-point code, crossover algorithm adopts mathematics to intersect and two kinds of modes that combine of heuristic intersection, and two kinds of cross methods are combined with the detectivity that can strengthen algorithm.For keeping the diversity of population and preventing from precocious phenomenon from need to the gene in former population being added to a random perturbation.Mutation operation in native system adopts polynary non-uniform mutation (multiNonUnJfMutafion) strategy generation mutant gene to form new population, respectively independent variable is carried out after non-uniform mutation in its solution space, then get at random a combination as variation result.Principle type is:

X′ _i＝X _i+(b _i-X _i)f(g)，r ₁＜0.5

X′ _i＝X _i-(X _i-a)f(g)，r ₁≥0.5 (12)

Wherein:

$f\left(g\right)={\left(r_2(1-\frac{g}{g_{\max }})\right)}^b---\left(13\right)$

R ₁, r ₂for even variable immediately between [0,1], a _i, b _ibe respectively variable upper and lower limit, g is current optimization algebraically, g _maxfor largest optimization algebraically, b is the systematic parameter of weighing level of disruption.

Based on foregoing description, show that to carry out by genetic algorithm the training step that online neural network weight coefficient optimizes as follows:

The input of-given neural network, output collection;

-determine the coded system of network weight coefficient, selected genetic manipulation, arranges genetic parameter;

-population scale N to set, produces initial population immediately;

Each position string in-decoding population, tries to achieve N group network weight coefficient, obtains having N network of same structure;

-by inputting sample set through forward direction propagation algorithm, try to achieve N the network output that N group network weights are corresponding;

The objective function of-setting network, is converted into fitness, and N network evaluated;

-select in hereditary space according to fitness to operate;

-operate accordingly according to selected intersection, variation and Some Related Algorithms, parameter, obtain population of new generation;

-return to step 4, until meet performance requirement, obtain one group of weight coefficient of optimizing.

(3) PI controls

After obtaining coordinating controlling elements, coordination controlling elements are inputed to PI controller, coordination controlling elements and its parameter are carried out further computing and combination by PI controller, draws control variable, thereby to coordinated effective control by control mechanism.

According to one embodiment of present invention, the governing equation of circulating cooling water flow is:

G _w＝K _cpZ _cve _t+K _ci∫(Z _cve _t)dt (14)

In formula:

G _w: the mass rate e of chilled water _t: the temperature difference

K _cp: the scale-up factor of recirculated cooling water water flow controller

K _ci: the integral coefficient of recirculated cooling water water flow controller

K _cd: the differential coefficient of recirculated cooling water water flow controller

The governing equation of electric heater heating power is:

Q _d＝K _tp(1-Z _cv)e _t+K _ti∫(1-Z _cv)e _tdt (15)

In formula:

Q _d: electric heater heating power

K _tp: the scale-up factor of electric heater controller

K _ti: the integral coefficient of electric heater controller

K _td: the differential coefficient of electric heater controller

(4) amplitude limiting processing

After controlled variable, have unavoidably some inexactnesies, therefore, also will carry out effective amplitude limiting processing to control variable.Here apply S type function and carry out the processing of amplitude limit, S function formula is:

${\mathrm{\&lambda;}}_{\mathrm{cv}}=\frac{1}{1+e^{-\mathrm{av}}}---\left(16\right)$

Obtaining amplitude limiting processing formula by above-mentioned formula is:

f(u)＝u ₂ u≥u ₂

f(u)＝u*λ _cv u ₁≤u＜u ₂

f(u)＝u ₁ u＜u ₁ (17)

In formula:

U: control variable Q _dor G _w;

U ₁, u ₂: amplitude limit threshold values;

F (u): the control variable after amplitude limiting processing

In a word, principle of the present invention is: the Temperature Intelligent Control System that the present invention proposes the sand dust environment test tunnel of a kind of PI based on neural network.PI controller due to its simplicity of design, be easy to realize, high reliability, be widely used in, in the process control and motion control of the industries such as electromechanics, metallurgy, machinery, chemical industry, being particularly useful for setting up the deterministic control system of mathematical models.But actual industrial process exists the interference of a lot of uncertain factors, the correction that uses classical PI controller to carry out control system does not often reach desirable control effect.And neural network has Nonlinear Mapping, self-learning capability, the features such as distributed store ability and process information, therefore combine neural network and PI controller, but are different from traditional combination.Combination is herein to be based upon to coordinate to control the combination on basis, due to uncertainty and the complicacy coordinating to control, therefore we utilize the non-linear of neural network, self-learning capability, a pre-judgement of coordinating to control before input signal enters PI controller, show that is coordinated controlling elements, determine in the control gear of all coordination controls, which control gear is major control mechanism, which is auxiliary controls, make like this control system not only possess processing inexactness, probabilistic ability, also there is the ability of coordinating control stiffness simultaneously.Again by self-learning capability, constantly revise neural network and connect weights simultaneously, adjust and coordinate controlling elements, make this factor coordinate parameter value under regulating and controlling for certain is optimum, to reaching the requirement of system performance index.This is coordinated to controlling elements and send into PI controller, carry out corresponding control operation, after the output of PI controller output signal, also to carry out corresponding amplitude limiting processing, control variable is further made to precision, finally reach harmony and the reliability of the temperature intelligent control of sand dust environment test tunnel.Reduce the fluctuating range in sand dust temperature adjustment process.

The neural network structure of the Temperature Intelligent Control System based on neuron network PI controller is to be made up of the three-layer network of a single hidden layer.The transforming function transformation function of hidden layer unit adopts the Sigmoid function in radial basis function, and it is a kind of non-negative nonlinear function to the decay of central point radial symmetry of local distribution.This neural network is nonlinear by the input space to the mapping in hidden layer space, and be linear from hidden layer space to insinuating of output layer space, such neural network has realized accelerates the feature of pace of learning, and has avoided certain oscillatory and local minimum problem.

It should be noted that, the neural network after learning training can be good at adjusting parameter, the nonlinear function of the high-precision input/output signal that approaches temperature intelligent control, and there is very strong generalization ability.The process of learning training is after initialization, enter judge module, differentiate temperature gap and whether exceeded a limit value, if exceeded this limit value, carry out off-line dynamic temperature model learning, before can carrying out on-line study like this, the margin of error is dwindled, shorten the time of on-line study.It is in fact to utilize improvement BP algorithm afterwards to realize the off-line global optimizing of weights that parameter is adjusted in off-line model study.Proceed afterwards judgement, until this difference within the scope of limit value after, carry out the on-line study training stage again, it is mainly to pass through genetic algorithm that weighting parameter is adjusted in on-line study, also start search in the multiple starting points of the random generation of higher-dimension solution space simultaneously, carry out guidance search direction by fitness function, finally obtain optimum weighting parameter.

According to one embodiment of present invention, when definition off-line learning, adopt the dynamic model of expert system:

$M_t{c}_t\frac{{\mathrm{d\&theta;}}_t}{\mathrm{dt}}=k_h{n}_f^3+G_a{c}_a{(\mathrm{\&theta;}}_a-{\mathrm{\&theta;}}_t)-k_s{F}_s\&CenterDot;({\mathrm{\&theta;}}_t-{\mathrm{\&theta;}}_s)+Q_d-G_w{c}_w({\mathrm{\&theta;}}_t+\mathrm{\&Delta;}{\mathrm{\&theta;}}_{\mathrm{wo}}-{\mathrm{\&theta;}}_{\mathrm{wi}})---\left(18\right)$

M _t: the equivalent metal quality c in circulation air path _t: mean specific heat

N _f ³: the rotation speed of fan k in circulation air path _h: scale-up factor

K _s: the coefficient of heat transfer F of circulation air path outside surface _s: area of dissipation

θ _s, θ _a, θ _t: environment temperature G _w: the mass rate of chilled water

C _w, c _a: specific heat θ _wi: inlet temperature

Δ θ _wo: the poor G of endpiece of chilled water _a: circulation air path secondary air mass rate

Q _d: the heating power of electric heater

Advantage of the present invention comprises:

(1) adopt neural network and PI to control the control method combining, profit is non-linear by neural network, self-learning capability, a pre-judgement of coordinating to control before input signal enters PI controller, show that is coordinated controlling elements, determine in the control gear of all coordination controls, which control gear is major control mechanism, which is auxiliary controls, to reach the harmony of control and to realize simple.

(2) the present invention can realize and controlling with software, and compensation way can on-line control, has feature simple to operate, with low cost.

(3) three-layer network of the present invention forms, and have good tracking characteristics, and the antijamming capability of system is strong, and travelling speed is fast, has very strong feasibility and practicality for real-time system.

(4) learning training of the present invention is after initialization, enter judge module, differentiate temperature gap and whether exceeded a limit value, if exceeded this limit value, carry out off-line dynamic temperature model learning, when difference is further dwindled, within the scope of limit value time, carry out on-line study, this mode has and well avoids oscillatory and improve stability for the learning training of parameter again.

(5) control signal of the present invention, after the output of PI controller, utilizes S function to carry out necessary amplitude limiting processing, makes to control the further precision of output signal, makes the harmony of control more stable.

Core concept of the present invention is to utilize ANN (Artificial Neural Network) Control to have good Nonlinear Mapping, the advantages such as self-learning capability, PI control and neural network are effectively combined, the pre-judgement that profit was coordinated before input signal enters PI controller by neural network to control, show that is coordinated controlling elements, determine in the control gear of all coordination controls, which control gear is major control mechanism, which is auxiliary controls, to reach the harmony of control.Take certain measure to accelerate response speed simultaneously and improve antijamming capability.Make temperature controlled harmony and the reliability of sand dust environment test tunnel obtain effective raising, when effectively having improved the interior sand dust temperature control effect of sand dust environment test tunnel, strengthened harmony and the stability controlled.Concrete steps are as follows:

(1) set up nerve network system structure

Native system is got constant 1, the temperature difference, wind speed is as the input variable of neural network, profit is non-linear by neural network, self-learning capability, the pre-judgement of of coordinating to control before input signal enters PI controller, show that is coordinated controlling elements, determine in the control gear of all coordination controls, which control gear is major control mechanism, and which is auxiliary controls.Selecting input layer number is 3, and the number of hidden nodes is 6, and output layer nodes is 1.

(2) blended learning of neural network parameter training

Only have the neural network after training just can draw optimum coordination controlling elements, the present invention will utilize dynamic model, adopt off-line learning and on-line study to train the pattern combining, the neural network that training is obtained by step (1), obtains coordinating controlling elements.The training process formula of neural network, by the weight coefficient between constantly adjusting each layer, makes neural network be output as optimum, makes to coordinate the maximum effect of control.

As shown in Figure 5, in the time carrying out neuroid Algorithm for Training, to carry out in advance judge module, in the time that temperature gap exceedes regulation boundary value, improve the off-line learning training of BP learning algorithm, proceed afterwards judgement, until the temperature difference is in the value of defining scope time, recycling genetic algorithm is adjusted online, by above-mentioned steps (1)-(9), makes weights optimization.Finally draw output quantity coordination controlling elements, export to PI controller, carry out effectively the control of coordination.

(3) PI controls

Will be by step (1), the coordination controlling elements that (2) obtain input to PI controller, carry out corresponding control strategy operation in conjunction with PI controller principle, draw control variable.

(4) amplitude limiting processing

After control variable output, application of formula (17) is carried out necessary amplitude limiting processing to it, and this temperature control to wind-tunnel under sand dust environmental test plays a part precision.

Should be understood that, the description of in above narration and explanation, the present invention being carried out just illustrates but not is determinate, and do not departing under the prerequisite of the present invention limiting as appended claims, can above-described embodiment carried out various changes, distortion and/or be revised.

Claims (10)

1. an intelligent temperature control system for PI environmental simulation system, described environmental simulation system comprises multiple control gears, it is characterized in that comprising:

A nerve network controller (107), for according to the measured air speed value obtaining of wind speed inductor that is arranged on described environmental simulation system, produces its output;

PI controller (102,109), for according to being arranged on the measured temperature value obtaining of temperature inductor of described environmental simulation system and the output of described nerve network controller (107), obtains corresponding rough controlled quentity controlled variable,

Wherein, described nerve network controller (107) is for showing that according to described air speed value is coordinated controlling elements, be defined as major control mechanism according to these coordination controlling elements one at least two temperature control devices, and remaining control gear in described at least two temperature control devices is defined as to auxiliary controls.

2. the intelligent temperature control system of environmental simulation system according to claim 1, is characterized in that:

PI controller is for obtaining described rough controlled quentity controlled variable according to described temperature value and described coordination controlling elements.

3. the intelligent temperature control system of environmental simulation system according to claim 2, is characterized in that:

Described at least two temperature control devices comprise an electric heater power regulating eqiupment control gear (105) and a circulating cooling water regulating valve control gear (103),

Described PI controller comprises:

The one PI controller (102), for according to described temperature value and described coordination controlling elements, produces the first rough controlled quentity controlled variable,

The 2nd PI controller (109), for according to described temperature value and described coordination controlling elements, produces the second rough controlled quentity controlled variable,

Described intelligent temperature control system further comprises:

The first limiter (108), for the described first rough controlled quentity controlled variable is carried out to amplitude limit and optimization process, thereby produces the meticulous controlled quentity controlled variable of optimization for described electric heater power regulating eqiupment control gear (105),

The second limiter (110), for the described second rough controlled quentity controlled variable is carried out to amplitude limit and optimization process, thereby produces the meticulous controlled quentity controlled variable of optimization for described circulating cooling water regulating valve control gear (103).

4. the intelligent temperature control system of environmental simulation system according to claim 3, is characterized in that:

The first limiter (108) and the second limiter (110) utilize S function to carry out described amplitude limit and optimization process,

Described nerve network controller (107) comprising:

Input layer, it comprises input, the corresponding temperature error e of corresponding air speed value _cinput, the input of corresponding constant 1, wherein constant plays an interference, its input mode vector is x=[v, e _c, 1], the input-output function of input layer is:

O _i=x(i) (1)

In formula, the number that i is input layer, i=1,2,3

Hidden layer, its neuronic being input as:

${\mathrm{net}}_j\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ji}}{O}_i---\left(2\right)$

In formula, w _jifor input layer is to the weights of hidden layer, and described hidden layer neuron is output as:

O _j(k)=f(net _j(k)) (3)

In formula, the number of j hidden layer neuron, the activation function that f is hidden layer, and the activation function of described hidden layer is got the Sigmoid function of Symmetrical:

$f\left(x\right)=\tanh \left(x\right)=\frac{\exp \left(x\right)-\exp (-x)}{\exp \left(x\right)+\exp (-x)}---\left(4\right)$

Output layer, the neuronic of described output layer is input as:

${\mathrm{net}}_l\left(k\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{O}_j\left(k\right)---\left(5\right)$

In formula, w _ijfor hidden layer is to the weights of output layer, and described output layer is neuronic is output as:

O _l(k)=g(net _l(k)) (6)

In formula,

$g\left(x\right)=\frac{1}{2}(1+\tanh \left(x\right))=\frac{\exp \left(x\right)}{\exp \left(x\right)+\exp (-x)}---\left(7\right)$

The neuronic output of described output layer corresponding coordination controlling elements Z _cv.

5. the intelligent temperature control system of environmental simulation system according to claim 4, is characterized in that further comprising:

Mixed learning algorithm module (501), it comprises:

Judge module, for judging described temperature error e _cwhether exceed a predetermined value,

An off-line dynamic temperature model learning module, at described temperature error e _cwhile having exceeded described predetermined value, carry out off-line dynamic temperature model learning, thereby utilize BP algorithm to realize the off-line global optimizing of weights, the margin of error is dwindled, and shorten the required time of on-line study,

On-line study training module, at described temperature error e _cwhile being no more than described predetermined value, carry out on-line study training, by genetic algorithm, also start search in the multiple starting points of the random generation of higher-dimension solution space simultaneously, carry out guidance search direction by fitness function, finally obtain optimum weighting parameter.

6. an intelligent temperature control method for environmental simulation system, described environmental simulation system comprises multiple control gears, it is characterized in that comprising:

According to the measured air speed value obtaining of wind speed inductor being arranged in described environmental simulation system, the output that produces a nerve network controller (107);

Utilize PI controller (102,109), according to being arranged on the measured temperature value obtaining of temperature inductor in described environmental simulation system and the output of described nerve network controller (107), obtain corresponding rough controlled quentity controlled variable,

Utilize described nerve network controller (107), according to described air speed value, show that is coordinated controlling elements,

Be defined as major control mechanism according to these coordination controlling elements one at least two temperature control devices,

Remaining control gear in described at least two temperature control devices is defined as to auxiliary controls.

7. the intelligent temperature control method of environmental simulation system according to claim 6, is characterized in that:

Utilize described PI controller, obtain described rough controlled quentity controlled variable according to described temperature value and described coordination controlling elements.

8. the intelligent temperature control method of environmental simulation system according to claim 7, wherein said at least two temperature control devices comprise an electric heater power regulating eqiupment control gear (105) and a circulating cooling water regulating valve control gear (103), and described PI controller comprises:

The one PI controller (102), for according to described temperature value and described coordination controlling elements, produces the first rough controlled quentity controlled variable,

The 2nd PI controller (109), for according to described temperature value and described coordination controlling elements, produces the second rough controlled quentity controlled variable,

Described intelligent temperature control method further comprises:

Utilize first limiter (108), the described first rough controlled quentity controlled variable carried out to amplitude limit and optimization process, thereby produce the meticulous controlled quentity controlled variable of optimization for described electric heater power regulating eqiupment control gear (105),

Utilize second limiter (110), the described second rough controlled quentity controlled variable is carried out to amplitude limit and optimization process, thereby produce the meticulous controlled quentity controlled variable of optimization for described circulating cooling water regulating valve control gear (103).

9. the intelligent temperature control method of environmental simulation system according to claim 8, is characterized in that:

In described the first limiter (108) and the second limiter (110), utilize S function to carry out described amplitude limit and optimization process,

Wherein said nerve network controller (107) comprising:

Input layer, it comprises input, the corresponding temperature error e of corresponding air speed value _cinput, the input of corresponding constant 1, wherein constant plays an interference, its input mode vector is x=[v, e _c, 1], the input-output function of input layer is:

O _i=x(i) (1)

In formula, the number that i is input layer, i=1,2,3

Hidden layer, its neuronic being input as:

${\mathrm{net}}_j\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ji}}{O}_i---\left(2\right)$

In formula, w _jifor input layer is to the weights of hidden layer, and described hidden layer neuron is output as:

O _j(k)=f(net _j(k)) (3)

In formula, the number of j hidden layer neuron, the activation function that f is hidden layer, and the activation function of described hidden layer is got the Sigmoid function of Symmetrical:

$f\left(x\right)=\tanh \left(x\right)=\frac{\exp \left(x\right)-\exp (-x)}{\exp \left(x\right)+\exp (-x)}---\left(4\right)$

Output layer, the neuronic of described output layer is input as:

${\mathrm{net}}_l\left(k\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{O}_j\left(k\right)---\left(5\right)$

In formula, w _ijfor hidden layer is to the weights of output layer, and described output layer is neuronic is output as:

O _l(k)=g(net _l(k)) (6)

In formula,

$g\left(x\right)=\frac{1}{2}(1+\tanh \left(x\right))=\frac{\exp \left(x\right)}{\exp \left(x\right)+\exp (-x)}---\left(7\right)$

The neuronic output of described output layer corresponding coordination controlling elements Z _cv.

10. the intelligent temperature control method of environmental simulation system according to claim 9, is characterized in that further comprising:

A determining step, judges described temperature error e _cwhether exceed a predetermined value,

An off-line dynamic temperature model learning step, at described temperature error e _cwhile having exceeded described predetermined value, carry out off-line dynamic temperature model learning, thereby utilize BP algorithm to realize the off-line global optimizing of weights, the margin of error is dwindled, and shorten the required time of on-line study,

An on-line study training step, at described temperature error e _cwhile being no more than described predetermined value, carry out on-line study training, by genetic algorithm, also start search in the multiple starting points of the random generation of higher-dimension solution space simultaneously, carry out guidance search direction by fitness function, finally obtain optimum weighting parameter.

Images