CN101458534A - Apparatus and method for controlling semiconductor equipment reaction chamber temperature - Google Patents

Abstract

The invention provides a device for controlling temperature of a semiconductor equipment reaction chamber and a method thereof, which includes a temperature control arithmetic module, at least two temperature detecting modules, at least two temperature control modules; wherein, the temperature detecting module detects temperature in detecting position of the semiconductor equipment reaction chamber, detected temperature value is input into the temperature control arithmetic module, the temperature control arithmetic module calculates control amount aiming at each temperature control module according with all detected temperature value and all input temperature pre-set value and using neural network control arithmetic; the temperature control module controls temperature in detecting position of the semiconductor equipment reaction chamber. The method and device provided by the invention can overcome defect that single circuit controls temperature in various part of the semiconductor device reaction chamber singly, accordingly temperature of the reaction chamber is controlled better.

Description

A kind of apparatus and method of controlling semiconductor equipment reaction chamber temperature

Technical field

The present invention relates to the semiconductor equipment reaction chamber temperature control technology, particularly relate to a kind of apparatus and method of controlling semiconductor equipment reaction chamber temperature.

Background technology

In semiconducter process,, need the quantity of parameters of strict control semiconductor equipment reaction chamber for stability and the consistance that guarantees the processing technology result.Wherein, temperature is exactly one of key parameter of semiconductor equipment reaction chamber.

Fig. 1 is a typical semiconductor equipment reaction chamber synoptic diagram.As shown in Figure 1, this reaction chamber can be divided into reaction chamber 1, following reaction chamber 2 and 3 three parts of pumping chamber, and each part comprises four well heaters and a temperature sensor.Wherein, small circle represents to be distributed in the well heater of reaction chamber, and little triangle is represented the temperature sensor of each several part.Each part of reaction chamber can be utilized temperature sensor detection temperature separately, utilizes extraneous temperature control algorithm to produce control signal more respectively, by the control signal that produces well heater is separately controlled, thereby realization is to the control of reaction chamber temperature.

For the method that illustrates that better prior art is controlled reaction chamber temperature, the temperature with a part of control reaction chamber is that example is described in detail below.

Fig. 2 is the device synoptic diagram that the temperature of a part of reaction chamber is controlled.As shown in Figure 2, this device can comprise temperature detecting module 201, temperature control algorithm module 202 and temperature control modules 203.

Wherein, temperature detecting module 201 can comprise temperature inductor 2011, amplifier 2012, mould/number conversion module 2013.Wherein, temperature inductor 2011 is placed in the reaction chamber, is used for the temperature of detection reaction chamber, and detected temperature is exported to amplifier 2012 with the form of simulating signal.Amplifier 2012 and mould/number conversion module 2013 is placed on the outside of reaction chamber, and amplifier 2012 is exported to mould/number conversion module 2013 after the simulating signal of temperature inductor 2011 inputs is amplified again; Mould/number conversion module 2013 is to export to temperature control algorithm module 202 after the digital signal with analog signal conversion.

Temperature control algorithm module 202 is according to the desired temperature and the temperature detecting module 201 detected temperature of input, and utilizes temperature control algorithm to obtain controlled quentity controlled variable, and the signal that carries controlled quentity controlled variable is exported to temperature control modules 203.Temperature control algorithm described here is the temperature control algorithm at the single output of single input, such as: fuzzy algorithm, proportional-integral-differential (PID) control algolithm etc.

Temperature control modules 203 is controlled the temperature of reaction chamber according to controlled quentity controlled variable.Particularly, be pulse signal if 202 outputs of temperature control algorithm module carry the signal of controlled quentity controlled variable, so, shown in Fig. 3 a, temperature control modules 203 can comprise pulse-width modulator 2031a, solid-state relay 2032a and well heater 2033.Wherein, export to solid-state relay 2032a after the pulse signal amplification of pulse-width modulator 2031a with input; Solid-state relay 2032a controls the power output of well heater 2033 according to the pulse signal of input; Described well heater 2033 is placed in the reaction chamber, can have one or morely, is used for providing power under the control of solid-state relay 2032a, with the temperature of control reaction chamber.

If it is digital signal that temperature control modules 203 outputs carry the signal of controlled quentity controlled variable, so, shown in Fig. 3 b, temperature control modules 203 can comprise D/A switch module 2031b, voltage regulator 2032b and well heater 2033.Wherein, the digital signal that D/A switch module 2031b will carry controlled quentity controlled variable is converted to simulating signal, and exports to voltage regulator 2032b; Voltage regulator 2032b carries out correspondent voltage according to the simulating signal that carries controlled quentity controlled variable to be regulated, and the power output of well heater 2033 is controlled; Described well heater 2033 provides power under the control of voltage regulator 2032b, with the temperature of control reaction chamber.No matter temperature control modules 203 is above-mentioned which kind of structure, how to utilize the method for pulse signal and Digital Signals reaction chamber temperature all to belong to prior art, repeat no more herein.

As can be seen from Figure 2, temperature detecting module 201, temperature control algorithm module 202 and temperature control modules 203 can constitute a control loop, utilize temperature control algorithm that the temperature maintenance of reaction chamber counterpart is arrived predefined temperature value.

The above-mentioned apparatus structure that the temperature of a part of reaction chamber is controlled of only having described.If reaction chamber is divided into a plurality of parts as shown in Figure 1, then need to utilize many covering devices as shown in Figure 2 respectively the temperature of various piece to be controlled, its situation is as shown in Figure 4.That is to say that each temperature control algorithm module 202 only utilizes the temperature detecting module 201 detected temperature and the predefined temperature value of self correspondence to calculate, and independently the temperature of reaction chamber various piece is controlled.

In the practical application, no matter reaction chamber is divided into several sections, various piece is not isolated in fact, and its temperature belongs to a kind of coupled relation, and perturbation is more intense.From Fig. 2 and Fig. 4 as can be known, prior art only adopts independently control loop respectively the temperature of reaction chamber each several part to be controlled, and does not consider the coupled relation of each several part thermal perturbation, the control of reaction chamber temperature is difficult to the effect that reaches good.

Summary of the invention

In view of this, fundamental purpose of the present invention is to provide a kind of device of controlling semiconductor equipment reaction chamber temperature, can overcome the defective that single control loop is controlled semiconductor equipment reaction chamber each several part temperature independently, better reaction chamber temperature be controlled.

In order to achieve the above object, the technical scheme of the present invention's proposition is:

A kind of device of controlling semiconductor equipment reaction chamber temperature, this device comprise temperature control algorithm module, at least two temperature detecting module, at least two temperature control modules; Described temperature detecting module is corresponding one by one with described temperature control modules;

Described temperature detecting module is used for the temperature of semiconductor equipment reaction chamber inspection positions is detected, and detected temperature value is exported to the temperature control algorithm module;

Described temperature control algorithm module, be used for desired temperature at inspection positions according to the detected temperature value of all temperature detecting module and all inputs, and utilize the ANN (Artificial Neural Network) Control algorithm computation to go out controlled quentity controlled variable, and export to corresponding temperature control modules at each temperature control modules;

Described temperature control modules is used for according to the controlled quentity controlled variable of input the temperature of temperature detecting module inspection positions being controlled.

In the such scheme, described temperature control algorithm module comprises:

The input layer module is used for all desired temperatures that receive are exported to a hiding layer module; The detected temperature value of all temperature detecting module that receives is exported to a hiding layer module;

Hide layer module, be used to receive desired temperature and detected temperature value from the input layer module; According to the input layer of first memory module preservation and the weights of hiding between the layer desired temperature and detected temperature value are weighted calculating, result of calculation is exported to the output layer module as hiding layer result of calculation;

The output layer module is used to receive a hiding layer result of calculation; Hiding layer and the weights between the output layer preserved according to second memory module are weighted calculating to hiding layer result of calculation, obtain the controlled quentity controlled variable at each temperature control modules, and export to corresponding temperature control modules;

First memory module is used to preserve the weights between input layer and the hiding layer;

Second memory module is used to preserve the weights of hiding between layer and the output layer.

In the such scheme, described input layer module comprises at least two input layer submodules, and each input layer submodule comprises first input layer and second input layer; Described hiding layer module comprises at least two hiding straton modules, and each hiding straton module comprises ratio neuron, integration neuron and differential neuron; Described output layer module comprises at least two output layer submodules, and each output layer submodule comprises an output layer neuron; Described input layer submodule, hiding straton module and output layer submodule are corresponding one by one;

For any input layer submodule, the hiding straton module corresponding and corresponding output layer submodule with it, described first input layer receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, and exports to ratio neuron, integration neuron and differential neuron; Described second input layer receives from the detected temperature value of described inspection positions, and exports to ratio neuron, integration neuron and differential neuron respectively; Described ratio neuron, integration neuron, differential neuron are weighted calculating according to the weights between the weights between the desired temperature, detected temperature value of input, self and first input layer, self and second input layer respectively, with the result that calculates respectively as self current input value; Again that self is the current input value of described ratio neuron multiply by preset proportional value and obtains the output valve of self, and exports to all output layer neurons; Described integration neuron input value that self is current and last output valve sum output valve as self, and export to all output layer neurons; Input value that described differential neuron is current with self and last input value sum be as the output valve of self, and export to all output layer neurons; Described output layer neuron basis is from all proportions neuron, integration neuron and the neuronic value of differential, and self respectively and the weights between all proportions neuron, integration neuron and the differential neuron be weighted calculating, acquisition is at the controlled quentity controlled variable of the temperature detecting module of self correspondence, and exports to the temperature detecting module of self correspondence.

In the such scheme, described temperature control algorithm module further comprises:

The weights study module is used to receive the desired temperature of all inputs, receives the detected temperature value of all temperature detecting module; Judge whether to reach the learning objective value that sets in advance according to all desired temperatures and all detected temperature values, if do not reach, then upgrade first memory module input layer of preserving and the weights of hiding between the layer to hiding layer study step-length, upgrade weights between hiding layer of the preservation of second memory module and the output layer to output layer study step-length according to predefined hiding layer according to predefined input layer.

The present invention also proposes a kind of method of controlling semiconductor equipment reaction chamber temperature, can overcome the defective that single control loop is controlled semiconductor equipment reaction chamber each several part temperature independently, better reaction chamber temperature is controlled.

At above-mentioned purpose, the technical scheme that the present invention proposes is:

A kind of method of controlling semiconductor equipment reaction chamber temperature, this method comprises:

According to the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, from the detected temperature value of described at least two inspection positions, and the ANN (Artificial Neural Network) Control algorithm computation goes out the controlled quentity controlled variable at each inspection positions; Respectively the temperature of each inspection positions of semiconductor equipment reaction chamber is controlled according to the controlled quentity controlled variable that calculates.

In the such scheme, the method for described calculation control amount comprises:

A1, input layer receive the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, receive the detected temperature value of described at least two inspection positions, and export to and hide layer;

A2, hiding layer are weighted calculating according to the weights between input layer of preserving and the hiding layer to described desired temperature and detected temperature value, and result of calculation is exported to output layer as hiding layer result of calculation;

A3, output layer are weighted calculating according to the weights between hiding layer of preserving and the output layer to hiding layer result of calculation, obtain the controlled quentity controlled variable at each inspection positions.

In the such scheme, described input layer comprises at least two sub-input layers, and each sub-input layer comprises first input layer and second input layer; Described hiding layer comprises the hiding layer of at least two sons, and each son is hidden layer and comprised ratio neuron, integration neuron and differential neuron; Described output layer comprises at least two sub-output layers, and each sub-output layer comprises an output layer neuron; Described sub-input layer, son hide layer and sub-output layer is corresponding one by one;

Described steps A 1 comprises:

First input layer of each sub-input layer receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, second input layer of each sub-input layer receive input from the detected temperature value of described inspection positions, and export to ratio neuron, integration neuron and the differential neuron that corresponding son is hidden layer respectively;

Described steps A 2 comprises:

Ratio neuron, integration neuron, differential neuron that each son is hidden layer respectively according to the desired temperature of input, detected temperature value, self with corresponding sub-input layer first input layer between weights, self be weighted calculating with weights between corresponding sub-input layer second input layer, with the result that calculates respectively as self current input value; Again that self is the current input value of ratio neuron that each son is hidden layer multiply by the neuronic output valve of preset proportional value acquisition ratio, and exports to all output layer neurons; Each son is hidden the integration neuron of layer input value that self is current and the last output valve sum output valve as self, and exports to all output layer neurons; Each son is hidden the differential neuron of the layer input value that self is current and last input value sum as self current output valve, and exports to all output layer neurons;

Described steps A 3 comprises:

The output layer neuron basis of each sub-output layer is from all proportions neuron, integration neuron, the neuronic value of differential, and self respectively and the weights between all proportions neuron, integration neuron and the differential neuron be weighted calculating, obtain the controlled quentity controlled variable of self pairing inspection positions.

In the such scheme, before the described steps A 1, this method further comprises:

X1, according to the desired temperature that receives, judge whether to reach the learning objective value that sets in advance from least two detected temperature values of inspection positions at least two inspection positions of semiconductor equipment reaction chamber, if do not reach, execution in step X2 then; Otherwise, execution in step A1;

X2, according to predefined input layer to hiding the weights of layer study step-length between upgrading the input layer preserved in advance and hiding layer, upgrade prior hiding layer of preserving and the weights between the output layer according to predefined hiding layer to output layer study step-length, return step X1.

In sum, the present invention proposes a kind of apparatus and method of controlling semiconductor equipment reaction chamber temperature, can be according at the desired temperature of at least two inspection positions in the semiconductor equipment reaction chamber, from the detected temperature value of described at least two inspection positions, and utilize the ANN (Artificial Neural Network) Control algorithm computation to go out controlled quentity controlled variable at each inspection positions.Like this, just the coupled relation of semiconductor equipment reaction chamber each several part thermal perturbation can be embodied in the neural network control method of multiple-input and multiple-output, overcome the defective that single control loop is controlled semiconductor equipment reaction chamber each several part temperature independently, thereby better reaction chamber temperature is controlled.

Description of drawings

Fig. 1 is a typical semiconductor equipment reaction chamber synoptic diagram;

Fig. 2 is the device synoptic diagram that prior art is controlled the temperature of a part of semiconductor equipment reaction chamber;

Fig. 3 a is a kind of inner structure synoptic diagram of temperature control modules in the prior art;

Fig. 3 b is the another kind of inner structure synoptic diagram of temperature control modules in the prior art;

Fig. 4 is the device synoptic diagram that prior art is controlled the temperature of a plurality of parts of semiconductor equipment reaction chamber;

Fig. 5 is the apparatus structure synoptic diagram that the present invention controls semiconductor equipment reaction chamber temperature;

Fig. 6 is a kind of inner structure synoptic diagram of temperature control algorithm module among the present invention;

Fig. 7 is based on the synoptic diagram of the ANN (Artificial Neural Network) Control algorithm of PID in the embodiment of the invention;

Fig. 7 b is the another kind of inner structure synoptic diagram of temperature control algorithm module among the present invention;

Fig. 8 is the inner structure synoptic diagram of input layer module in the embodiment of the invention;

Fig. 9 is an inner structure synoptic diagram of hiding layer module in the embodiment of the invention;

Figure 10 is the inner structure synoptic diagram of output layer module in the embodiment of the invention;

Figure 11 is the method flow diagram that the present invention controls semiconductor equipment reaction chamber temperature;

Figure 12 is the method embodiment process flow diagram that the present invention controls semiconductor equipment reaction chamber temperature.

Embodiment

For making the purpose, technical solutions and advantages of the present invention clearer, the present invention is described in further detail below in conjunction with the accompanying drawings and the specific embodiments.

In order to overcome the defective that single control loop is controlled semiconductor equipment reaction chamber each several part temperature independently, the present invention is embodied in the coupled relation of semiconductor equipment reaction chamber each several part temperature in the neural network control method of multiple-input and multiple-output, thereby better reaction chamber temperature is controlled.

Fig. 5 is the apparatus structure synoptic diagram that the present invention controls semiconductor equipment reaction chamber temperature.As shown in Figure 5, this device can comprise: at least two temperature detecting module 501, temperature control algorithm module 502, at least two temperature control modules 503, there is relation one to one in temperature detecting module 501 with temperature control modules 503.

Temperature detecting module 501 is used for the temperature of semiconductor equipment reaction chamber inspection positions is detected, and detected temperature value is exported to temperature control algorithm module 502.

Temperature control algorithm module 502, be used for desired temperature at inspection positions according to all temperature detecting module 501 detected temperature values and all inputs, and utilize the ANN (Artificial Neural Network) Control algorithm computation to go out controlled quentity controlled variable, and export to corresponding temperature control modules 503 at each temperature control modules.

Temperature control modules 503 is used for according to the controlled quentity controlled variable of input the temperature of temperature detecting module inspection positions being controlled.

The value that inputs to temperature control algorithm module 502 has two classes: a class be at least two temperature detecting module 501 from the detected temperature value of semiconductor equipment reaction chamber various piece, another kind of is direct at least two desired temperatures of input.That is to say that the value that inputs to temperature control algorithm module 502 can have a plurality of, also can have a plurality ofly from the value of temperature control algorithm module 502 outputs that what temperature control algorithm module 502 adopted is the ANN (Artificial Neural Network) Control algorithm.

As seen, what at least two temperature detecting module 501 shown in Figure 5, temperature control algorithm module 502, at least two temperature control modules 503 constituted is the control loop of multiple-input and multiple-output, can embody the coupled relation of reaction chamber various piece temperature, thereby better temperature be controlled.

Among the present invention, the ANN (Artificial Neural Network) Control algorithm that temperature control algorithm module 502 can adopt is many, so long as the ANN (Artificial Neural Network) Control algorithm of multiple-input and multiple-output all can.

The ANN (Artificial Neural Network) Control algorithm generally can comprise input layer, hide layer and output layer, and correspondingly, the inner structure of temperature control algorithm module 502 can comprise as shown in Figure 6:

Input layer module 5021 is used for all desired temperatures that receive are exported to a hiding layer module 5022; Receive all temperature detecting module 501 detected temperature values, and export to and hide layer module 5022.

Hide layer module 5022, be used to receive the desired temperature and the described detected temperature value of input layer module 5021 inputs; According to the input layer of first memory module, 5024 preservations and the weights of hiding between the layer desired temperature and detected temperature value are weighted calculating, result of calculation is exported to output layer module 5023 as hiding layer result of calculation.

Output layer module 5023 is used to receive a hiding layer result of calculation; Hiding layer and the weights between the output layer preserved according to second memory module 5025 are weighted calculating to hiding layer result of calculation, obtain the controlled quentity controlled variable at each temperature control modules, and export to corresponding temperature control modules 503.

First memory module 5024 is used to preserve the weights between input layer and the hiding layer.

Second memory module 5025 is used to preserve the weights of hiding between layer and the output layer.

In the practical application, input layer with hide inside annexation and weights between the layer, to hide inside annexation between layer and the output layer and weights relevant with concrete ANN (Artificial Neural Network) Control algorithm.But, generally all be that input layer is exported to hiding layer with the value of receiving no matter be which kind of ANN (Artificial Neural Network) Control algorithm; The value that hiding layer will be imported is weighted calculating, and exports to output layer; The value that output layer will be imported is weighted calculating, with the result output of result of calculation as the ANN (Artificial Neural Network) Control algorithm.

The present invention program be described in detail below with preferred embodiment in order to illustrate better.

Present embodiment utilization is come the constructing neural network control algolithm based on proportional-integral-differential (PID) algorithm.For better explanation present embodiment scheme, below earlier the ANN (Artificial Neural Network) Control algorithm based on PID is described in detail.

Fig. 7 is based on the synoptic diagram of the ANN (Artificial Neural Network) Control algorithm of PID in the present embodiment.As shown in Figure 7, the present embodiment algorithm can have m part of coupled relation at temperature, and each part comprises 2 inputs and 1 output.Such as: for first, comprise u ₁₁And u ₁₂These two inputs also comprise

This output; For second portion, comprise u ₂₁And u ₂₂These two inputs also comprise

This output, the rest may be inferred for other parts.No matter be which part, can be divided into three levels from being input to output, i.e. input layer, hiding layer and output layer.The input layer of each part comprises two input layers, and the hiding layer of each part comprises ratio neuron P, integration neuron and differential neuron I, and the output layer of each part comprises an output layer neuron D.

In the present embodiment, the external world imports two class values to PID ANN (Artificial Neural Network) Control algorithm, and a class is predefined input value r _d, as each desired temperature of input, d=1 wherein, 2,3...m; Another kind of is the y as a result that controlling object is detected _d, as the temperature of reaction chamber each several part, d=1 wherein, 2,3...m.This two classes input value will be respectively as two input values of each part of PID ANN (Artificial Neural Network) Control algorithm.Such as: the r of extraneous input ₁Just can be used as the u of first's input layer ₁₁This input value, the y of extraneous input ₁Just can be used as the u of first's input layer ₁₂This input value, the rest may be inferred for the situation of other parts.

In the present embodiment, the input/output relation in the PID neural network input layer can be represented with formula one:

x _Di(k)=u _Di(k) i=1,2 d=1,2..., m formula one

Wherein, i represents the sequence number of each part input layer, d represent at the part sequence number, u _Di(k) input value of the corresponding input layer of expression, x _Di(k) output valve of the corresponding input layer of expression.That is to say that the input layer of each part is not done any change to the input value of self, but, export to corresponding ratio neuron P, integration neuron and the differential neuron I that hides layer respectively directly with the output valve of input value as self.

In the present embodiment, the PID neural network is hidden each neuronic input value of layer and can be represented with formula two:

$u_{\mathrm{dj}}^{\′}\left(k\right)=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{w}_{\mathrm{dij}}{x}_{\mathrm{di}}\left(k\right)$ I=1,2 j=1,2,3 d=1,2,3...m formula two

Wherein, i represents the sequence number of each part input layer; J represents the sequence number of every part hidden layer neuron, such as: 1 expression ratio neuron, 2 expression integration neurons, 3 expression differential neurons; D represent at the part sequence number; w _DijRepresent that i input layer of d part and d partly hide the weights between j neuron of layer; x _Di(k) represent the value that i input layer of d part exported; Represent that d partly hides j input value that neuron obtained of layer.Such as: the value of the 1st input layer output of part 1 is x ₁₁(k), the value of the 2nd input layer output of part 1 is x ₁₂(k), the weights between the ratio neuron of the hiding layer of the 1st input layer of part 1 and part 1 are w ₁₁₁, the weights between the ratio neuron of the hiding layer of the 2nd input layer of part 1 and part 1 are w ₁₂₁, so, the neuronic input value of ratio that part 1 is hidden layer just should for $u_{11}^{\′}\left(k\right)=w_{11}{x}_{11}\left(k\right)+w_{121}{x}_{12}\left(k\right).$ Can analogize according to above-mentioned example as for hiding other neuronic input value of layer, repeat no more herein.

In the present embodiment, the PID neural network is hidden the neuronic output valve of ratio of layer and can be represented with formula three:

$x_{d1}^{\′}\left(k\right)={\mathrm{Ku}}_{d1}^{\′}\left(k\right)$ D=1,2,3...m formula three

Wherein, k represents predefined ratio value,

Represent that d partly hides the neuronic input value of ratio of layer,

Represent that d partly hides the neuronic output valve of ratio of layer.That is to say that the ratio neuron of d part as output valve, and is exported to all output layer neurons with the product of current input value and predefined ratio value.

In the present embodiment, the PID neural network is hidden the neuronic output valve of integration of layer and can be represented with formula four:

$x_{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="A200710179613E00311.png"\; state="\{\{state\}\}">d</figure-callout>}2}^{\&prime;}\left(k\right)={\mathrm{<figure-callout\; id="2"\; label="x\; d"\; filenames="A200710179613E00311.png"\; state="\{\{state\}\}">x</figure-callout>}}_d^2(k-1)+{\mathrm{<figure-callout\; id="2"\; label="u\; d"\; filenames="A200710179613E00311.png"\; state="\{\{state\}\}">u</figure-callout>}}_d^2\left(k\right)$ D=1,2,3...m formula four

Wherein,

Represent that d partly hides the output valve of the integration neuron last time of layer,

Represent that d partly hides the current input value of integration neuron of layer.That is to say that the integration neuron that d partly hides layer as output valve, and is exported to all output layer neurons with self last output valve and current input value sum.

In the present embodiment, the PID neural network is hidden the neuronic output valve of differential of layer and can be represented with formula five:

$x_{d3}^{\&prime;}\left(k\right)=u_{d3}^{\&prime;}\left(k\right)+u_{d3}^{\&prime;}(k-1)$ D=1,2,3...m formula five

Wherein,

Represent that d partly hides the neuronic current input value of differential of layer,

Represent that d partly hides the neuronic last input value of differential of layer.That is to say that the differential neuron that d partly hides layer as output valve, and is exported to all output layer neurons with self current input value and last input value sum.

In the present embodiment, the neuronic input value of PID neural network output layer can be represented with formula six:

$u_h^{\&prime;\&prime;}\left(k\right)=\underset{d=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{\mathrm{djh}}^{\&prime;}{x}_{\mathrm{dj}}^{\&prime;}\left(k\right)$ D=1,2,3...m j=1,2,3 h=1,2,3...m formula six

Wherein, d and h all represent at the part sequence number, j represents the sequence number of hidden layer neuron.For h part output layer neuron,

Represent that d partly hides the weights between j neuron of layer and the h part output layer neuron,

Represent that d partly hides j neuron of layer and exports to the neuronic output valve of h part output layer.Such as: for part 1 output layer neuron, part 1 is hidden the ratio neuron of layer and the weights between the part 1 output layer neuron are

Part 1 is hidden the integration neuron of layer and the weights between the part 1 output layer neuron are

..., m partly hides the differential neuron of layer and the weights between the part 1 output layer neuron and is

And the hiding neuronic output valve of stratum proportion of part 1 is

Part 1 is hidden lamination and is divided neuronic output valve to be

..., m partly hides the neuronic output valve of layer differential and is

So, the input value of part 1 output layer neuron acquisition just can be expressed as:

$u_1^{\&prime;\&prime;}\left(k\right)=w_{111}^{\&prime;}{x}_{11}^{\&prime;}\left(k\right)+w_{121}^{\&prime;}{x}_{12}^{\&prime;}\left(k\right)+...+w_{m31}^{\&prime;}{x}_{m3}^{\&prime;}\left(k\right)$

The input value that other parts output layer neuron obtains can be analogized according to following formula, repeats no more herein.

In the present embodiment, the neuronic input/output relation of PID neural network output layer can be represented with formula seven:

$x_h^{\&prime;\&prime;}\left(k\right)=u_h^{\&prime;\&prime;}\left(k\right)$ H=1,2,3...m formula seven

Wherein,

Represent the neuronic input value of h part output layer,

Represent the neuronic output valve of h part output layer.That is to say that the output layer neuron can be with the input value that obtains directly as output valve.

Above-mentioned formula one to formula seven is present embodiment principles based on the ANN (Artificial Neural Network) Control algorithm of PID, wherein, and the weight w between input layer and the hiding layer _DijAnd the weights between hiding layer and the output layer

Can obtain by study.Such as: w is set in advance _DijWith

Initial value, according to the initial value, the input value r that are provided with _dAnd y _dAnd formula one to formula seven obtains output valve

According to the output valve that obtains

Controlling object is controlled, as reaction chamber each several part temperature is controlled; Again detect the result of controlling object, with testing result y _dAs input value.Like this, if testing result y _dWith predefined r _dGap is bigger, does not reach predefined learning objective value in other words, just can upgrade w _DijWith

And the rest may be inferred, up to reaching predefined learning objective value for extremely.

In order to judge testing result y _dWith predefined r _dGap, the method for calculating its gap can be set in advance, such as: with y _dWith r _dBetween mean square deviation as the method for weighing its difference.y _dWith r _dBetween mean square deviation can represent with formula eight:

$J=\frac{1}{n}\underset{d=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{[r_d\left(k\right)-y_d\left(k\right)]}^2$ D=1,2,3...m k=1,2 ... n formula eight

Wherein, d represent at the part sequence number, k represents sampled point, n represents number of samples.Like this, suppose that the learning objective value is study number of times X, and w is set _DijWith

Initial value, so, obtain w by study _DijWith

Value method can for:

A, according to input value r _d, y _dAnd formula one to formula seven obtains output valve

B, according to output valve

Controlling object is controlled, such as the temperature of reaction chamber each several part is controlled, and regained testing result y _d

C, judge whether to reach study number of times X,, utilize the study step-length refreshing weight w that sets if do not arrive _DijWith

And return step a; Otherwise, withdraw from this flow process.

Repeatedly learn by said method, can finally obtain weight w _DijWith

After this, just can directly utilize the weight w of learning _DijWith

And needn't be to w _DijWith

Upgrade once more.In the practical application, the study step-length can be a fixing step-length, also can be the step-length of a variation, specifically how to be provided with and can be determined voluntarily by the user who uses the present embodiment scheme, repeats no more herein.

According to the principle of above-mentioned present embodiment based on PID ANN (Artificial Neural Network) Control algorithm, present embodiment correspondingly proposes a kind of device of controlling semiconductor equipment reaction chamber temperature.The basic structure of the device of present embodiment is identical with Fig. 5, comprising: 501,1 temperature control algorithm module 502 of m temperature detecting module, a m temperature control modules 503, m is equal to or greater than 2.Wherein, the inner structure of temperature control algorithm module 502 is shown in Fig. 7 b.Fig. 7 b is similar to Fig. 6, comprises input layer module 5021, hides layer module 5022, output layer module 5023, first memory module 5024 and second memory module 5025, also comprises study module 5026.

Wherein, input layer module 5021, hide the inner structure of layer module 5022 and output layer module 5023 can be respectively as 8, Fig. 9 and shown in Figure 10.

As can be seen from Figure 8, input layer module 5021 can comprise m input layer submodule 5021X, and each input layer submodule 5021X comprises the first input layer 5021Xa and the second input layer 5021Xb.Wherein,

The first input layer 5021Xa is used to receive the desired temperature of input, and exports to ratio neuron 5022Ya, integration neuron 5022Yb and the differential neuron 5022Yc of the corresponding straton module 5022Y of hiding respectively.

The second input layer 5021Xb is used to receive corresponding temperature detection module 501 detected temperature values, and exports to ratio neuron 5022Ya, integration neuron 5022Yb and the differential neuron 5022Yc of the corresponding straton module 5022Y of hiding respectively.

As can be seen from Figure 9, hide layer module 5022 and comprise m hiding straton module 5022Y, each hiding straton module 5022Y comprises ratio neuron 5022Ya, integration neuron 5022Yb and differential neuron 5022Yc.Wherein,

Ratio neuron 5022Ya, be used for desired temperature according to the first input layer 5021Xa input of corresponding input layer submodule 5021X, the detected temperature value of second input layer 5021Xb input, weights between the first input layer 5021Xa and the ratio neuron 5022Ya, weights between the second input layer 5021Xb and the ratio neuron 5022Ya are weighted calculating, with the result that calculates as the current input value of ratio neuron 5022Ya, the current input value of ratio neuron 5022Ya be multiply by the output valve that preset proportional value obtains ratio neuron 5022Ya, and export to all output layer neuron 5023Za.

Integration neuron 5022Yb is according to the desired temperature of the first input layer 5021Xa input of corresponding input layer submodule 5021X, the detected temperature value of second input layer 5021Xb input, weights between the first input layer 5021Xa and the integration neuron 5022Yb, weights between the second input layer 5021Xb and the integration neuron 5022Yb are weighted calculating, with the result that calculates as the current input value of integration neuron 5022Yb, the input value that integration neuron 5022Yb is current and last output valve sum be as the current output valve of integration neuron 5022Yb, and export to all output layer neuron 5023Za.

Differential neuron 5022Yc, be used for desired temperature according to the first input layer 5021Xa input of corresponding input layer submodule 5021X, the detected temperature value of second input layer 5021Xb input, weights between the first input layer 5021Xa and the differential neuron 5022Yc, weights between the second input layer 5021Xb and the differential neuron 5022Yc are weighted calculating, with the result that calculates as the current input value of differential neuron 5022Yc, input value that differential neuron 5022Yc is current and last input value sum be as the current output valve of differential neuron 5022Yc, and export to all output layer neuron 5023Za.

As can be seen from Figure 10, the output layer module comprises m output layer submodule 5023Z, and each output layer submodule 5023Z comprises an output layer neuron 5023Za.

Output layer neuron 5023Za, be used for ratio neuron 5022Ya, the integration neuron 5022Yb according to all hiding straton module 5022Y, the value of differential neuron 5022Yc input, and the weights between the weights between the weights between all proportions neuron 5022Ya and the output layer neuron 5023Za, integration neuron 5022Yb and the output layer neuron 5023Za, differential neuron 5022Yc and the output layer neuron 5023Za are weighted calculating, obtain controlled quentity controlled variable, the signal that carries controlled quentity controlled variable is exported to corresponding temperature control modules 503.

In the present embodiment, weights between the first input layer 5021Xa and the ratio neuron 5022Ya, weights between the second input layer 5021Xb and the ratio neuron 5022Ya, weights between the first input layer 5021Xa and the integration neuron 5022Yb, weights between the second input layer 5021Xb and the integration neuron 5022Yb, weights between the first input layer 5021Xa and the differential neuron 5022Yc, weights between the second input layer 5021Xb and the differential neuron 5022Yc are exactly the weights between input layer and the hiding layer, i.e. w _Dij, can be kept in first memory module 5024.Correspondingly, weights between weights between weights between ratio neuron 5022Ya and the output layer neuron 5023Za, integration neuron 5022Yb and the output layer neuron 5023Za, differential neuron 5022Yc and the output layer neuron 5023Za are exactly the weights of hiding between layer and the output layer, promptly

Can be kept in second memory module 5025.

That is to say, for each any input layer submodule 5021X, hiding straton module 5022Y corresponding and corresponding output layer submodule 5023Z with it, the first input layer 5021Xa receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, and exports to ratio neuron 5022Ya, integration neuron 5022Yb and differential neuron 5022Yc; The second input layer 5021Xb receives from the detected temperature value of inspection positions, and exports to ratio neuron 5022Ya, integration neuron 5022Yb and differential neuron 5022Yc respectively; Described ratio neuron 5022Ya, integration neuron 5022Yb and differential neuron 5022Yc are weighted calculating according to the weights between the weights between the desired temperature, detected temperature value of input, self and the first input layer 5021Xa, self and the second input layer 5021Xb respectively, with the result that calculates respectively as self current input value; Again that self is the current input value of described ratio neuron 5022Ya multiply by preset proportional value and obtains the output valve of self, and exports to all output layer neuron 5023Za; Described integration neuron 5022Yb input value that self is current and last output valve sum output valve as self, and export to all output layer neuron 5023Za; Input value that described differential neuron 5022Yc is current with self and last input value sum be as the output valve of self, and export to all output layer neuron 5023Za; Described output layer neuron 5023Za is according to the input value from all proportions neuron 5022Ya, 5022Yb and differential neuron 5022Yc, and self respectively and the weights between all proportions neuron 5022Ya, 5022Yb and the differential neuron 5022Yc be weighted calculating, acquisition is at the controlled quentity controlled variable of the temperature detecting module 501 of self correspondence, and exports to the temperature detecting module 501 of self correspondence.

In addition, present embodiment also comprises a weights study module 5026, is used to receive m desired temperature of input, receives m temperature detecting module 501 detected temperature values; Judge whether to reach the learning objective value that sets in advance according to m desired temperature and m detected temperature value, if do not reach, then upgrade first memory module 5024 input layer of preserving and the weights of hiding between the layer to hiding layer study step-length, upgrade weights between hiding layer of 5025 preservations of second memory module and the output layer to output layer study step-length according to predefined hiding layer according to predefined input layer.Process as for concrete study can repeat no more with reference to following method embodiment herein.

Certainly, same as the prior art, the temperature detecting module 501 in the present embodiment also can comprise temperature inductor, amplifier, mould/number conversion module, and its function and structure are same as the prior art, repeat no more herein.

In the present embodiment, if the signal that carries controlled quentity controlled variable of temperature control algorithm module 502 outputs is a pulse signal, so, temperature control modules 503 can comprise: pulse-width modulator, solid-state relay and well heater.If the signal that carries controlled quentity controlled variable of temperature control algorithm module 502 outputs is a digital signal, temperature control modules 503 can comprise: D/A switch module, voltage regulator and well heater.No matter which kind of signal 502 outputs of temperature control algorithm module is, the inner structure of temperature control modules 503 and function can be same as the prior art, repeat no more herein.

At said apparatus, the present invention also proposes a kind of method of controlling semiconductor equipment reaction chamber temperature.Figure 11 is the method flow diagram that the present invention controls semiconductor equipment reaction chamber temperature.As shown in figure 11, the present invention includes:

Step 1101: according to the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, from least two detected temperature values of inspection positions, and the ANN (Artificial Neural Network) Control algorithm computation goes out the controlled quentity controlled variable at each inspection positions.

If the ANN (Artificial Neural Network) Control algorithm comprises input layer, hides layer and output layer, so, the method for this step calculation control amount can specifically comprise:

A1: input layer receives the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, receives the detected temperature value of described at least two inspection positions, and exports to and hide layer.

A2: hide layer and described desired temperature and detected temperature value are weighted calculating, result of calculation is exported to output layer as hiding layer result of calculation according to the weights between input layer of preserving and the hiding layer.

A3: output layer is weighted calculating according to the weights between hiding layer of preserving and the output layer to hiding layer result of calculation, obtains the controlled quentity controlled variable at each inspection positions.

Step 1102: respectively the temperature of each inspection positions of semiconductor equipment reaction chamber is controlled according to the controlled quentity controlled variable that calculates.

In the practical application, input layer and the weights of hiding between the weights between the layer, hiding layer and the output layer can obtain by study in advance.In this case, before the step 1101, can further include:

X1, according to the desired temperature that receives, judge whether to reach the learning objective value that sets in advance from least two detected temperature values of inspection positions at least two inspection positions of semiconductor equipment reaction chamber, if do not reach, execution in step X2 then; Otherwise, execution in step 1101.

X2, according to predefined input layer to hiding the weights of layer study step-length between upgrading the input layer preserved in advance and hiding layer, upgrade prior hiding layer of preserving and the weights between the output layer according to predefined hiding layer to output layer study step-length, return step X1.

Study step-length described here can also can be the step-length of a variation for fixing step-length.

For the weight w between input layer and the hiding layer _Dij, suppose that input layer is redefined for η to hiding layer study step-length _DijIf adopt fixed step size to come refreshing weight w _Dij, so, before upgrading and the relation after upgrading can be expressed as:

w _Dij(l+1)=w _Dij(l)-η _DijFormula nine

Wherein, w _DjhWeights before upgrading when (l) expression is learnt at every turn, w _DjhWeights after upgrading when (l+1) expression is learnt at every turn.

Certainly, if the step-length that adopt to change, so, before upgrading and the relation after upgrading can be expressed as:

$w_{\mathrm{dij}}(l+1)=w_{\mathrm{dij}}\left(l\right)-{\mathrm{\&eta;}}_{\mathrm{dij}}\frac{\&PartialD;J}{{\&PartialD;w}_{\mathrm{dij}}}$ Formula ten

Wherein, from formula eight as can be known,

Be a value that changes according to actual conditions, so

Just can be regarded as the study step-length of a variation.

Similarly, for the weights of hiding between layer and the output layer , suppose to hide layer a study step-length and be redefined for to output layer

If adopt fixed step size to come refreshing weight

So, upgrade before and upgrade after relation can be expressed as:

$w_{\mathrm{djh}}^{\&prime;}(l+1)=w_{\mathrm{djh}}^{\&prime;}\left(l\right)-{\text{\&eta;}}_{\mathrm{djh}}^{\&prime;}$ Formula 11

Wherein,

Weights before upgrading when expression is learnt at every turn,

Weights after upgrading when expression is learnt at every turn.

Certainly, if the step-length that adopt to change, so, before upgrading and the relation after upgrading can be expressed as:

$w_{\mathrm{djh}}^{\&prime;}(l+1)=w_{\mathrm{djh}}^{\&prime;}\left(l\right)-{\mathrm{\&eta;}}_{\mathrm{djh}}^{\&prime;}\frac{\&PartialD;J}{\&PartialD;w_{\mathrm{djh}}^{\&prime;}}$ Formula 12

Wherein, from formula eight as can be known, Be a value that changes according to actual conditions, so

Just can be regarded as the study step-length of a variation.

For the present embodiment method is described better, be elaborated with a preferred embodiment below.

Present embodiment adopts and based on the ANN (Artificial Neural Network) Control algorithm of PID reaction chamber each several part temperature is controlled, and its principle can be with reference to the introduction of above-mentioned PID ANN (Artificial Neural Network) Control algorithm.In the present embodiment, suppose that input layer comprises at least two sub-input layers, each sub-input layer comprises first input layer and second input layer; Hiding layer comprises the hiding layer of at least two sons, and each son is hidden layer and comprised ratio neuron, integration neuron and differential neuron; Output layer comprises at least two sub-output layers, and each sub-output layer comprises an output layer neuron.Sub-input layer described here, son hide layer and sub-output layer is one to one.

The method of present embodiment can comprise as shown in figure 12:

Step 1201: first input layer of each sub-input layer receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, second input layer of each sub-input layer receive input from the detected temperature value of described inspection positions, and export to ratio neuron, integration neuron and the differential neuron that corresponding son is hidden layer respectively.

In this step, described detected temperature value is by the various piece inspection positions detected temperature value of temperature detecting module from reaction chamber.

Step 1202: ratio neuron that each son is hidden layer according to the desired temperature of input, detected temperature value, self with corresponding sub-input layer first input layer between weights, self be weighted calculating with weights between corresponding sub-input layer second input layer, with the result that calculates respectively as self current input value, the input value that self is current multiply by preset proportional value and obtains the output valve of self, and exports to all output layer neurons.

Step 1203: integration neuron that each son is hidden layer according to the desired temperature of input, detected temperature value, self with corresponding sub-input layer first input layer between weights, self be weighted calculating with weights between second input layer of corresponding sub-input layer, with the result that calculates as self current input value, current input value and last output valve sum as self current output valve, and are exported to all output layer neurons.

Step 1204: differential neuron that each son is hidden layer according to the desired temperature of input, detected temperature value, self with corresponding sub-input layer first input layer between weights, self be weighted calculating with weights between corresponding sub-input layer second input layer, with the result that calculates as self current input value, input value that self is current and last input value sum be as self current output valve, and export to all output layer neurons.

Step 1205: the output layer neuron of each sub-output layer is according to the value of all proportions neuron, integration neuron, the input of differential neuron, and self respectively and the weights between all proportions neuron, integration neuron and the differential neuron be weighted calculating, obtain self controlled quentity controlled variable at each inspection positions.

Step 1206: respectively the temperature of each inspection positions of semiconductor equipment reaction chamber is controlled according to the controlled quentity controlled variable that calculates.

In the practical application, be pulse signal if carry the signal of controlled quentity controlled variable, then step 1206 can be specially: the pulse signal that pulse-width modulator will carry controlled quentity controlled variable amplifies, and exports to solid-state relay; Solid-state relay is controlled the power output of well heater according to the pulse signal of input, with the temperature of control semiconductor equipment reaction chamber.

If carrying the signal of controlled quentity controlled variable is digital signal, then step 1206 can be specially: the digital signal that the D/A switch module will carry controlled quentity controlled variable is converted to the module by signal that carries controlled quentity controlled variable, and exports to voltage regulator; Voltage regulator carries out correspondent voltage according to the simulating signal that carries controlled quentity controlled variable to be regulated, and the power output of well heater is controlled, with the temperature of control semiconductor equipment reaction chamber.

Use the embodiments of the invention scheme, owing to the coupled relation of semiconductor equipment reaction chamber each several part temperature is embodied in the neural network control method of multiple-input and multiple-output, overcome the defective that single control loop is controlled semiconductor equipment reaction chamber each several part temperature independently, thereby can control the temperature of semiconductor equipment reaction chamber better.

In sum, more than be preferred embodiment of the present invention only, be not to be used to limit protection scope of the present invention.Within the spirit and principles in the present invention all, any modification of being done, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1, a kind of device of controlling semiconductor equipment reaction chamber temperature is characterized in that, this device comprises temperature control algorithm module, at least two temperature detecting module, at least two temperature control modules; Described temperature detecting module is corresponding one by one with described temperature control modules;

Described temperature detecting module is used for the temperature of semiconductor equipment reaction chamber inspection positions is detected, and detected temperature value is exported to the temperature control algorithm module;

Described temperature control algorithm module, be used for desired temperature at inspection positions according to the detected temperature value of all temperature detecting module and all inputs, and utilize the ANN (Artificial Neural Network) Control algorithm computation to go out controlled quentity controlled variable, and export to corresponding temperature control modules at each temperature control modules;

Described temperature control modules is used for according to the controlled quentity controlled variable of input the temperature of temperature detecting module inspection positions being controlled.

2, device according to claim 1 is characterized in that, described temperature control algorithm module comprises:

The input layer module is used for all desired temperatures that receive are exported to a hiding layer module; The detected temperature value of all temperature detecting module that receives is exported to a hiding layer module;

Hide layer module, be used to receive desired temperature and detected temperature value from the input layer module; According to the input layer of first memory module preservation and the weights of hiding between the layer desired temperature and detected temperature value are weighted calculating, result of calculation is exported to the output layer module as hiding layer result of calculation;

The output layer module is used to receive a hiding layer result of calculation; Hiding layer and the weights between the output layer preserved according to second memory module are weighted calculating to hiding layer result of calculation, obtain the controlled quentity controlled variable at each temperature control modules, and export to corresponding temperature control modules;

First memory module is used to preserve the weights between input layer and the hiding layer;

Second memory module is used to preserve the weights of hiding between layer and the output layer.

3, device according to claim 2 is characterized in that, described input layer module comprises at least two input layer submodules, and each input layer submodule comprises first input layer and second input layer; Described hiding layer module comprises at least two hiding straton modules, and each hiding straton module comprises ratio neuron, integration neuron and differential neuron; Described output layer module comprises at least two output layer submodules, and each output layer submodule comprises an output layer neuron; Described input layer submodule, hiding straton module and output layer submodule are corresponding one by one;

For any input layer submodule, the hiding straton module corresponding and corresponding output layer submodule with it, described first input layer receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, and exports to ratio neuron, integration neuron and differential neuron; Described second input layer receives from the detected temperature value of described inspection positions, and exports to ratio neuron, integration neuron and differential neuron respectively; Described ratio neuron, integration neuron, differential neuron are weighted calculating according to the weights between the weights between the desired temperature, detected temperature value of input, self and first input layer, self and second input layer respectively, with the result that calculates respectively as self current input value; Again that self is the current input value of described ratio neuron multiply by preset proportional value and obtains the output valve of self, and exports to all output layer neurons; Described integration neuron input value that self is current and last output valve sum output valve as self, and export to all output layer neurons; Input value that described differential neuron is current with self and last input value sum be as the output valve of self, and export to all output layer neurons; Described output layer neuron basis is from all proportions neuron, integration neuron and the neuronic value of differential, and self respectively and the weights between all proportions neuron, integration neuron and the differential neuron be weighted calculating, acquisition is at the controlled quentity controlled variable of the temperature detecting module of self correspondence, and exports to the temperature detecting module of self correspondence.

4, device according to claim 2 is characterized in that, described temperature control algorithm module further comprises:

The weights study module is used to receive the desired temperature of all inputs, receives the detected temperature value of all temperature detecting module; Judge whether to reach the learning objective value that sets in advance according to all desired temperatures and all detected temperature values, if do not reach, then upgrade first memory module input layer of preserving and the weights of hiding between the layer to hiding layer study step-length, upgrade weights between hiding layer of the preservation of second memory module and the output layer to output layer study step-length according to predefined hiding layer according to predefined input layer.

5, a kind of method of controlling semiconductor equipment reaction chamber temperature is characterized in that, this method comprises:

According to the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, from the detected temperature value of described at least two inspection positions, and the ANN (Artificial Neural Network) Control algorithm computation goes out the controlled quentity controlled variable at each inspection positions; Respectively the temperature of each inspection positions of semiconductor equipment reaction chamber is controlled according to the controlled quentity controlled variable that calculates.

6, method according to claim 5 is characterized in that, the method for described calculation control amount comprises:

A1, input layer receive the desired temperature at least two inspection positions of semiconductor equipment reaction chamber of input, receive the detected temperature value of described at least two inspection positions, and export to and hide layer;

A2, hiding layer are weighted calculating according to the weights between input layer of preserving and the hiding layer to described desired temperature and detected temperature value, and result of calculation is exported to output layer as hiding layer result of calculation;

A3, output layer are weighted calculating according to the weights between hiding layer of preserving and the output layer to hiding layer result of calculation, obtain the controlled quentity controlled variable at each inspection positions.

7, method according to claim 6 is characterized in that, described input layer comprises at least two sub-input layers, and each sub-input layer comprises first input layer and second input layer; Described hiding layer comprises the hiding layer of at least two sons, and each son is hidden layer and comprised ratio neuron, integration neuron and differential neuron; Described output layer comprises at least two sub-output layers, and each sub-output layer comprises an output layer neuron; Described sub-input layer, son hide layer and sub-output layer is corresponding one by one;

Described steps A 1 comprises:

First input layer of each sub-input layer receives the desired temperature at an inspection positions of semiconductor reaction chamber of input, second input layer of each sub-input layer receive input from the detected temperature value of described inspection positions, and export to ratio neuron, integration neuron and the differential neuron that corresponding son is hidden layer respectively;

Described steps A 2 comprises:

Ratio neuron, integration neuron, differential neuron that each son is hidden layer respectively according to the desired temperature of input, detected temperature value, self with corresponding sub-input layer first input layer between weights, self be weighted calculating with weights between corresponding sub-input layer second input layer, with the result that calculates respectively as self current input value; Again that self is the current input value of ratio neuron that each son is hidden layer multiply by the neuronic output valve of preset proportional value acquisition ratio, and exports to all output layer neurons; Each son is hidden the integration neuron of layer input value that self is current and the last output valve sum output valve as self, and exports to all output layer neurons; Each son is hidden the differential neuron of the layer input value that self is current and last input value sum as self current output valve, and exports to all output layer neurons;

Described steps A 3 comprises:

The output layer neuron basis of each sub-output layer is from all proportions neuron, integration neuron, the neuronic value of differential, and self respectively and the weights between all proportions neuron, integration neuron and the differential neuron be weighted calculating, obtain the controlled quentity controlled variable of self pairing inspection positions.

8, method according to claim 6 is characterized in that, before the described steps A 1, this method further comprises:

X1, according to the desired temperature that receives, judge whether to reach the learning objective value that sets in advance from least two detected temperature values of inspection positions at least two inspection positions of semiconductor equipment reaction chamber, if do not reach, execution in step X2 then; Otherwise, execution in step A1;

X2, according to predefined input layer to hiding the weights of layer study step-length between upgrading the input layer preserved in advance and hiding layer, upgrade prior hiding layer of preserving and the weights between the output layer according to predefined hiding layer to output layer study step-length, return step X1.

Images