CN102759430A - BP (Back Propagation) neural network based high-precision correction and test method for resonance cylinder pressure sensor - Google Patents

Abstract

The invention belongs to the technical field of test and measurement and particularly relates to a BP (Back Propagation) neural network based high-precision correction and test method for a resonance cylinder pressure sensor, aiming to increase the correction and test precision for the resonance cylinder pressure sensor. The method comprises the steps of: structuring a sensor BP neural network of a dual-implicit-strata network structure, wherein input variables of the network structure are respectively the output period T and the temperature voltage V of the sensor, and output variables of the network structure is the pressure value P; acquiring output periods and temperature voltages of the sensor at different temperature under different pressure input conditions; and collecting output quantities of the sensor at different temperature under different pressure conditions to serve as correction and test as well as inspection sample points. According to the method, with the dual-implicit-strata network structure, the number of network parameters is reduced while the output precision is guaranteed; and the correction and test precision of the resonance cylinder pressure sensor is increased by 25%, and the high-precision correction and test of the sensor is realized.

Description

Resonant vibration barrel pressure pickup high precision school method for testing based on the BP neural network

Technical field

The invention belongs to a kind of thermometrically technical field, relate to resonant vibration barrel pressure pickup high precision school method for testing, be specifically related to a kind of resonant vibration barrel pressure pickup high precision school method for testing based on the BP neural network.

Background technology

For realizing the accurate control of guided missile, improve day by day for tonometric range and accuracy requirement, therefore need the development high-precision pressure sensor.To present good stability, resonant vibration barrel pressure pickup that precision is high; The level of high precision school examination technology becomes the key factor that influences sensor accuracy; The polynomial fitting method that in the past adopted usually based on the sensor physical model; Vibration cylinder pressure transducer output periodic signal T and temperature voltage signal V, according to the sensor physical model, can set up following nonlinear multivariable regression equation between they and the input measurement pressure P:

$P-{\mathrm{<figure-callout\; id="0"\; label="P"\; filenames="HDA00001825131900012.png"\; state="\{\{state\}\}">P</figure-callout>}}_0=\underset{i=0}{\overset{m}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{n}{\mathrm{\&Sigma;}}}{K}_{\mathrm{Ij}}{(T-T_0)}^i{(V-V_0)}^j$ Formula (1)

In the formula: P---input measurement air pressure;

T---the sensor output cycle;

V---baroceptor temperature voltage;

P ₀---RP air pressure;

T ₀---the RP sensor output cycle;

V ₀---the RP temperature voltage;

K _Ij---the characteristic coefficient of sensor.

All a different set of K can be arranged for each vibration cylinder pressure transducer _IjSystem of parameters.

Traditional fitting of a polynomial technology is on the basis of formula (1), selects different polynomial expression orders for use, utilizes least square method to obtain K _IjCoefficient.This method can not satisfy the accuracy requirement of model 0.028%FS on fitting precision.

Summary of the invention

The objective of the invention is to improve resonant vibration barrel pressure pickup school examination precision, a kind of resonant vibration barrel pressure pickup high precision school method for testing based on the BP neural network is provided.

The technical scheme that the present invention adopted is:

A kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network comprises the sensor BP neural network that makes up two latent layer network structures, and the input variable that makes network structure is the output cycle T and the temperature voltage V of sensor, and output variable is a pressure value P; Output cycle and the temperature voltage of pick-up transducers under different temperatures, different pressures initial conditions; Gather different temperatures and pressure condition lower sensor output quantity as school examination and test samples point.

Aforesaid a kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network; Wherein: in the said sensor BP neural network; Utilize tanh S type transport function as the transport function between input layer and first middle layer, first middle layer and second middle layer, utilize linear transfer function as the transport function between second middle layer and the output layer.

Aforesaid a kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network, wherein: when said sample point is gathered, in-45 ℃ to 80 ℃ scopes, choose a plurality of temperature spots, each temperature spot is chosen a plurality of sampled points.

Aforesaid a kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network; Wherein: in the said sensor BP neural network; The network weight of input layer and first middle layer node is a k-factor; The network weight of first middle layer node and second middle layer node is the W coefficient, and the network weight of second middle layer node and output layer node is the s coefficient; And the node number in each middle layer is 4; And tanh S type transport function is following:

$y_i=\frac{2}{1+e^{-2(\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}{k}_{\mathrm{ij}}{x}_j-{\mathrm{\&theta;}}_i)}}-1(i=\mathrm{1,2}...4;j=\mathrm{1,2})$

$Y_m=\frac{2}{1+e^{-2(\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}{W}_{\mathrm{mi}}{y}_i-{\mathrm{\&omega;}}_m)}}-1(m=\mathrm{1,2}...4;i=\mathrm{1,2}...4)$

Wherein: y _iOutput for first middle layer; x _jBe the input variable value; k _IjNetwork weight for input layer and first middle layer; θ _iThreshold value for first each node of middle layer;

Y _mBe the output in second middle layer; W _MiNetwork weight for first middle layer and second middle layer; ω _mIt is the threshold value of second each node of middle layer;

Linear transfer function is following:

$O=\underset{n=1}{\overset{4}{\mathrm{\&Sigma;}}}{s}_{\ln }{Y}_n-{\mathrm{\&beta;}}_l(l=1;n=\mathrm{1,2}...4)$

Wherein: O is model output; s _LnIt is the network weight of second middle layer and output layer; β _lThreshold value for the output layer node.

Aforesaid a kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network; Wherein: adopt the training method of fast algorithm as the BP neural network; Threshold value and weights to each layer in the training are revised, and make error function descend along the negative gradient direction.

The invention has the beneficial effects as follows:

School provided by the invention method for testing has used two latent layer network structures, when guaranteeing output accuracy, has reduced network parameter quantity; Adopt the method for ground training, effectively controlled the reliability of training process; Through solidifying the initialization training parameter, solved the problem reproduced of training process; And made up S type fitting function, and strengthened the approximation capability of matched curve, improved fitting precision than fitting of a polynomial.

The present invention has improved 25% through making up the BP neural network model with resonant vibration barrel pressure pickup school examination precision, has solved sensor high precision school and may well ask and topic can satisfy guided missile accurate pressure measurement demand.

Description of drawings

Fig. 1 is the sensor BP network model figure that the present invention adopts;

Fig. 2 is a pressure error of fitting comparison diagram.

Embodiment

Below in conjunction with accompanying drawing and embodiment a kind of resonant vibration barrel pressure pickup high precision school method for testing based on the BP neural network provided by the invention is introduced:

A kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network comprises the sensor BP neural network that makes up two latent layer network structures, and the input variable that makes network structure is the output cycle T and the temperature voltage V of sensor, and output variable is a pressure value P; Output cycle and the temperature voltage of pick-up transducers under different temperatures, different pressures initial conditions; Gather different temperatures and pressure condition lower sensor output quantity as school examination and test samples point.

In the sensor BP neural network; Utilize tanh S type transport function as the transport function between input layer and first middle layer, first middle layer and second middle layer, utilize linear transfer function as the transport function between second middle layer and the output layer.

When sample point is gathered, in-45 ℃ to 80 ℃ scopes, choose a plurality of temperature spots, each temperature spot is chosen a plurality of sampled points.

The present invention adopts the BP neural network to make up resonant vibration barrel pressure pickup school die trial type, and particular content is following:

At first carry out the sample data collection, gather output cycle and the temperature voltage of pressure transducer under different temperatures, different pressures initial conditions.For example the test temperature point is chosen 14 temperature spots in-45 ℃ to the 80 ℃ scopes; Under each temperature spot; Pressure measurement point is chosen from high to low respectively, and pressure amounts to 35 spot pressures; Treat to write down test figure after the sensor stable output signal, making wherein, a part is the test samples point for school examination sample point, a part.The sample number of gathering should be as much as possible.

In sensor BP neural network makes up, select two latent layer network structures, the input variable that makes network structure is the output cycle T and the temperature voltage V of sensor, output variable is a pressure value P; Variable T, V, P are treated to variable X 1, X2, O through normalization, and figure is as shown in Figure 1 for the BP network model.

Because confirm that the number of plies in middle layer and the node number in each middle layer are very complicated problems, the guidance on the gear shaper without theoretical still often needs rule of thumb and repeatedly tests to confirm always.The present invention is through data simulation repeatedly, confirm the middle layer number of plies be preferably 2 and the node number in each middle layer be preferably at 4 o'clock (for example, 2,4,4,1 structures), fitting effect is better, also can adopt other numerical value.

In Fig. 1, the network weight of input layer and first middle layer node is a k-factor, and the network weight of first middle layer node and second middle layer node is the W coefficient, and the network weight of second middle layer node and output layer node is the s coefficient.Network model utilizes tanh S type transport function as the transport function between input layer and first middle layer, first middle layer and second middle layer, and formula (2) and formula (3) are seen in the function design:

$y_i=\frac{2}{1+e^{-2(\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}{k}_{\mathrm{Ij}}{x}_j-{\mathrm{\&theta;}}_i)}}-1(i=\mathrm{1,2}...4;j=\mathrm{1,2})$ Formula (2)

$Y_m=\frac{2}{1+e^{-2(\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}{W}_{\mathrm{Mi}}{y}_i-{\mathrm{\&omega;}}_m)}}-1(m=\mathrm{1,2}...4;i=\mathrm{1,2}...4)$ Formula (3)

Wherein: y _iOutput for first middle layer; x _jBe the input variable value; k _IjNetwork weight for input layer and first middle layer; θ _iThreshold value for first each node of middle layer.Y _mBe the output in second middle layer; W _MiNetwork weight for first middle layer and second middle layer; ω _mIt is the threshold value of second each node of middle layer.

Choose linear transfer function as the transport function between second middle layer and the output layer, formula (4) is seen in the function design:

$O=\underset{n=1}{\overset{4}{\mathrm{\&Sigma;}}}{s}_{\mathrm{Ln}}{Y}_n-{\mathrm{\&beta;}}_l(l=1;n=\mathrm{1,2}...4)$ Formula (4)

Wherein: O is model output; Can obtain output pressure P through anti-normalization processing; s _LnIt is the network weight of second middle layer and output layer; β _lThreshold value for the output layer node.

Utilize sample data to carry out model training based on the neural network tool platform, threshold value and weights to each layer in the training are revised, and make error function descend along the negative gradient direction.Train the value that obtains network parameter through sample data, wherein certain training result is following:

k _ij=

-3.19283709 -7.16055966

3.07537706 -0.04163307

6.06282238 -9.60802625

-0.99771530 6.42799254

θ _i=

12.46850700

-3.48749525

-2.23594779

-1.30846304

W _mi=

0.06106101 -2.72750488?0.00198993 0.16069464

0.03575073 1.17094160 0.02075128 0.08536263

-0.23359155?0.61221741 1.11657418 0.15811123

0.41401469 0.41214660 1.96806393 0.08391720

ω _m=

2.75440664

1.79626179

-0.94786308

1.49824852

s _ln=

-0.02919945 0.06017406 0.05557906 0.00022849

β _l=

0.88376202

Preferably adopt fast algorithm (LM algorithm) as the training method of BP neural network and progressively with parameter adjustments such as e-learning rate, momentum term, iterations and targets to optimum condition.Utilize sample data to carry out model training based on MATLAB network tool platform, threshold value and weights to each layer in the training are revised, and make error function descend along the negative gradient direction.Obtain the value of each network parameter through the sample data training.

In order to compare the precision of BP neural network model of fit and fitting of a polynomial model, calculate 2 times of residual standard deviation 2S of test sample _P, the fitting of a polynomial model is 0.250kPa, and BP neural network model of fit is 0.188kPa, and it is about 25% that fitting precision has improved, as shown in Figure 2.Fig. 2 is different temperatures environment downforce measuring error contrast effect figure, and solid line is represented the fitting of a polynomial pressure measurement errors, and dotted line is represented the neural network model measuring error.

The present invention selects for use the BP neural network to make up resonant vibration barrel pressure pickup school die trial type the examination in the sensor school of artificial intelligence Application of Neural Network, realizes the output of sensor high-precision pressure, satisfies the accurate demand for control of guided missile.

Claims (5)

1. the resonant vibration barrel pressure pickup school method for testing based on the BP neural network comprises the sensor BP neural network that makes up two latent layer network structures, and the input variable that makes network structure is the output cycle T and the temperature voltage V of sensor, and output variable is a pressure value P; Output cycle and the temperature voltage of pick-up transducers under different temperatures, different pressures initial conditions; Gather different temperatures and pressure condition lower sensor output quantity as school examination and test samples point.

2. a kind of resonant vibration barrel pressure pickup school method for testing according to claim 1 based on the BP neural network; It is characterized in that: in the said sensor BP neural network; Utilize tanh S type transport function as the transport function between input layer and first middle layer, first middle layer and second middle layer, utilize linear transfer function as the transport function between second middle layer and the output layer.

3. a kind of resonant vibration barrel pressure pickup school method for testing based on the BP neural network according to claim 2 is characterized in that: when said sample point is gathered, in-45 ℃ to 80 ℃ scopes, choose a plurality of temperature spots, each temperature spot is chosen a plurality of sampled points.

4. a kind of resonant vibration barrel pressure pickup school method for testing according to claim 3 based on the BP neural network; It is characterized in that: in the said sensor BP neural network; The network weight of input layer and first middle layer node is a k-factor; The network weight of first middle layer node and second middle layer node is the W coefficient, and the network weight of second middle layer node and output layer node is the s coefficient; And the node number in each middle layer is 4; And tanh S type transport function is following:

$y_i=\frac{2}{1+e^{-2(\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}{k}_{\mathrm{ij}}{x}_j-{\mathrm{\&theta;}}_i)}}-1(i=\mathrm{1,2}...4;j=\mathrm{1,2})$

$Y_m=\frac{2}{1+e^{-2(\underset{i=1}{\overset{4}{\mathrm{\&Sigma;}}}{W}_{\mathrm{mi}}{y}_i-{\mathrm{\&omega;}}_m)}}-1(m=\mathrm{1,2}...4;i=\mathrm{1,2}...4)$

Wherein: y _iOutput for first middle layer; x _jBe the input variable value; k _IjNetwork weight for input layer and first middle layer; θ _iThreshold value for first each node of middle layer;

Y _mBe the output in second middle layer; W _MiNetwork weight for first middle layer and second middle layer; ω _mIt is the threshold value of second each node of middle layer;

Linear transfer function is following:

$O=\underset{n=1}{\overset{4}{\mathrm{\&Sigma;}}}{s}_{\ln }{Y}_n-{\mathrm{\&beta;}}_l(l=1;n=\mathrm{1,2}...4)$

Wherein: O is model output; s _LnIt is the network weight of second middle layer and output layer; β _lThreshold value for the output layer node.

5. a kind of resonant vibration barrel pressure pickup school method for testing according to claim 4 based on the BP neural network; It is characterized in that: adopt the training method of fast algorithm as the BP neural network; Threshold value and weights to each layer in the training are revised, and make error function descend along the negative gradient direction.

Images