CN103411628A - Processing method for random drift error of MEMS gyroscope - Google Patents

Abstract

The invention provides a processing method for a random drift error of an MEMS gyroscope. The method comprises the following steps: determining a RBF neural network structure; then acquiring a learning sample; optimizing and training the RBF neural network structure by using the learning sample through genetic algorithm (GA); and finally calculating angular velocity data after suppression of the random drift error. According to the invention, directed at the random drift error of the MEMS gyroscope, real time mean algorithm is employed for suppression of the random drift error, and calculation step length of real time mean algorithm is controlled by using the RBF neural network structure optimized based on GA; modeling of the random drift error is not needed, computational complexity is small, and real-time suppression of the random drift error of the MEMS gyroscope can be conveniently realized.

Description

A kind of disposal route of MEMS Gyroscope Random Drift Error

Technical field

The invention belongs to the inertial technology field, particularly a kind of disposal route of Gyroscope Random Drift Error.

Background technology

In recent years, microelectromechanical systems (MEMS) gyro, as the very important branch in one, inertia field, has obtained significant progress.Due to its characteristic that cost is low, size is little, lightweight and reliability is high, in low-cost inertia system, obtained application more and more widely.But MEMS gyroscope performance is also lower at present, how to reduce the gyrostatic drift error of MEMS, Random Drift Error especially wherein, become the key that improves MEMS gyroscope precision.In order to improve the gyrostatic precision of MEMS, be to improve constantly the gyrostatic device precision of MEMS on the one hand, but due to the restriction that is subject to manufacturing process, improving fast in a short time the device precision is to be difficult to realize; Be exactly to set up rational Random Drift Error transaction module on the other hand, suppress in real time the gyrostatic Random Drift Error of MEMS.The gyrostatic Random Drift Error of existing MEMS is processed basic ideas and is: the model error model, then adopt certain filtering technique it is estimated and compensate.But there is following problem in this thinking: 1. because the random character of the gyrostatic Random Drift Error of MEMS is difficult to set up rational error model; 2. employing filtering technique, calculated amount is larger, and the real-time while using the MEMS gyroscope can be restricted.

Summary of the invention

In order to overcome the deficiencies in the prior art, the invention provides a kind of disposal route of MEMS Gyroscope Random Drift Error, adopt real-time averaging method to suppress the gyrostatic Random Drift Error of MEMS, utilization is based on the step-length of the calculating of the real-time averaging method of RBF ANN (Artificial Neural Network) Control of genetic algorithm optimization, calculated amount is little, can conveniently realize that the real-time Random Drift Error of MEMS gyroscope suppresses.

The technical solution adopted for the present invention to solve the technical problems comprises the following steps:

Step 1: it is single-input single-output that the RBF neural network is set, and input quantity is x=[Δ w], Δ w is MEMS gyroscope angular velocity varies amount, output quantity is y=[Y _Step], Y _StepFor the calculating step-length of real-time averaging method, radial basis vector h=[h ₁h ₂H _jH _n] ^T, h wherein _jFor the gaussian basis function, n is the hidden layer unit number,

$h_j=\exp (-\frac{{\left|\right|x-m_j\left|\right|}^2}{2{{\mathrm{\σ}}_j}^2}),j=\mathrm{1,2}\·\·\·n$

In formula, x is the input quantity of RBF neural network, m _jAnd σ _j ²Respectively center and the variance of j hidden layer unit gaussian basis function of RBF neural network;

The weight vector W=[w of RBF neural network ₁, w ₂..., w _n] ^T,

The output y=[Y of RBF neural network _Step]=W ^TH; Obtain the RBF neural network initial model of 1-n-1, wherein the center m of gaussian basis function _j, variances sigma _j ²With initial value data after in step 3, initial population being decoded of weight vector W, determine;

Step 2: the MEMS gyroscope is fixed on the single shaft rate table, then exists

In scope, uniformly-spaced give respectively the input of single shaft rate table k angular acceleration, when inputting each angular acceleration for the single shaft rate table, gather the angular velocity of MEMS gyroscope output and the angular velocity of turntable output, obtain k group training sample; Wherein Be the maximum angular acceleration that the MEMS gyroscope can measure, uniformly-spaced described

K is the group number of image data, 20≤k≤50;

Step 3: RBF neural network initial model training step 1 obtained with the training sample that step 2 obtains, and adopt genetic algorithm to be optimized center, variance and the hidden layer of the gaussian basis function of the RBF neural network connection weights to output layer, finally obtain optimum RBF neural network, specifically comprise the following steps:

3.1: chromosome adopts binary coding, and each chromosomal binary coding includes the center m of n gaussian basis function _j, a n gaussian basis function variances sigma _j ²With the connection weights W of n hidden layer to output layer _j, j=1,2 ... n; Center m _j, variances sigma _j ², connect weights W _jAll adopt the p bit to mean, a chromosomal total length is 3*p*n, 4≤p≤8;

3.2: the initialization of population generates N initial chromosome, 30≤N≤80;

3.3: the chromosome decoding is converted to respectively decimal number by binary-coded each chromosomal three parts;

3.4: calculate each chromosomal fitness, concrete steps are as follows:

1) will the decode center m of gaussian basis function corresponding to each group chromosome of obtaining _jAnd variances sigma _j ², and hidden layer is to the connection weights W of output layer _jSubstitution RBF neural network, obtain N RBF neural network;

2) use the training sample obtained in step 2, according to the N obtained in step 1) RBF neural network, obtain the calculating step-length Y of different real-time averaging methods _Step, adopt respectively real-time averaging method to process sample data, obtain the angular velocity data after the n group suppresses Random Drift Error

3) calculate j chromosome fitness function $f_j=\frac{1}{E[(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})*{(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})}^T]},$ In formula,

Angular velocity for turntable output;

3.5: judge whether to reach any one in the middle of two end conditions, if meet, the data that optimum chromosome is corresponding form optimum RBF neural network; If do not meet, carry out Population Evolution and operate and return step 3.3;

Described end condition comprises:

(1). the number of times of Population Evolution reaches predefined cycle index NumCycle, 50≤NumCycle≤100;

(2). fitness meets $f_j>\frac{1}{\mathrm{\α}*E[(w_j^o-w_j^{\mathrm{ZT}})*{(w_j^o-w_j^{\mathrm{ZT}})}^T]},$ In formula,

For the angular velocity of the original output of gyroscope, α is a proportional control factor, 0.01≤α<1;

Described evolution operation comprises the following steps:

1) retain the individuality of fitness front 3 in the parent population, directly copy as filial generation; And then utilize roulette method to select remaining individuality, until produce N individuality;

2) according to the crossover probability P set _cDetermine whether chromosome will intersect, 0.4≤P _c≤ 0.8, crossover operator adopts 2 bracketing methods;

If two the chromosome of intersection is respectively

With ${\mathrm{<figure-callout\; id="2"\; label="PG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">PG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}],$ In formula,

With

Represent respectively the gene on homologue;

Two of random generations are less than or equal to 90 positive integer r ₁, r ₂, r ₁<r ₂, will be more than or equal to r ₁And be less than or equal to r ₂Chromosome as exchanging object, obtain two new filial generations:

${\mathrm{<figure-callout\; id="1"\; label="CG"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">CG</figure-callout>}}_1=[A_1^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_2^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{r_1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]$

${\mathrm{<figure-callout\; id="2"\; label="CG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">CG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{r_1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]$

3) according to the variation probability P of setting _mDetermine whether chromosome will make a variation, 0.001≤P _m≤ 0.2, mutation operator adopts basic mutation operator, and the one or more locus of random choose carry out the genic value negate on chromosome;

Step 4: the gyrostatic Output speed variable quantity of Real-time Collection MEMS, be input to the RBF neural network of the optimum of step 3, obtain in real time the calculating step-length of the real-time average under different situations; Then according to calculating step-length, adopt real-time averaging method to process the original output data of MEMS gyro, just can obtain the angular velocity data after Random Drift Error suppresses.

The invention has the beneficial effects as follows: the present invention is directed to the gyrostatic Random Drift Error of MEMS, adopt real-time averaging method to suppress Random Drift Error, utilize the calculating step-length based on the real-time averaging method of RBF ANN (Artificial Neural Network) Control of genetic algorithm optimization.The present invention does not need the Random Drift Error modeling, and calculated amount is little, can conveniently realize that the real-time Random Drift Error of MEMS gyroscope suppresses.

The accompanying drawing explanation

Fig. 1 is the process flow diagram of implementation method of the present invention;

Fig. 2 is RBF neural network structure schematic diagram

Fig. 3 is the gyrostatic output of MEMS Data Comparison figure before and after averaging method processing in real time.

Embodiment

The present invention is further described below in conjunction with drawings and Examples, the present invention includes but be not limited only to following embodiment.

The present invention includes following steps:

Step 1: determine the RBF neural network structure

It is single-input single-output (be input layer, the output layer unit number is all 1) that the RBF neural network is set, and input quantity is x=[Δ w], Δ w is MEMS gyroscope angular velocity varies amount, output quantity is y=[Y _Step], Y _StepFor the calculating step-length of real-time averaging method, radial basis vector h=[h ₁h ₂H _jH _n] ^T, h wherein _jFor the gaussian basis function, that is:

$h_j=\exp (-\frac{{\left|\right|x-m_j\left|\right|}^2}{2{{\mathrm{\&sigma;}}_j}^2}),j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;n---\left(1\right)$

In formula, x is the input quantity of RBF neural network, m _jAnd σ _j ²Respectively center and the variance of j hidden layer unit gaussian basis function of RBF neural network.

The weight vector of RBF neural network is:

W＝[w ₁,w ₂,…,w _n] ^T (2)

The RBF neural network is output as:

y＝[Y _step]＝W ^Th (3)

Obtain the RBF neural network initial model of 1-n-1, wherein the center m of gaussian basis function _j, variances sigma _j ²With initial value data after in step 3, initial population being decoded of weight vector W, determine.

Step 2: obtain learning sample

At first the MEMS gyroscope is fixed on the single shaft rate table by fixture, then exists

( The maximum angular acceleration that the MEMS gyroscope can measure) in scope, by uniformly-spaced (

K is the group number of image data, general 20≤k≤50), respectively to k angular acceleration of turntable input, when to turntable, inputting each angular acceleration, gather the angular velocity of MEMS gyroscope output and the angular velocity of turntable output.After the data acquisition that the MEMS gyroscope is exported and turntable is exported completed under all different angular acceleration, can obtain k group training sample.

Step 3: adopt genetic algorithm (GA) optimization, Training RBF Neural Network

With the learning sample that step 2 obtains, the Initial R BF neural network that step 1 obtains is trained, and adopt genetic algorithm to be optimized center, variance and the hidden layer of the gaussian basis function of the RBF neural network connection weights to output layer, finally obtain optimum RBF neural network.

3.1: chromosome adopts binary coding, and each chromosome coding comprises three parts, and specific definition is:

The center m of unit 1:n gaussian basis function _j, j=1,2 ... n;

The variances sigma of unit 2:n gaussian basis function _j ², j=1,2 ... n;

A 3:n hidden layer in unit is to the connection weights W of output layer _j, j=1,2 ... n;

Center m _j, variances sigma _j ², connect weights W _jAll adopt p position (generally getting 4≤p≤8) binary number representation, so a chromosomal total length is 3*p*n.

3.2: the initialization of population generates initial N(and generally gets 30≤N≤80) individual chromosome.

3.3: the chromosome decoding is converted to respectively decimal number by binary-coded each chromosomal three parts.

3.4: calculate each chromosomal fitness, concrete steps are as follows:

The center m of the gaussian basis function that each group chromosome 1) decoding in 3.3 obtained is corresponding _jAnd variances sigma _j ², and hidden layer is to the connection weights W of output layer _jSubstitution RBF neural network, can obtain N RBF neural network.

2) use the learning sample obtained in step 2, according to 1) in N RBF neural network obtaining, can obtain the calculating step-length Y of different real-time averaging methods _Step, adopt respectively real-time averaging method to process sample data, can obtain the angular velocity data after the N group suppresses Random Drift Error

J=1,2 ..., N.

3) calculate fitness

J chromosome fitness function:

$f_J=\frac{1}{E[(w_J^{\mathrm{RT}}-w_J^{\mathrm{ZT}})*{(w_J^{\mathrm{RT}}-w_J^{\mathrm{ZT}})}^T]},J=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N---\left(4\right)$

In formula,

For the angular velocity of turntable output,

For adopting the MEMS gyroscope angular velocity data after real-time averaging method.

3.5: judge whether to reach end condition.If meet, the data that optimum chromosome is corresponding form optimum RBF neural network; If do not meet, carry out Population Evolution (select, recombinate, make a variation) operation, return to step 3.3.3.5 in, end condition (consist of two conditions, arriving following arbitrary condition is end loop):

(1). the number of times of Population Evolution reaches the general 50≤NumCy of predefined cycle index NumCycle(≤cl1e);

(2). fitness meets:

$f_J>\frac{1}{\mathrm{\&alpha;}*E[(w_J^o-w_J^{\mathrm{ZT}})*{(w_J^o-w_J^{\mathrm{ZT}})}^T]},J=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N---\left(5\right)$

In formula,

For the angular velocity of turntable output,

For the angular velocity of the original output of gyroscope, α is a proportional control factor (generally getting 0.01≤α<1), E[] function is to ask expectation.

3.5 in, the operation of developing comprises the following steps:

1) select

Adopt optimum individual reservation principle and roulette wheel selection to carry out individual choice.At first retain the individuality of fitness front 3 in the parent population, directly copy as filial generation; And then utilize roulette method to select remaining individuality, until produce N individuality;

2) intersect

According to the crossover probability P set _c(P _cBy expertise, determined general span 0.4≤P _c≤ 0.8) determine whether chromosome will intersect, crossover operator adopts 2 bracketing methods.

If two the chromosome of intersection is respectively:

${\mathrm{<figure-callout\; id="1"\; label="PG"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">PG</figure-callout>}}_1=[A_1^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_2^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]---\left(6\right)$

${\mathrm{<figure-callout\; id="2"\; label="PG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">PG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]---\left(7\right)$

In formula,

Represent respectively i, a j gene (binary representation) on homologue.

Two positive integer r of random generation ₁, r ₂(0<r ₁, r ₂≤ 90), will be in r ₁～r ₂Between (comprise r ₁, r ₂, below with r ₁<r ₂For example) chromosome as exchanging object, can obtain like this two new filial generations:

${\mathrm{<figure-callout\; id="1"\; label="CG"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">CG</figure-callout>}}_1=[A_1^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_2^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{r_1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]---\left(8\right)$

${\mathrm{<figure-callout\; id="2"\; label="CG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">CG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{r_1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]---\left(9\right)$

3) variation

According to the variation probability P of setting _m(P _mBy expertise, determined general span 0.001≤P _m≤ 0.2) determine whether chromosome will make a variation, mutation operator adopts basic mutation operator.Namely the one or more locus of random choose change (the present invention adopts binary coding, and change is exactly the genic value negate, namely 0 → 1 or 1 → 0) on chromosome.

The inhibition of step 4:MEMS Gyroscope Random Drift Error

The gyrostatic Output speed variable quantity of Real-time Collection MEMS, be input to the RBF neural network that step 3 trains, and just can obtain in real time the calculating step-length of the real-time average under different situations.Then according to calculating step-length, adopt real-time averaging method to process the original output data of MEMS gyro, just can obtain the angular velocity data after Random Drift Error suppresses.

Embodiment:

Step 1: set up RBF neural network initial model

It is single-input single-output that the RBF neural network is set, and input quantity is x=[Δ w] (Δ w is the MEMS gyroscope angular velocity varies amount of former and later two sampling instants), output quantity is y=[Y _Step] (Y _StepCalculating step-length for real-time averaging method), radial basis vector h=[h ₁h ₂H _jH ₅] ^T, h wherein _jFor gaussian basis function (establishing n=5, i.e. 5 hidden layer unit), that is:

$h_j=\exp (-\frac{{\left|\right|x-m_j\left|\right|}^2}{2{{\mathrm{\&sigma;}}_j}^2}),j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;,5---\left(1\right)$

In formula, x is the input quantity of RBF neural network, m _jAnd σ _j ²Respectively center and the variance of j hidden layer unit gaussian basis function of RBF neural network.

The weight vector of RBF neural network is:

W＝[w ₁,w ₂,…,w ₅] ^T (2)

The RBF neural network is output as:

y＝[Y _step]＝W ^Th (3)

Obtain the RBF neural network initial model of 1-5-1, wherein the center m of gaussian basis function _j, variances sigma _j ²With initial value data after in step 3.3, initial population being decoded of weight vector W, determine.

Step 2: obtain learning sample

The present embodiment adopts ADIS16375 inertia system (a built-in three-axis gyroscope and a three axis accelerometer), and the MEMS gyroscope is fixed on the single shaft rate table by fixture, gathers the angular velocity of single shaft, then exists

(

Be the maximum angular acceleration that the MEMS gyroscope can measure, the present embodiment is got ) in scope by uniformly-spaced

To 50 angular acceleration of turntable input, when to turntable, inputting each angular acceleration, gather the angular velocity of MEMS gyroscope output and the angular velocity of turntable output respectively.After the data acquisition that the MEMS gyroscope is exported and turntable is exported completed under all different angular acceleration, can obtain 50 groups of training samples.

Step 3: adopt genetic algorithm (GA) optimization, Training RBF Neural Network

3.1: chromosome adopts binary coding, and each chromosome coding comprises three parts, and specific definition is:

The center m of unit 1:5 gaussian basis function _j, j=1,2 ..., 5;

The variances sigma of unit 2:5 gaussian basis function _j ², j=1,2 ..., 5;

A 3:5 hidden layer in unit is to the connection weights W of output layer _j, j=1,2 ..., 5;

Center m _j, variances sigma _j ², connect weights W _jAll adopt 6 bits to mean, so a chromosomal total length is 90.

3.2: the initialization of population generates 30 initial chromosomes.

3.3: the chromosome decoding is converted to respectively decimal number by binary-coded each chromosomal three parts.

3.4: calculate each chromosomal fitness, concrete steps are as follows:

The center m of the gaussian basis function that each group chromosome 1) decoding in 3.3 obtained is corresponding _jAnd variances sigma _j ², and hidden layer is to the connection weights W of output layer _jSubstitution RBF neural network, can obtain 30 RBF neural networks.

2) use the learning sample obtained in step 2, according to 1) in 30 RBF neural networks obtaining, can obtain the calculating step-length Y of different real-time averaging methods _Step, adopt respectively real-time averaging method to process sample data, can obtain the angular velocity data after the n group suppresses Random Drift Error

J=1,2 ..., 30.

3) calculate fitness

J chromosome fitness function:

$f_j=\frac{1}{E[(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})*{(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})}^T]},j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,30---\left(4\right)$

In formula,

For the angular velocity of turntable output,

For adopting the MEMS gyroscope angular velocity data after real-time averaging method.

3.5: judge whether to reach end condition.If meet, the data that optimum chromosome is corresponding form optimum RBF neural network; If do not meet, carry out Population Evolution (select, recombinate, make a variation) operation, return to step 3.3.3.5 in, end condition (consist of two conditions, arriving following arbitrary condition is end loop):

(1). the number of times of Population Evolution reaches predefined cycle index NumCycle(and should embodiment selects NumCycle=50);

(2). fitness meets:

$f_j>\frac{1}{\mathrm{\&alpha;}*E[(w_j^o-w_j^{\mathrm{ZT}})*{(w_j^o-w_j^{\mathrm{ZT}})}^T]},j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n---\left(5\right)$

In formula,

For the angular velocity of turntable output,

For the angular velocity of the original output of gyroscope, α is a proportional control factor (the present embodiment is got α=0.1), E[] function is to ask expectation.

3.5 in, the operation of developing comprises the following steps:

1) select

Adopt optimum individual reservation principle and roulette wheel selection to carry out individual choice.At first retain the individuality of fitness front 3 in the parent population, directly copy as filial generation; And then utilize roulette method to select remaining individuality, until produce 30 individualities;

2) intersect

According to the crossover probability P set _c(the present embodiment is chosen P _c=0.5) determine whether chromosome will intersect, crossover operator adopts 2 bracketing methods.

If two the chromosome of intersection is respectively:

${\mathrm{<figure-callout\; id="1"\; label="PG"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">PG</figure-callout>}}_1=[A_1^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_2^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]---\left(6\right)$

${\mathrm{<figure-callout\; id="2"\; label="PG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">PG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]---\left(7\right)$

In formula,

Represent respectively i, a j gene (binary representation) on homologue.

Two positive integer r of random generation ₁, r ₂(0<r ₁, r ₂≤ 90), will be in r ₁～r ₂Between (comprise r ₁, r ₂, below with r ₁<r ₂For example) chromosome as exchanging object, can obtain like this two new filial generations:

${\mathrm{<figure-callout\; id="1"\; label="CG"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">CG</figure-callout>}}_1=[A_1^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_2^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{r_1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]---\left(8\right)$

${\mathrm{<figure-callout\; id="2"\; label="CG"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">CG</figure-callout>}}_2=[A_1^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_2^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},A_{r_1}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA0000366764420000011.png,HDA0000366764420000021.png"\; state="\{\{state\}\}">P</figure-callout>}}_1},A_{{\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">r</figure-callout>}}_2+1}^{{\mathrm{<figure-callout\; id="2"\; label="P"\; filenames="HDA0000366764420000011.png"\; state="\{\{state\}\}">P</figure-callout>}}_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]---\left(9\right)$

3) variation

According to the variation probability P of setting _m(the present embodiment is chosen P _m=0.05) determine whether chromosome will make a variation, mutation operator adopts basic mutation operator.Namely the one or more locus of random choose change (the present invention adopts binary coding, and change is exactly the genic value negate, namely 0 → 1 or 1 → 0) on chromosome.

The inhibition of step 4:MEMS Gyroscope Random Drift Error

4.1 the angular velocity of Real-time Collection MEMS gyroscope output.

4.2 using the variable quantity of MEMS gyroscope angular velocity as input, be updated to the optimum RBF neural network that step 3 training obtains, can obtain in real time the calculating step-length L of real-time averaging method.

4.3, according to the 4.2 calculating step-length L that determine, adopt real-time averaging method to process the angular velocity data of the gyrostatic output of MEMS, the angular velocity data after the Random Drift Error that can be inhibited.

As shown in Figure 3, be the Data Comparison figure after the original output data of MEMS gyroscope and employing method of the present invention suppress random drift, can obviously see and adopt method of the present invention well to suppress Random Drift Error, can get a desired effect.

The present invention, according to the gyrostatic different operating state of MEMS, utilizes the RBF neural network rationally to adjust the calculating step-length of real-time averaging method, can effectively suppress the gyrostatic Random Drift Error of MEMS.This invention calculated amount is little, can be easily on engineering.

The content be not described in detail in instructions of the present invention belongs to the known prior art of professional and technical personnel in the field.

Claims (1)

1. the disposal route of a MEMS Gyroscope Random Drift Error, is characterized in that comprising the steps:

Step 1: it is single-input single-output that the RBF neural network is set, and input quantity is x=[Δ w], Δ w is MEMS gyroscope angular velocity varies amount, output quantity is y=[Y _Step], Y _StepFor the calculating step-length of real-time averaging method, radial basis vector h=[h ₁h ₂H _jH _n] ^T, h wherein _jFor the gaussian basis function, n is the hidden layer unit number,

$h_j=\exp (-\frac{{\left|\right|x-m_j\left|\right|}^2}{2{{\mathrm{\&sigma;}}_j}^2}),j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;n$

In formula, x is the input quantity of RBF neural network, m _jAnd σ _j ²Respectively center and the variance of j hidden layer unit gaussian basis function of RBF neural network;

The weight vector W=[w of RBF neural network ₁, w ₂..., w _n] ^T,

The output y=[Y of RBF neural network _Step]=W ^TH; Obtain the RBF neural network initial model of 1-n-1, wherein the center m of gaussian basis function _j, variances sigma _j ²With initial value data after in step 3, initial population being decoded of weight vector W, determine;

Step 2: the MEMS gyroscope is fixed on the single shaft rate table, then exists

In scope, uniformly-spaced give respectively the input of single shaft rate table k angular acceleration, when inputting each angular acceleration for the single shaft rate table, gather the angular velocity of MEMS gyroscope output and the angular velocity of turntable output, obtain k group training sample; Wherein

Be the maximum angular acceleration that the MEMS gyroscope can measure, uniformly-spaced described

K is the group number of image data, 20≤k≤50;

Step 3: RBF neural network initial model training step 1 obtained with the training sample that step 2 obtains, and adopt genetic algorithm to be optimized center, variance and the hidden layer of the gaussian basis function of the RBF neural network connection weights to output layer, finally obtain optimum RBF neural network, specifically comprise the following steps:

3.1: chromosome adopts binary coding, and each chromosomal binary coding includes the center m of n gaussian basis function _j, a n gaussian basis function variances sigma _j ²With the connection weights W of n hidden layer to output layer _j, j=1,2 ... n; Center m _j, variances sigma _j ², connect weights W _jAll adopt the p bit to mean, a chromosomal total length is 3*p*n, 4≤p≤8;

3.2: the initialization of population generates N initial chromosome, 30≤N≤80;

3.3: the chromosome decoding is converted to respectively decimal number by binary-coded each chromosomal three parts;

3.4: calculate each chromosomal fitness, concrete steps are as follows:

1) will the decode center m of gaussian basis function corresponding to each group chromosome of obtaining _jAnd variances sigma _j ², and hidden layer is to the connection weights W of output layer _jSubstitution RBF neural network, obtain N RBF neural network;

2) use the training sample obtained in step 2, according to the N obtained in step 1) RBF neural network, obtain the calculating step-length Y of different real-time averaging methods _Step, adopt respectively real-time averaging method to process sample data, obtain the angular velocity data after the n group suppresses Random Drift Error

3) calculate j chromosome fitness function $f_j=\frac{1}{E[(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})*{(w_j^{\mathrm{RT}}-w_j^{\mathrm{ZT}})}^T]},$ In formula, Angular velocity for turntable output;

3.5: judge whether to reach any one in the middle of two end conditions, if meet, the data that optimum chromosome is corresponding form optimum RBF neural network; If do not meet, carry out Population Evolution and operate and return step 3.3; Described end condition comprises:

(1). the number of times of Population Evolution reaches predefined cycle index NumCycle, 50≤NumCycle≤100;

(2). fitness meets $f_j>\frac{1}{\mathrm{\&alpha;}*E[(w_j^o-w_j^{\mathrm{ZT}})*{(w_j^o-w_j^{\mathrm{ZT}})}^T]},$ In formula,

For the angular velocity of the original output of gyroscope, α is a proportional control factor, 0.01≤α<1;

Described evolution operation comprises the following steps:

1) retain the individuality of fitness front 3 in the parent population, directly copy as filial generation; And then utilize roulette method to select remaining individuality, until produce N individuality;

2) according to the crossover probability P set _cDetermine whether chromosome will intersect, 0.4≤P _c≤ 0.8, crossover operator adopts 2 bracketing methods;

If two the chromosome of intersection is respectively

With ${\mathrm{PG}}_2=[A_1^{P_2},A_2^{P_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}],$ In formula,

With

Represent respectively the gene on homologue;

Two of random generations are less than or equal to 90 positive integer r ₁, r ₂, r ₁<r ₂, will be more than or equal to r ₁And be less than or equal to r ₂Chromosome as exchanging object, obtain two new filial generations:

${\mathrm{CG}}_1=[A_1^{P_1},A_2^{P_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{P_1},A_{r_1}^{P_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{P_2},A_{r_2+1}^{P_1},\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_1}]$

${\mathrm{CG}}_2=[A_1^{P_2},A_2^{P_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_1-1}^{P_2},A_{r_1}^{P_1},\&CenterDot;\&CenterDot;\&CenterDot;,A_{r_2}^{P_1},A_{r_2+1}^{P_2},\&CenterDot;\&CenterDot;\&CenterDot;,A_{90}^{P_2}]$

3) according to the variation probability P of setting _mDetermine whether chromosome will make a variation, 0.001≤P _m≤ 0.2, mutation operator adopts basic mutation operator, and the one or more locus of random choose carry out the genic value negate on chromosome;

Step 4: the gyrostatic Output speed variable quantity of Real-time Collection MEMS, be input to the RBF neural network of the optimum of step 3, obtain in real time the calculating step-length of the real-time average under different situations; Then according to calculating step-length, adopt real-time averaging method to process the original output data of MEMS gyro, just can obtain the angular velocity data after Random Drift Error suppresses.

Images