CN106597852A - MEMS gyroscope temperature compensation method based on RBF neural network - Google Patents

Abstract

The invention relates to an MEMS (Micro-Electromechanical System) gyroscope temperature compensation method based on an RBF (Radius Basis Function) neural network. Sample data is processed using a k-means clustering algorithm and a recursive least square method. Modeling is carried out based on an RBF neural network algorithm principle which is more accurate for nonlinear mapping to make compensation for the temperature. Thus, the problem on how to make compensation for the temperature of an MEMS gyroscope under strong nonlinear characteristic is solved.

Description

A kind of MEMS gyroscope temperature compensation based on RBF neural

Technical field

The present invention relates to MEMS gyroscope technical field, specially the MEMS gyroscope temperature based on RBF neural are mended Compensation method, it is adaptable to the temperature-compensating of MEMS gyroscope.

Background technology

Temperature is the key factor for affecting MEMS inertia system attitude output accuracies, and composition can be made under different temperatures environment The material of MEMS inertia devices produces different deformation, and then the additional moment for causing causes output bias, while other circuit elements Part characteristic also can produce impact with the change of ambient temperature on MEMS inertia devices.Even if from environment temperature in MEMS gyro structure There is no big change in degree, gyro itself work long period, its internal temperature also can be varied widely, Jing experimental studies, The main device of its internal influence temperature be power module, design circuit when MEMS gyroscope away from power supply near, but Because MEMS inertia measurement product structures are little, MEMS gyro is still affected larger by power supply heating.

Gyroscope gauge outfit barycenter shifts in temperature change, so as to cause specific force and angular acceleration disturbance torque, makes top The output of spiral shell instrument is drifted about, therefore under dynamic environment, temperature is received in particularly low input speed area, the output of MEMS inertia devices Change affects maximum, and under various circumstances zero bias shows very strong nonlinear characteristic.

The present invention is directed to sensitivity and its nonlinear characteristic of the MEMS gyroscope to temperature, adopts to nonlinear mapping more Accurately neural network algorithm principle carries out temperature-compensating to it, and the temperature solved under MEMS gyroscope strong nonlinearity characteristic is mended Repay problem.

The content of the invention

The technical problem to be solved

In order to avoid the deficiencies in the prior art part, the present invention proposes a kind of MEMS gyroscope based on RBF neural Temperature compensation.

Technical scheme

A kind of MEMS gyroscope temperature compensation based on RBF neural, it is characterised in that step is as follows：

Step 1：Dynamic and static state performance testing experiment is carried out to MEMS gyroscope by single axle table, is obtained under each temperature spot not Compensation gyro output valve and turntable rotating speed true value test data；Experimental data is divided into into two parts, a part is used for training pattern Parameter, a part is for the checking of last model and network test；

Step 2：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural, turntable Output layer of the rotating speed true value as RBF neural；Input quantity is clustered using K mean cluster algorithm so that cost letter Number J (C) is minimized, and obtains cluster centre

Wherein：J (C) is the cost function of encoder C；To belong to the estimation mean vector of cluster j, i.e. cluster centre, x_i For x₁,x₂Matrix；

Step 3：Calculate each cluster centreThe distance between, select ultimate range d_max：

Step 4：By ultimate range d between each cluster centre_maxAnd each cluster centre number K obtains Gaussian function width；

Step 5：Output layer is trained by recurrent least square method, obtains weights ω_i：

R(n)ω_i=r (n), n=1,2 ...

Wherein：R (n) is K × K correlation functions of hidden unit output；R (n) is the Expected Response and hidden of RBF networks output Hide K × 1 cross correlation vector between unit；

Step 6：According to the weights ω for obtaining_i, width cs, center in Gaussian functionWith radial direction baseBuild RBF neural Network model：

Wherein, x is input quantity；F (x) is output；σ be withCentered on Gaussian function Wide observation；

Step 7：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural network model, Output obtains carrying out MEMS gyroscope output valve f (x) after temperature-compensating.

Beneficial effect

A kind of MEMS gyroscope temperature compensation based on RBF neural proposed by the present invention, has the beneficial effect that：

1) for practical situation and MEMS gyroscope structure in experimental data, select using simple efficient K mean cluster method And recurrent least square method, effectively increase the efficiency of RBF mixture of networks learning trainings, significant increase work efficiency；

2) RBF neural is good to the degree of fitting of non-linear temperature model, effectively improves the output essence of MEMS gyroscope Degree.

Description of the drawings

Fig. 1 RBFs (RBF) neural network structure, input in the present invention is x₁,x₂That is temperature and uncompensated MEMS Gyro output valve, is output as the measuring turntable measured values (true value) of f (x) i.e. MEMS.

Specific embodiment

It is of the invention mainly right using RBF (radial basis function, RBF) neutral net (see Fig. 1) MEMS gyro carries out model of temperature compensation foundation (see formula 1), so as to realize x ∈ R^m→f(x)∈RⁿI.e. gyro output parameter and Nonlinearity in parameters mapping after temperature parameter and temperature-compensating.

Wherein：ω_i(1≤i≤m) is weighted value；

x∈R^mFor input quantity (gyro output parameter, and temperature parameter)；

f(x)∈RⁿFor output (parameter, i.e. test data intermediate station rotating speed after gyro output compensation)；

c_iCentered on

For radial direction base, this programme adopts Gaussian functionσ is with c_iCentered on Gaussian function Several wide observation.

It is thus determined that data center c in function_i, width cs and weights ω_iIt is achieved with model.And RBF neural parameter It is by obtained from the training study to sample data.

RBF neural is the feedforward network of a three-decker：Input layer, non-linear hidden layer and linear convergent rate layer. The training of RBF neural in this programme was carried out by following two stages：

1) total data sample is constructed using K mean cluster, trains non-linear hidden layer, obtain wide in this stage Degree σ and radial direction base center c_i；

2) linear convergent rate layer weights ω is carried out using recurrent least square method_iCalculating.

Input quantity is clustered using K mean cluster method, the sample set is carried out using recurrent least square method excellent Change and select, obtain radial direction base.So that it is determined that the desired temperature-compensating mould for obtaining of radial basis function neural network, i.e. this programme Type.Detailed process is as follows：

Step 1：Dynamic and static state performance testing experiment is carried out to MEMS gyroscope by single axle table：

Specific experiment content：

1) Temperature of Warm Case is made to be respectively maintained at -18 DEG C, -10 DEG C, -5 DEG C, 0 DEG C, 10 DEG C, 25 DEG C, 35 DEG C, 50 DEG C, 68 DEG C of temperature Under degree, turntable rotating speed is 0 °/s；Read every a timing at a temperature of each and do not compensate gyro output original value, obtain each temperature Spend next group of data.

2) Temperature of Warm Case is made to be respectively maintained at -18 DEG C, -10 DEG C, -5 DEG C, 0 DEG C, 10 DEG C, 25 DEG C, 35 DEG C, 50 DEG C, 68 DEG C of temperature Under degree, turntable is made respectively with ± 0.1 °/s, ± 0.5 °/s, ± 1 °/s, ± 10 °/s, ± 50 °/s, the rotation of the angular speed of ± 150 °/s Turn；Read every a timing at a temperature of each and do not compensate gyro output original value, obtain next group of data of each temperature.

It is derived from not compensating gyro output original value and turntable rotating speed (true value) under the i.e. each temperature spot of test sample data One group of experimental data under each temperature value is divided into into two parts, a part is used for training pattern parameter, and a part is for most Afterwards model is verified and network test；

Step 2：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural, turntable Output layer of the rotating speed true value as RBF neural；Input quantity is clustered using K mean cluster algorithm so that cost letter Number J (C) is minimized, and obtains cluster centre

Wherein：J (C) is the cost function of encoder C；To belong to the estimation mean vector of cluster j, i.e. cluster centre, x_i For x₁,x₂Matrix；

Step 3：Using seeking each cluster centre of the norm calculation between matrixThe distance between, select ultimate range d_max：

Step 4：By ultimate range d between each cluster centre_maxAnd each cluster centre number K obtains Gaussian function width；

Step 5：Output layer is trained by recurrent least square method, obtains weights ω_i：

R(n)ω_i=r (n), n=1,2 ...

Wherein：R (n) is K × K correlation functions of hidden unit output；R (n) is the Expected Response and hidden of RBF networks output Hide K × 1 cross correlation vector between unit；

Step 6：According to the weights ω for obtaining_i, width cs, center in Gaussian functionWith radial direction baseBuild RBF neural Network model：

Wherein, x is input quantity；F (x) is output；σ be withCentered on Gaussian function Wide observation；

Step 7：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural network model, Output obtains carrying out MEMS gyroscope output valve f (x) after temperature-compensating.

Claims (1)

1. a kind of MEMS gyroscope temperature compensation based on RBF neural, it is characterised in that step is as follows：

Step 1：Dynamic and static state performance testing experiment is carried out to MEMS gyroscope by single axle table, is obtained and is not compensated under each temperature spot Gyro output valve and turntable rotating speed true value test data；Experimental data is divided into into two parts, a part is used for training pattern parameter, A part is for the checking of last model and network test；

Step 2：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural, turntable rotating speed Output layer of the true value as RBF neural；Input quantity is clustered using K mean cluster algorithm so that cost function J (C) minimize, obtain cluster centre

$J\left(C\right)=\underset{j=1}{\overset{K}{\Σ}}\underset{C\left(i\right)=j}{\Σ}\left|\right|x_i-{\hat{c}}_j|{|}^2$

Wherein：J (C) is the cost function of encoder C；To belong to the estimation mean vector of cluster j, i.e. cluster centre, x_iFor x₁,x₂Matrix；

Step 3：Calculate each cluster centreThe distance between, select ultimate range d_max：

Step 4：By ultimate range d between each cluster centre_maxAnd each cluster centre number K obtains Gaussian function width；

$\mathrm{\σ}=\frac{d_{\max }}{\sqrt{2K}}$

Step 5：Output layer is trained by recurrent least square method, obtains weights ω_i：

R(n)ω_i=r (n), n=1,2 ...

Wherein：R (n) is K × K correlation functions of hidden unit output；R (n) is the Expected Response of RBF networks output and hides single K × 1 cross correlation vector between unit；

Step 6：According to the weights ω for obtaining_i, width cs, center in Gaussian functionWith radial direction baseBuild RBF neural Model：

Wherein, x is input quantity；F (x) is output；σ be withCentered on Gaussian function width Observation；

Step 7：By temperature and uncompensated MEMS gyro output valve x₁,x₂As the input layer of RBF neural network model, output Obtain carrying out MEMS gyroscope output valve f (x) after temperature-compensating.