Research on Signal Processing of MEMS Gyro Array

A new random drift model and the measured angular rate model of MEMS gyro are presented. Based on such models, signal processing techniques are used to decrease gyro drift. Kalman filtering equations have been built for static measurement and dynamic measurement of the gyro array, which combines N individual gyros into a single rate estimate. By selecting the favorable cross correlation coefficient between individual gyros in the noise correlation matrix, the gyro array performance can be significantly improved over that of any individual component device. A new gyro array dynamic measurement procession is also presented. Data fusion of the difference between individual gyro dynamic measurements can identify every gyro real-time drift out and get its noisy test. Based on the laws of the gyro curve motion, the tested dynamic signal is filtered to improve the gyro accuracy. All these processings have been implemented by digital signal processor. Simulation results show that the static drift can decrease from 22.1/h to 0.184/h and the dynamic drift can decrease from 22.1/h to 8.98/h.


Introduction
Recently, the development of microelectromechanical system (MEMS) technologies [1][2][3] enables us to have a MEMS gyro small enough, which has a great potential to be applied to many applications such as virtual reality, car navigation, inertial navigation for small air vehicles, and space avionics.Moreover, for conventional gyros, such as mechanical or optical gyros, they have other good properties including (1) compact size, (2) small weight, (3) low power consumption, (4) low cost micromachining process, and (5) ease of mass production [4,5].These factors offer a wide range of applications for MEMS gyros, ranging from stability and navigation control in spacecraft to rollover detection for automotive applications, consumer electronics, robotics, and a variety of military applications.Compared to mechanical or optical gyros, MEMS gyro present performance is not high enough to substitute them.So it is necessary to improve the performance of MEMS gyros to extend their utilities.
Gyro drift is the main factor that affects its performance.Therefore, how to eliminate the gyro drift effectively is the essential problem to guarantee the gyro accuracy.In general, the gyro drift is classified into systematic and random drift.By use of proper mathematical model and drift compensation calculation, the effect of systematic drift on accuracy can be eliminated.However, random drift is weak.And it is a slow time-varying signal.It is often affected by some indeterminate factors such as exterior environment noise.Simple methods to compensate the random drift are not effective.The random drift error includes the drift instability, rate random walk (RRW), and angular random walk (ARW).Theoretically, the drift instability and the measured true rate can be modeled as random walk process, driven by the white noise.
In the area of MEMS gyro, significant accuracy improvements in the designing and manufacturing process will require increased costs and are time-consuming.Thus, signal processing techniques applied to multiple sensor configurations are being examined and are expected to improve the accuracy of existing MEMS gyro.Appropriate processing of microsensor outputs appears to be the quickest and the most feasible method of improving sensor accuracy.Up to present, there appear many approaches proposed for signal processing techniques of the gyro drift data.Sensor modeling and the active control of the sensor dynamics for model identification and angular rate sensing were focused on to compensate the gyro drift [2,[6][7][8][9][10].ARMA time series autoregression model of the gyro original drift or its first-order difference was proposed to decrease the noise and predict the gyro measurement [11][12][13][14][15][16][17].Neural network structure with the name of RBF [18,19] and GRBF [20,21] was applied to reduce drift influence on the gyro accuracy.The waveletbased denoising was employed to reconstruct outputs of the MEMS gyro and to increase the gyro accuracy [22][23][24].All these processing techniques were based on single MEMS gyro output and often aimed at its static measurement, so the performance improvement of the gyro dynamic measurement was limited.
In this paper, a signal processing technique for a MEMS gyro array, combining  individual gyros into a single rate estimation [4,5], is brought up to improve gyro accuracy.A new random drift model of gyro is presented firstly.The true angular rate is also modeled as a random walk driven by white noise.Based on these models, Kalman filtering equations are built to process gyro array static measurement and dynamic measurement.By setting favorable cross coefficients between individual gyros of the noise correlation matrix, the optimal procession result will be gotten from the gyro array static measurement.To the gyro array dynamics measurement, difference between individual gyros outputs can be served as a new variable and Kalman filtering equations will be built in order to get every gyro real-time drift and its noisy measurement indirectly.Based on the laws on the dynamic motion, the further Kalman filtering procession of the gyro noisy measurement can decrease noise of the detrained dynamic measurement and improve its signal-to-noise rate (SNR) evidently.
The remainder of this paper is organized as follows.In Section 2, we present a new gyro random drift model of the single gyro and then get the Kalman filtering equations for the gyro array static measurement.New Kalman filtering equations for gyro array dynamic measurement have also been built to identify the individual gyro real-time drift and the measured true rate in this section.Subsequently, in order to decrease noise of the identified true rate, the filtering equation is also given out.All simulation examples together with performance comparison of different Kalman filters are contained in Section 3. Section 4 presents discussion, conclusions, and future research issues.

Signal Processing Technique of Static Drift.
The gyro noisy measurement  of the true angular rate  can be described as where  is equal to slowly variant random quantity, which is denoted as the gyro drift, and  is white noise process denoted as ARW noise.Mathematically, the gyro drift can be modeled as a random walk, driven by term   , denoted as RRW noise [4,5].Consider where   is delta-correlated process with correlation function   ; that is, The true rate  can be modeled as a random walk driven by a noise with intensity   ; that is, [4,5,25,26] One of the main virtues of the silicon micromachined gyro is its easy mass production.Gyro array measurement can be used to increase the gyro accuracy.With the above statistical description, the gyro array measure equation can be written in the vector form as where   is measure result of every gyro and   is its drift and V  is its measuring noise and is a continuous-time  correlative process.The measuring noise vector correlation matrix  is as follows: where When correlation matrix  ∈  × is a non-diagonal matrix, ARW noise among every gyro is correlated with each other.Correlation scale function   from every   is the correlation, that is, where the matrix   ∈  × is nondiagonal matrix; the RRW noises among every gyro are correlated with each other.Based on measured model, the following discrete Kalman filtering equations can be gotten by (8).Consider where  = sampling internal of the measured signal;   = system white noise series;   = measurement noise series.Consider They have the following property: where   is covariance of the ARW noise and   () denotes noise intensity of a random walk, which is modeled as the true rate.

Real-Time Identification of the Random Drift from the Dynamic Measurement.
When the gyro is used to measure dynamics signal, the true rate  cannot be gotten by the output of the Kalman filtering of (8).
As seen from ( 1),  is the measurement of the true rate and is similar to every gyro measurement.Difference between the two individual gyros measurements will become as Difference between the two gyros measurements is equal to the difference between their drifts.When   () and   +   are, respectively, adopted as state variables of Kalman filter, Kalman filter state equation will become the same.Thus, real-time gyro drift can be identified and the true angular rate  is the  value between the individual gyro original measurement   and its detrained drift   +   .For the  = 3 gyro array dynamic measurement, Kalman filtering equations will be as state equation as follows: Measurement equation is as follows: All parameters of the filter are described as In this Kalman filtering equation, input variables are not the individual gyros direct measurement but are their difference.The output variables are not  but   +  .The noisy  is as follows: Further Kalman filtering can reduce the noise of the noisy tested out .The gyro rotation is curve motion.When the angle   , angular rate   , and rate acceleration   are the state variables of the Kalman filter, the state equations are as follows: = first-order derivative of   and is often regarded as the white noise.When  = [  ,   ,   ], the tested out   can be denoised by following Kalman filtering equations.State equation is as follows: Measurement equations are as follows: Thus,   can be Kalman-filtered to decrease the noise and increase gyro measurement SNR.

Simulation
A simulation study, using the preliminary Kalman filtering, is conducted to testify the analytical results.array, covariance matrices   and   will be computed by the following equations: where  = ( 1 +  2 +  3 )/3 is mean of the gyro array measurement.Var() is the variance of the signal .Data of the simulation are from our self-exploited micromachined gyro measurement.Every individual gyro drift of the gyro array is 22.1 ∘ /h and the scale factor is 4.3759 mv/ ∘ /sec.
The initial filtering value  0 is set as [0 0 0 0]  .The same correlation coefficient  is set with −1 ⩽  ⩽ 1 and its 0.05 interval during the filtering.Relations between the times of the zero-drift stability increase and the  variation is shown in Figure 1.As seen from Figure 1, variation of  affects the variation of gyro drift greatly.When  = 0.5, the filtering result is the most optimal.The drift error decreases from 22.1 ∘ /h to 0.184 ∘ /h.
When  = 0.5, every gyro drift and gyro array overall drift are shown in Figure 2.
Thus, when gyros are correlated and correlation coefficient  is selected favorably, gyro array drift will be reduced greatly.

Simulation of the Dynamic Measurement.
Gyro performance of the dynamic measurement simulation is the same as that of the static simulation.The detrained drift of one gyro from the gyro array and its original drift are compared in Figure 3.
The sinusoid, whose amplitude is 0.5 V and its frequency is 120 Hz, is added to the  value between original drift and detached gyro drift as the true noisy gyro dynamic signal .The filtered result is compared with the unprocessed dynamic signal shown in Figure 4. Finally, the drift reduces from 22.1 ∘ /h to 8.89 ∘ /h.It shows the effectiveness of such Kalman filtering of the dynamic measurement.The simulation also shows that the accuracy improvement of the gyro dynamic measurement nearly has no relation to the correlation coefficient .
The real-time of the Kalman filtering has been verified on the digital signal processor TMS320C67X.

Discussion and Conclusion
A random drift model and the measured true rate model of single gyro are presented.Gyro array is used to increase gyro performance.Different Kalman filtering equations are built for the gyro static measurement, dynamic measurement, and the tested out noisy rate would be further processed.Setting the favorable correlation coefficient , the gyro static drift can decrease from 22.1 ∘ /h to 0.184 ∘ /h.As sinusoid is used as the dynamics signal, the gyro dynamics drift decreases from 22.1 ∘ /h to 8.89 ∘ /h and the performance improvement nearly has no relation to the correlation coefficient .The db4    wavelet is used as wavelet basis; the 6-level wavelet transformation (WT) denoising of gyro array measurement can make its static drift decrease from 22.1 ∘ /h to 5.1 ∘ /h and can make the dynamic drift decrease from 22.1 ∘ /h to 15.1 ∘ /h.Processing time of every signal is 0.13 ms.So the new approach in this paper is effective and will have its further application.
In this paper, single gyros are made up of individual gyro and all data are also from every individual gyro measurement.The actual degree of performance improvement remains to be determined by experimental validation after the actual gyro array is manufactured.True noise model is not white noise but will be further ascertained.More emphases will be put on the establishment of accurate gyro array dynamic measurement model.

3. 1 .
Simulation of the Static Drift.Kalman filtering can give out the globally optimal minimum-variance estimate.Correlation among the individual gyros measurements leads to correlation among the RRW noises and among the ARW noises.Noise variance matrix of   and   can be gotten by the Allan variance computation of gyro drift data [27].By multiplying with , unit of   can be changed from deg 2 /sec 3 to deg 2 /sec 2 .Variance matrix   of the white noise, driven by the true rate , is decided by the rate stability of the tested platform.In this paper, correction coefficients among all the gyros are set as the same constant .For  = 3 gyro

Figure 1 :
Figure 1: Relation between cross correlation  and the times of the gyro drift reduction.
Component gyros drift for  = 3 correlated gyros simulation Signal magnitude (V) Kalman filtering result for  = 3 correlated gyros simulation

Figure 4 :
Figure 4: Simulation of the further Kalman filtering of the noisy tested out dynamic rate.