Improving MEMS Gyroscope Performance using Homogenous Sensor Fusion
Authors
Totu, Luminita Cristiana ; Konge Koldbæk, Simon
Term
10. term
Publication year
2011
Submitted on
2011-05-31
Pages
161
Abstract
Projektet undersøger, om man kan forbedre et array af små, chip-baserede gyroskoper (MEMS-gyroskoper) ved at udnytte, at deres støj kan være korreleret, og ved at kombinere målingerne med Kalman-filtrering (en metode til at fusionere støjende data over tid). Arbejdet bygger på Bayard og Ploen, der i simuleringer viser, at ydeevnen kan forbedres, når målinger fra gunstigt korrelerede gyroskoper kombineres. Projektet er gennemført i samarbejde med CDL Scotland, der udvikler løsninger til inertiel navigation under vand. CDL har designet et specialbygget sensorboard med otte gyroskoper i mellemklassen samt nødvendig interfacehardware. Med Allan-varians (en standardmetode til at karakterisere sensorstøj over forskellige tidsskalaer) og klassiske signalanalysemetoder er der udviklet en enkel stokastisk model for den tilfældige bias (vinkeldrift) i gyroskopets udgangssignal og implementeret i MATLAB. Modellen er valideret ved komparativ analyse af Root Allan Variance. Expectation-Maximization (EM, en statistisk algoritme til at estimere skjulte parametre) er implementeret og testet i MATLAB for at identificere støjkorrelationer. Flere metoder til at forbedre et gyroskoparray er undersøgt: Kalman-filterbaserede estimationsstrategier er sammenlignet med simpel gennemsnitsfiltrering. Simuleringerne viser, at forbedringer—især i estimering af vinkeldrift—er mulige, hvis der findes gunstige korrelationer mellem gyroskopernes støjprocesser. Kalman-filteret anvendes her atypisk på en tilstandsmodel, der ikke er observerbar eller detekterbar (dvs. ikke alle tilstande kan udledes robust fra målingerne), og brugen analyseres. EM-algoritmen kunne under simuleringerne ikke identificere alle relevante støjkorrelationer tilstrækkeligt præcist. Hovedproblemet er forholdet mellem målenøjagtighed og systemstøj: systemstøjen er to størrelsesordener mindre end målestøjen. Derfor er identifikation af støjkorrelationer fortsat uafklaret, og forbedringspotentialet for gyroskopboardet kan ikke vurderes tilfredsstillende på nuværende tidspunkt.
This project explores whether the performance of an array of small, chip-based gyroscopes (MEMS gyroscopes) can be improved by exploiting correlations in their noise and by combining measurements with Kalman filtering (a method for fusing noisy data over time). The work builds on Bayard and Ploen, who showed in simulations that performance can improve when measurements from favorably correlated gyroscopes are combined. The project was proposed and carried out with CDL Scotland, a developer of subsea inertial navigation sensors and solutions. CDL designed a custom sensor board with eight medium-grade gyroscopes and interface hardware. Using Allan variance (a standard way to characterize sensor noise across time scales) and classical signal analysis, we developed a simple stochastic model of the random bias (angle drift) in the gyroscope output and implemented it in MATLAB. The model was validated through comparative analysis of the Root Allan Variance. Expectation-Maximization (EM, a statistical algorithm for estimating hidden parameters) was implemented and tested in MATLAB to identify noise correlations. Several array-processing methods were investigated: Kalman filter–based estimation strategies were benchmarked against a simple averaging filter. Simulations indicate performance gains—especially in angle drift estimation—are possible if favorable correlations exist between the noise processes of the gyroscopes. The Kalman filter is applied here in an unusual way to a state-space model that is not observable or detectable (meaning not all states can be reliably inferred from measurements), and this usage is analyzed. The implemented EM algorithm could not identify all relevant noise correlations with sufficient accuracy in simulations. The main obstacle is the ratio of measurement noise to system noise: the system noise is two orders of magnitude smaller than the measurement noise. As a result, identifying the noise correlations remains an open problem, and the improvement potential of the gyroscope board cannot yet be assessed satisfactorily.
[This abstract was generated with the help of AI]
Documents