The scope of the project is to investigate the possibilities of using noise correlations and
Kalman filtering to improve the performance of a sensor array containing multiple MEMS
gyroscope. The project is based on the work of Bayard and Ploen whom have showed,
trough simulation, that the performance of MEMS gyroscopes can be improved by combining
measurements from favorably correlated gyroscopes. In addition, the project also investigates
the possibility of identifying noise correlations by using Expectation-Maximization. The
project has been proposed and carried out in collaboration with CDL Scotland, which is
developer and provider of subsea inertial navigation sensors and solutions. A custom sensor
board containing eight medium grade gyroscopes and additional interface hardware has been
designed for the project by CDL.
Based on Allan Variance analysis and classical signal analysis methods, a simple stochastic
model for the random bias component in the gyroscope output signal has been developed and
implemented in MATLAB. The model has been validated through comparative analysis of
the Root Allan Variance. The Expectation-Maximization algorithm has been implemented
in and tested in MATLAB. Several methods of improving the performance of a sensor array
containing multiple MEMS gyroscope has been investigated.
Using Kalman filter based estimation strategies and benchmarking against a simple signal
averaging filter, the group has shown through simulations that performance improvement,
especially in the angle drift estimation, is possible assuming that favorable correlations exist
between the noise processes of the gyroscope signals. The Kalman filter algorithm is atypically
applied to a state space model that is not observable nor detectable, and an interesting
analysis of this usage is discussed. The implemented Expectation-Maximization algorithm
was not able to identify all the relevant noise correlations with sufficient accuracy during
simulations. The main problem is related to the ratio between the measurement noise and the
system noise, as the system noise is two orders of magnitude smaller than the measurement
noise. As such, the problem of identifying the noise correlations remains open, and the
improvement potential of the gyroscope board can not be assessed in a satisfactory manner
at the present time.