Fault Detection for Mobile Robots based on Integrated Sensor Systems

Most fault detection algorithms are based on residuals, i.e. the difference between a measured signal and the corresponding model based prediction. However, in many more advanced sensors the raw measurements are internally processed before refined information is provided to the user. The contribution of this thesis is a study of the fault detection problem when only the state estimate from an observer/Kalmanfilter is available and not the measured residual/innovation. The idea is to look at an extended state space model where the true states and the observer states are combined. This extended model is then used to generate residuals viewing the observer outputs as measurements. Results for fault observability of such extended models are given. The approach is rather straightforward in case the internal structure of the observer is exactly known. For the Kalman filter this corresponds to knowing the observer gain. If this is not the case certain model approximations can be done to generate a simplified model to be used for standard fault detection. Our motivating application has been mobile robots where the so-called pose, the position and orientation of the robot, is an important quantity that has to be estimated. The pose can be measured indirectly from several different sensor systems such as odometry, computer vision, sonar and laser. The output from these so-called pose providers are often state estimates together with, in best cases, an error covariance matrix estimate from which it might be difficult or even impossible to access the raw sensor data since the sensor and state estimator/observer are often integrated and encapsulated. In this thesis we discuss some of the main pose estimators in mobile robots and validate a fault detector filter through experiments using the our proposed framework.