Model-based fault diagnosis of the sensors for lateral vehicle dynamics

Modern driver assistance systems simultaneously increase the safety and the comfort of driving. These systems are designed to fulfil routine tasks and thus aid the driver. Critical situations are detected early and are prevented by direct intervention, i.e. active braking and in the future also active steering. With the complexity of the assistance systems the information demand increases as well. It must be ensured, that the information used, in particular the signals measured by the sensors, are proper. As redundant sensors are generally not acceptable for cost reasons, the sensors must be supervised reliably. In this work the application of model-based sensor-fault diagnosis is examined at the example of the sensors for lateral vehicle dynamics. The aim is to achieve a fast, model-based detection of sensor-faults while avoiding false alarms in any driving situation. This includes driving situations, where the models used are imprecise, and under the influence of an unknown input, the road bank angle. The basis for model-based fault diagnosis are suitable models of the vehicle with adequate accuracy. The process of modelling is thus revisited carefully leading to the two-track model with non-linear modelling of the tyre forces according to the HSRI-model. From this model the one-track model with consideration of vehicle rolling motion is deducted. The essential part of this work are two methods, that are developed. The first method is the normbased estimation of the effect of modelling errors on the state-variables, which are estimated with a model or an observer. The modelling errors of the one-track model in comparison to the two-track model with nonlinear modelling of tyre forces are described as a funcion of measured sensor signals. The method provides an estimate of the faults of the system variables, which can be utilized as well for the definition of a confidence interval as for the generation of adaptive thresholds for fault diagnosis. Subsequently a novel concept for fault diagnosis is presented, which is based on the evaluation of two residuals fro each sensor to be supervised. A detection residual is sensitive to the fault to be detected while a checking residual does not react on this fault but only on any other fault as well as on model uncertainties and unknown input variables. The checking residual is used to define a threshold for the detection residual. This threshold is based on the definition of fault areas in the vector-space spanned by the residuals. To avoid false alarms, the threshold is automatically increased in driving situations with restricted model accuracy. For the realisation of the concept the residuals are generated using Kalman-Filters. The validation of the method is accomplished using measurements from real driving maneuvers