Online data-driven fault detection for robotic systems

Abstract

Rollover is the principal cause of serious accidents for All-Terrain Vehicles (ATV), especially for light vehicles (e.g. quad bikes). In order to reduce this risk, the development of active devices, contributes a promising solution. With this aim, this paper proposes an algorithm allowing to predict the rollover risk, by means of an on-line estimation of a stability criterion. Among several rollover indicators, the Lateral Load Transfer (LLT) has been chosen because its estimation needs only low cost sensing equipment compared to the price of a light ATV. An adapted backstepping observer associated to a bicycle model is first developed, allowing the estimation of the grip conditions. In addition, the lateral slope is estimated thanks to a classical Kalman filter relying on measured acceleration and roll rate. Then, an expression of the LLT is derived from a roll model taking into account the grip conditions and the slope. Finally, the LLT value is anticipated by means of a prediction algorithm. The capabilities of this system are investigated thanks to full scale experiments with a quad bike.