Rollover Control in Heavy Vehicles via Recurrent High Order Neural Networks

Abstract

Heavy vehicles, such as tractor-semitrailers, play an important role in transportation systems. They present more complex dynamical behavior than passenger cars, due to their high centers of gravity, which can vary depending on the load conditions, and are highly susceptible to rollover during cornering. Heavy vehicle rollover on highways occurs as a result of cornering with excessively high speed, cornering on a small radius curve or sudden lane change. However, if rollover threat is predicted using an appropriate algorithm, then the accident can be prevented by the driver's corrective maneuvers. For situations where rollover warning is ineffective, active rollover control is necessary. Most of the rollover warning algorithms use instantaneous rollover-threat index to identify the rollover threat. Since a rollover warning may be issued at 75 % of the rollover threshold acceleration, the time from warning to rollover is too short for the driver to respond effectively. However, if the rollover threat is predicted using the expected maneuvers, a warning can be issued sufficiently in advance of the event. This fact implies that warning systems based on predicted rollover threat can be more effective. Many control strategies have been designed to prevent rollover, most of them based on active speed control and active roll control. However, active roll control is ineffective for sharp turns, since it does not reduce the lateral acceleration, and requires hydraulic actuators which increase the cost considerably. On the other hand, the use of differential braking prevent jack-knifing and loss of direction generated by sudden braking during cornering. Different loading configurations produce different reaction forces on each wheel. This motivates the use of nonlinear robust controllers which have to be able to deal with parametric uncertainties, but most controllers are based on reduced models, in order to lessen the computational requirements. Many mathematical models for tractor semitrailers have been developed in order to derive active control algorithms. The Automotive Research Center of the University of Michigan developed the 33 degrees-of-freedom ArcSim model (UMTRI, 1997) to study the acceleration/braking and handling responses of an US Army 6-axle tractor-semitrailer. In (Hyun & Langari, 2003), the vehicle model for single-unit heavy