Backlash Estimation for Heavy Truck Steering Systems

Abstract

Mechanical backlash in the steering gear of heavy-duty trucks introduces nonlinearities and dead zones that degrade steering precision and complicate the control algorithms required for Advanced Driver Assistance Systems (ADAS) and autonomous operation. As components wear over their operational lifespan, backlash increases, motivating the need for online estimation to enable adaptive control and predictive maintenance. While dedicated load-side angle sensors (located on the side of the steering gear subject to the road load) could directly quantify this wear, their cost and reliability in harsh environments make a software-based solution preferable. This thesis proposes a Switched Kalman Filter (SKF) for estimating both the size and position of the steering gear backlash using only sensors already existing on the truck. The method adapts the framework of Lagerberg and Egardt to the topology of a heavy-duty truck steering system, where no physical sensor exists on the load side of the steering gear. A key contribution is the reformulation of the backlash offset states into a center position and a half-size state, which decouples slowly varying measurement biases from the physically meaningful backlash size estimate and resolves a systematic directional error present in the original formulation. The absent load-side measurement is addressed by estimating the pitman arm angle from yaw rate using a steady-state bicycle model with lead compensation. The observability and practical stability of the switched system are analyzed theoretically. The estimator is validated against highway driving data from a heavy-duty truck with known backlash levels. Using a physical pitman arm sensor, the worst-case estimation error is 11% with an absolute error of 0.50◦. Using the yaw-rate-based angle estimate on the same hardware configuration, the worst-case error increases to 35% with an absolute error of 0.70◦. Parameter sensitivity analysis shows the steering model to be robust to large parameter variations, while the bicycle model represents the primary source of estimation uncertainty.