PB model-based FEL and its application to Driving Pattern Learning

Abstract

A significant part of the vehicle control software is based on Look-Up-Tables (LUTs) that define the mappings from different combinations of multiple independent (input) variables to the values of one dependent (output). LUTs are used as feedforward controllers or as gain-scheduling parameters for feedback controllers. The LUT feedforward controllers can be viewed as inverse vehicle models capturing the strong nonlinearity and multimodal behaviors that can be (in many cases) formalized only by experimentally measured data under different operating conditions. The paper proposes a Feedback Error Learning (FEL) based method for adaptation of the LUT feedforward controllers in order to match the desired and actual vehicle performance. The FEL is an on-line learning strategy acquiring an inverse model of a plant through feedback control actions. In this paper we consider the driver demand LUTs as a feedforward controller defining the relationship between the accelerator pedal position, the engine speed, and corresponding brake torque and the driver as a feedback controller. The FEL scheme have been implemented through Piecewise Bilinear (PB) models which can be expressed as LUTs and are very convenient with regard to nonlinear modeling, control objective and on-line learning capability.