Target State Optimization: Drivability Improvement for Vehicles with Dual Clutch Transmissions

Abstract

Vehicles with dual clutch transmissions (DCT) are well known for their comfortable drivability since gear shifts can be performed jerklessly. The ability of blending the torque during gear shifts from one clutch to the other, making the type of automated transmission a perfect alternative to torque converters, which also comes with a higher efficiency. Nevertheless, DCT also have some drawbacks. The actuation of two clutches requires an immense control effort, which is handled in the implementation of a wide range of software functions on the transmission control unit (TCU). These usually contain control parameters, which makes the behavior adaptable to different vehicle and engine platforms. The adaption of these parameters is called calibration, which is usually an iterative time-consuming process. The calibration of the embedded software solutions in control units is a widely known problem in the automotive industry. The calibration of any vehicle subsystem (e.g., engine, transmission, suspension, driver assistance systems for autonomous driving, etc.) requires costly test trips in different ambient conditions. To reduce the calibration effort and the accompanying use of professionals, several approaches to automize the calibration process are proposed. Due to the fact that a solution is desired which can optimize different calibration problems, a generic metaheuristic approach is aimed. Regardless, the scope of the current research is the optimization of the launch behavior for vehicles equipped with DCT since, particularly at low speeds, the transmission behavior must meet the intention of the driver (drivers tend to be more perceptive at low speeds). To clarify the characteristics of the launch, several test subject studies are performed. The influence factors, such as engine sound, maximal acceleration, acceleration build-up (mean jerk), and the reaction time, are taken into account. Their influence on the evaluation of launch with relation to the criteria of sportiness, comfort, and jerkiness, are examined based on the evaluation of the test subject studies. According to the results of the study, reference values for the optimization of the launch behavior are derived. The research contains a study of existing approaches for optimizing driving behavior with metaheuristics (e.g., genetic algorithms, reinforcement learning, etc.). Since the existing approaches have different drawbacks (in scope of the optimization problem) a new approach is proposed, which outperforms existing ones. The approach itself is a hybrid solution of reinforcement learning (RL) and supervised learning (SL) and is applied in a software in the loop environment, and in a test vehicle.