This dissertation focuses on predicting the system responses and using them to improve the automotive system performance based on the data-driven based algorithms. Two applications included are multivariable borderline knock prediction and control and tire-road friction coefficient estimation. Internal combustion engines are core components of traditional and hybrid passenger vehicles and also widely used for off road applications. When the combustion is limited by the engine knock, it is desired to operate it as close to its borderline knock limit as possible to optimize combustion efficiency. Traditionally, this limit is detected by sweeping tests of related control parameters, which is expensive and time-consuming; and also, the detected borderline knock limit often is relatively conservative. When more advanced control parameters (subsystems) are added, these sweeping tests lead to tremendous higher test cost. An intelligent and efficient way to predict borderline knock without detailed knowledge of combustion dynamics is proposed. This supervised-learning based Bayesian optimization method is assisted by a surrogate model trained based on the system statistic properties. A two-control-parameter (spark timing and intake valve timing) case is demonstrated for optimizing two competing objectives (knock intensity (KI) and fuel economy).A complete borderline knock control structure is proposed and divided into three parts. The first part is about offline training with necessary modifications of the Bayesian optimization algorithm. Engine tests are conducted under two different operational conditions to obtain knock borderline limit, indicating the proposed algorithm is able to reduce required experimental budget (cost and time) significantly. The predicted mean Pareto front and its variance can be used to find the optimum control parameters at borderline knock limit for the best fuel economy possible. Smooth response surfaces of surrogate models can also be used as the initial model to be updated in real-time. The second part is an online updating process, based on the offline-trained surrogate model, using modified likelihood ratio controller. Principal component analysis indicated that spark timing is the most sensitive factor affecting the Pareto front. A two-buffer design was proposed to update the surrogate model under different rates so that both short-term compensation for environment changes and long-term for slow engine aging effect are covered. Both simulation and engine test results indicate that the proposed control strategy is able to update the machine-learned surrogate models in real-time, which outperforms the conventional knock control strategy and offline-trained knock limit, and especially reduces the conservativeness of borderline knock control significantly. Finally, to reduce cycle-to-cycle combustion variations, a real-time cycle-wised knock compensation scheme is developed based on the measured exhaust temperature when the engine is operated close to its knock borderline. To make model-based control possible, ?-Markov COVER (COVariance Equivalent Realization) system identification was used to obtain a linearized engine exhaust temperature model from change of spark timing to associated variations of exhaust temperature and knock intensity (KI). Accordingly, a Linear-Quadratic-Gaussian (LQG) controller is designed to minimizing the KI fluctuations based on change (?) of exhaust temperature. For the entire control architecture, results of three test scenarios indicated that the spark timing can be further advanced while maintaining the same knock intensity level due to reduced knock combustion variations.For the vehicle dynamics research, estimation of tire-road friction coefficient is very important due to new active safety control systems, especially for autonomous vehicles that rely on the accurate estimation of road surface conditions to find vehicle operational boundary and achieve the best performance possible. Several cause- and effect-based methods were proposed with their own limitations. A new evaluation criterion associated with slip-ratio is found based on CarSim simulation data on different road conditions; and strong correlation between proposed criterion and tire-road friction under different road surface conditions is observed. Note that the data-driven based method proposed in this dissertation only utilizes the statistic information from existing production vehicle sensors without increasing hardware cost. A computational cheap black-box model of proposed criterion and tire-road friction can be obtained and augmented with the existing dual-Kalman filter estimation algorithm, which improves tire-road friction estimation.
Read
