Automated Risk-Aware Calibration of Internal Combustion Engines

A major contributor to the emission of harmful greenhouse gasses and other pollutants is the transportation sector; therefore, the automotive industry strives to reduce the emission of these gasses by developing cleaner, more fuel efficient and sustainable combustion engines. These improved and more advanced engines results in a more complex and time consuming calibration process for the engine controller. Nowadays, this calibration process is performed by a calibration engineer. During this process, the actuator settings, feedback and feedforward controller parameters for a required driver demand are determined to maximize the thermal efficiency and limit mechanical stresses, pollutant emissions and cycle-to-cycle variation. In this thesis, a novel method is developed to automate the calibration process of internal combustion engines. This proposed calibration method does not require a first-principle model or prior knowledge about the combustion process. Instead, it learns the combustion behaviour during the calibration process using in-cylinder pressure measurements. The calibration process is guided by constrained Bayesian Optimization. To optimize the thermal efficiency, the method minimizes the difference between the measured in-cylinder pressure and an idealized thermodynamic cycle. This is equivalent to maximizing the thermal efficiency; however, this gives a more flexible description to the engine designer if other optimization properties are of interest that can be captured by the in-cylinder pressure. The constraints are formulated as the probability of violating safety and combustion stability limits. This serves two purposes: 1) to limit the mechanical stresses and cycle-to-cycle variations of the final calibration of the engine, and 2) to mitigate harmful behaviour during the calibration process by preventing exploration into regions with a large model uncertainty. A model is used to provide the required information for the optimization process. This model has been developed to predict the in-cylinder pressure during the compression and combustion stroke. The in-cylinder pressure is used to predict the combustion efficiency and constraint violation. The in-cylinder pressure is modelled using Principal Component Decomposition (PCD) and Gaussian Process Regression (GPR). This model includes information on the cycle-to-cycle variation of the combustion process. The calibration method and its individual parts are demonstrated on a Reactivity Controlled Compression Ignition (RCCI) engine. It shows the effectiveness of the proposed calibration and modelling method. The PCD/GPR model captures the behaviour of the modified single cylinder PACCAR/DAF MX13 engine present in the Zero-Emission Lab at the Eindhoven University of Technology. The automated calibration process is succesfully demonstrated on a simulated RCCI engine. The calibration method presented in this thesis proposes an innovative tool to reduce development time and cost in future engine development. Moving forward, the proposed method needs to be validated on a physical engine setup.