| Internal combustion engines experience a wide range of operating conditions, thus requiring design compromises to achieve satisfactory overall performance. However, there is strong motivation to make fixed parameters variable, permitting design constraints to be relaxed. This increases the system complexity due to increased degrees of freedom and complicated interactions among these independent control variables. Developing controllers and calibration maps becomes increasingly challenging, as the total number of experiments required for calibration increases exponentially with the number of independent control variables. Hence, the traditional calibration methodology, which generates look-up tables through systematic sweep tests, becomes prohibitively expensive. This study answers the challenge imposed by high degrees of freedom through development of a simulation-based algorithm.; A high-fidelity engine simulation tool is developed to predict engine performance corresponding to different control variable combinations. Pre-optimality studies are conducted to generate high-fidelity simulation benchmarks. Since optimization is very computation-intensive, it is not feasible to use the high-fidelity tool for solving optimization problems directly. Instead, Artificial Neural Networks (ANN) trained with high-fidelity simulation results are used as surrogate models. The ANNs are shown to be capable of representing complex relationships between multiple independent variables and selected engine performance indicators, such as brake torque, fuel consumption, NOx emissions, etc. Finally, the ANN surrogate models are employed in the optimization framework that searches the optimal combination of setpoints for any given driving condition.; The proposed algorithm is demonstrated on a conventional port-injected Spark-Ignition (SI) engine with two additional degrees of freedom introduced by the dual independent Variable Valve Timing (VVT) mechanism. The intake and exhaust camshaft positions are optimized for both wide open throttle and part load, using the appropriate combinations of optimization objectives and constraints. In addition, the capability of generating fast ANN models is utilized for developing a real-time air mass flow rate estimator for a VVT engine. With proper adaptation, the algorithm can be extended for complex engine and powertrain systems with even more degrees of freedom. |
