Automotive Engine Calibration with Experiment-Based Evolutionary Multi-objective Optimization


Abstract

The aim of this thesis is establishment of an overall framework of a novel control parameter optimization of automotive engine. Today, control parameters of an automotive engine have to be adjusted adequately and simultaneously to achieve plural criteria such as environmental emissions, fuel-consumption and engine torque. This process is called 'engine calibration'. Because many electronic control devices have been adopted for engine to satisfy these objectives, the complexity of engine calibration is increasing year to year. Recent progress in automatic control and instrumentation provides a smart environment called Hardware In the Loop Simulation (HILS) for engine calibration. In addition, Response Surface Methodology (RSM) based on statistical model is currently employed as the optimization method. Nevertheless, this approach is complicated by adequate model selection, precise model construction, and close model validation to confirm the precision of the model output. To cope with these problems, we noticed experiment-based optimization via HILS environment based on Multi-objective Evolutionary Algorithms (MOEAs), that is expected to be a powerful optimization framework for real world problems such as engineering design, as another automatic calibration approach. In experiment-based optimization, the parameters of a real system are optimized directly by optimization techniques in real time through experimentation. In this thesis, this approach is called Experiment-Based Evolutionary Multi-objective Optimization (EBEMO) and it is proposed as a novel automatic engine calibration technique. This approach can release us from burdens of model selection, construction, and validation. When using this technique, calibration can be done immediately after specifications have been changed after optimization. Hence, EBEMO promises to be an effective approach to automatic engine calibration. However, since conventional MOEAs face several difficulties, it is not easy to apply it to real engines. On the one hand, deterioration factors of the search performance of MOEAs in real environments have to be considered. For example, the observation noise of sensors included in output interferes with convergence of MOEAs. In addition, transient response by parameter switching also has similar harmful effects. Moreover, the periodicity of control inputs increase the complexity of the problems. On the other hand, the search time of MOEAs in real environments has to reduce because MOEAs require a tremendous number of evaluations. While we can obtain many measurements with HILS, severe limitations in the number of fitness evaluations still exist because the real experiments need real-time evaluations. Therefore, it is difficult to obtain a set of Pareto optimal solutions in practical time with conventional MOEAs. Additionally, plural MOPs defined by plural operating conditions of map-based controllers has to be optimized. In this thesis, to overcome the difficulties and to make EBEMO using the HILS environment feasible, five techniques are proposed. Each technique is developed through problem formulation, and their effectiveness are confirmed via numerical and real engine experiments. First, observation noise handling technique for MOEAs is considered. Because observation noise deteriorates the search ability of MOEAs, a memorybased fitness estimation method to exclude observation noise is introduced. Then, a crossover operator for periodic functions is proposed. Periodicity exists in engineering problems and leads to harmful effects on the performance of evolutionary algorithms. Moreover, the influence of transient response caused by parameter switching for dynamical systems is considered. In order to solve this problem, a solver of traveling salesman problems is used to determine the evaluation order of individuals. In addition, Pre-selection as acceleration method of MOEAs is proposed. In this technique, the generated offspring are pre-evaluated in the approximation model made by the search history, and then the promising offspring are evaluated in a real environment. Finally, parameterization of multi-objective optimization problems is considered. In engine calibration for maps, optimal control parameters have to be obtained at each operating condition such as engine speed and torque. This problem can be formulated in a form that needs to solve all of the plural multi-objective optimization problems defined by plural conditional variables. To solve this problem effectively, an interpolative initialization method is proposed. Through the real engine experiments, it was confirmed that EBEMO can achieve a practical search accuracy and time by using proposed techniques. In conclusion, the contribution of EBEMO for engine calibration is discussed. Additionally, the directions for future work are outlined.