Development and Trial of a Non-Dominated Sorting Genetic Algorithm for Rapid Engine Calibration

In the quickly changing automotive powertrain landscape, calibration processes are a major bottleneck to production and new technology development. The implementation of rapid powertrain calibration processes can reduce development resource requirements, improve performance, and reduce risk. In previous works the author successfully developed an internal combustion engine calibration strategy using a Lagrange-Multiplier genetic algorithm for multi-constraint optimization. In the present work, this previous algorithm is expanded to create and demonstrate a new non-dominated sorting genetic algorithm which is aimed at identifying the entire Pareto front of the relationship between two engine performance parameters. Accurate and rapid identification of Pareto fronts enables designers to effectively and efficiently balance performance tradeoffs and identify optimal calibration settings. The algorithm developed here optimized the calibration of three combustion control parameters (stoichiometry, spark timing, and fuel injection timing) to define the Pareto front of the relationship between NMEP and NOx. The engine system consists of a 1.0L Ford Fox engine modified to operate on a single cylinder, which is operated at 1000 RPM, 1 bar manifold air pressure. First a design-of-experiments (DoE) test strategy was conducted to map the engine performance for a total of 100 steady-state engine experiments. These results were then used to train a neural network engine model, which was used to optimize different logic and parameter settings of the non-dominated sorting genetic algorithm. In trials using the neural network engine model, the optimized algorithm was able to successfully identify 60+% of the Pareto front for 100 total experiments, a factor of four improvement over the accuracy of the DoE approach (approx. 15%). Experimental validation of this performance is ongoing and initial results indicate similar levels of improvement over typical DoE approaches. Early experimental results indicated high sensitivity to the treatment of measurement uncertainty and therefore specific logic to address this uncertainty was added to optimize final algorithm performance.