Development cycle time reduction using design of experiments and machine learning-based optimization framework

Abstract

In this study, we have developed an in-house design of experiments (DoE) and machine learning (ML) based optimization framework for the engine calibration process. The new framework was compared to the traditional “one factor at a time” (OFAT) calibration process. The calibration process was carried out in a single-cylinder research engine for 7 bar indicated mean effective pressure (IMEP) at 1500 rpm. The engine control parameters such as the rail pressure, exhaust gas recirculation (EGR), start, and duration of injections for main and pilot injections were optimized to minimize indicated specific fuel consumption (ISFC) and smoke under constraint limitations of indicated specific oxides of Nitrogen (ISNOx), maximum in-cylinder pressure (Pmax), peak pressure rise rate (PPRR) and IMEP. The engine control parameters were optimized for the same load and speed conditions individually by two approaches for comparison. Machine learning models were built on the data obtained from both approaches. The ML models were used as a surrogate for experiments to screen the entire design space to optimize the objective function. Then, the optimized control parameters were validated through experiments. The machine learning-based optimization approach facilitated unraveling the global optimum faster. The DoE-ML-based process was able to identify a combination of control parameters which yielded a 14.7 g/kWh improvement in ISFC with other target parameters within the acceptable limits. Further, the DoE-ML-based optimization framework reduced the total number of experimental runs by about 40% compared to the OFAT-ML-based calibration process.