Computational Optimization of a Diesel Engine Calibration Using a Novel SVM-PSO Method

Accelerated computational optimization of a diesel engine calibration was achieved by combining Support Vector Regression models with the Particle Swarm Optimization routine. The framework utilized a full engine simulation as a surrogate for a real engine test with test parameters closely resembling a typical 4.5L diesel engine. Initial tests were run with multi-modal test problems including Rastragin's, Bukin's, Ackely's, and Schubert's functions which informed the ML model tuning hyper-parameters. To improve the performance of the engine the hybrid approach was used to optimize the Fuel Pressure, Injection Timing, Pilot Timing and Fraction, and EGR rate. Nitrogen Oxides, Particulate Matter, and Specific Fuel Consumption are simultaneously reduced. As expected, optimums reflect a late injection strategy with moderately high EGR rates. The study shows that the optimization can be accelerated by approximately 75% while improving the ability to avoid local trapping using this novel Machine Learning - Optimization scheme.