Abstract
The paper deals with CCV knowledge transfer from reference data (either experiments or 3-D CFD data) into system simulation SW tools (based on 0-D/1-D CFD). It was verified that CCV phenomenon can be modeled by means of combustion model perturbations. The proposed methodology consists of two major steps. First, individual cycle data have to be matched with the 0-D/1-D model, i.e., combustion model parameters are varied to achieve the best possible match of in-cylinder pressure traces. Second, the combustion model parameters (obtained in previous step) are statistically evaluated to obtain PDFs and cross-correlations. Then such information is imposed to the 0-D/1-D tool to mimic pressure traces CCV. Good correspondence with the reference data is achieved only if both PDFs and cross-correlations are imposed simultaneously.
Different engine operating points were evaluated to draw some general conclusions in terms of CCV. It was confirmed that turbulence properties and initial flame kernel development are the dominant factors. However, these factors are neither independent nor random – they seem to be correlated. Operating points with high CCV are more organized in terms of the statistics – they exhibit strong cross-correlations of combustion model parameters.
Knowledge Transfer for Heterogeneous Robots based on KL Divergence Regularization between Model Parameters
