Data Reduction Methods to Improve Computation Time for Calibration of Piston Thermal Models

Fatigue analysis of pistons is reliant on an accurate representation of the high temperatures to which they are exposed. It can be difficult to represent this accurately, because instrumented tests to validate piston thermal models typically include only measurements near the piston crown and there are many unknown backside heat transfer coefficients (HTCs). Previously, a methodology was proposed to aid in the estimation of HTCs for backside convection boundary conditions of a stratified charge compression ignition (SCCI) piston. This methodology relies on Bayesian inference of backside HTC using a co-simulation between computational fluid dynamics (CFD) and finite element analysis (FEA) solvers. Although this methodology primarily utilizes the more computationally efficient FEA model for the iterations in the calibration, this can still be a computationally expensive process. In this paper, several data reduction methods, such as principal component analysis, data clustering and resampling, sensor reduction, and uniform bin sampling are investigated to improve computation time while minimizing reduction in accuracy of the inference results. Each data reduction method is compared to a control case to determine change in accuracy and improvement in run time. Results indicate that most reduction methods were no more effective than using a smaller Latin hypercube design to inform the Gaussian process within the Bayesian inference code. Reduced error was observed for the structured sensor reduction method, indicating that further studies on the value of individual sensor locations to the overall calibration might be a viable path to reduce the computation time of the calibration methodology without compromising accuracy.