Research and development (R&D) programmes to optimise engine performance and emissions involve a large number of experimental variables and the optimum solution will normally be a trade-off between several measured responses (e.g. fuel consumption, exhaust emissions and combustion noise). The increasing number of experimental variables and the search for smaller improvements make identification of optimum configurations and robust solutions more demanding. Empirical models are routinely used to explore the trade-offs and identify the optimum engine hardware build and parameter settings.
The use of Bayesian methods enables prior engineering knowledge to be explicitly incorporated into the model generation process, which allows useful models to be developed at an earlier stage in the test programme. It also enables a sequential approach to experimental design to be adopted in which the ultimate engineering objectives can be more effectively taken into account.
Software has been developed to help engineers describe their prior knowledge in an intuitive manner, to update this prior information sequentially in the light of test results and to review the updated models graphically.
This paper reports on the use of Bayesian techniques in an engine test programme, with prior information being supplied by two engineers with different backgrounds and experience. A Ricardo “Ceres” high speed direct injection (HSDI) diesel engine has been comprehensively mapped with three independent variables. The resulting database has been used to demonstrate the Bayesian methodology (prior information gathering, model updating, model visualisation and data checking) and sensitivity to the quality of the prior information.