Abstract

A recently developed non-parametric regression (NPR) method, namely the component selection and smoothing operator (COSSO), was applied to a high-speed direct injection (HSDI) diesel engine optimization problem to analyse the complex correlations between NO_x and soot emissions, and fuel consumption responses and control factors. Seven control factors were considered, including swirl ratio, fuel injection pressure, intake air boost pressure, exhaust gas recirculation (EGR), start-of-injection timings for each pulse of two-pulse injections, and the dwell between the injections. The responses are non-linearly and non-monotonically correlated to the control factors, which feature complex interactions. The models were constructed by extracting information from an undesigned and unevenly distributed data sample, whose size is small with regard to the high dimensionality of the inputs. The models perform well, as shown by cross-validations. This success reveals the potential of NPR methods as powerful analysis tools for problems of this type.

Based on the current proposed well-constructed models, the importance of the control factors could be ranked according to their significance in the models, and interactions between control factors and their influence on responses could be quantitatively assessed. For example, a two-pulse injection featuring a very early first injection and a second injection well past top dead centre of compression provides excellent performance. The start-of-second-injection timing is shown to be most influential for fuel consumption and interactions between the boost pressure and the start-of-first-injection timing are important. Similarly, as expected, EGR is the most important main factor for NO_x emissions, while the main and interaction effects of the second-injection timing and EGR are also very significant. All main and interaction effects were significant for soot emissions. With a complete and continuous description of response surfaces available using the method, response values at untried parameter points can also be estimated, and possibly optimal solutions can be derived without being restrained by specified search grid resolution. Moreover, the solutions can be diversified to meet various criteria. The potential of using the proposed NPR method for assisting optimal parameter search is also discussed.