A robustified Newton based extremum seeking for engine optimization

Extremum seeking (ES) is a well known approach for online optimization of control parameters, e.g. in engine control. While the basic idea of ES is rather straightforward, in practice its application suffers from the problems related to determine the optimum numerically using measurements corrupted by noise. In addition, nonlinearities of the system under scrutiny, e.g. engines, can lead to a non convex objective function and thus to numerical problems. The purpose of this paper is to introduce a simply implementable extension to Newton based methods to improve the robustness of the convergence under real world conditions and to test it on a production Diesel engine. The extension is based on the regularization idea, and does not introduce significant additional tuning and setup effort. The results clearly show the improvements with respect to standard gradient and Newton based ES algorithms. The key advantage of this method is to provide convergence properties independently from the operating point and without re-tuning.