Yoshihiro Sakayanagi

On-line Feedforward Map Generation for Engine Ignition Timing Control

This paper newly proposes an on-line learning method for control maps by using Gaussian filters. In the present method, man-hours for calibration of control maps can be decreased, and complicated structures of control maps are learned without prior knowledge. Moreover, by the effect of Gaussian filtering, smoothed maps can be created even under noisy conditions or few measured points. We also introduce improvements of the algorithm to cope with engine deterioration due to aging. In this work, the proposed method is applied to minimum advance for best torque control on actual vehicles with just one driving data, and the accuracy of the learned map is verified through simulation and experiments.

Knocking detection in gasoline engines based on probability density functions: A mixed Gaussian distribution approach

This paper proposes an engine knocking detection approach based on probability density functions (PDFs). In this work, we suppose that the PDF of knocking intensity distribution can be approximated by a mixture of two Gaussian functions due to normal combustion and abnormal one. We first apply EM (Expectation-Maximization) algorithm to actual engine data to show that the knocking intensity probability distribution can be successfully estimated by the mixed Gaussian distribution. We next try to apply the existing online EM algorithm in view of the actual implementation by an engine control unit. However, since the existing algorithm is built for the estimation of a fixed PDF, the effect of new input data is gradually attenuated and vanishes after long time. We thus newly propose a modified online EM algorithm so that the effect of the new input data is not attenuated. Finally, we perform simulations and an actual engine bench test for the validity of the present approach.