Makoto Nagaoka

This paper considers machine learning based virtual design process of engine control system, and the demonstration in a diesel engine air path control is shown. This process contains two steps of machine learning. In the first step, a control-oriented forward model that predicts the transient behavior of the engine is learned from detailed engine model by using recurrent neural network (RNN). In the second step, an inverse model that determines the optimal control inputs to follow the references is learned from the numerical computation results of the offline model predictive control (MPC). The forward and inverse models could be used as a state observer and a controller, respectively, in a control system. An experiment of a diesel air path control system designed by the process was conducted using rapid control prototyping (RCP), and its following capability to the reference was demonstrated.

A direct numerical simulation of compressible turbulent channel flow at low Mach number(Ma = 0.3) subject to strong temperature gradient is conducted. The wall temperature ratios which are defined by a temperature on the upper wall divided by that on the lower one, are set to 2 and 3. It is shown that the flow of high temperature wall side is laminarized as increasing the wall temperature ratio. Moreover, compressible and dilatational motions are induced on the low and high temperature wall sides, respectively. The identity of friction coefficient originally derived by Fukagata et al.(2002) (so-called ‘FIK identity’) is extended to those of friction coefficient and Nusselt number on the compressible channel flow. It is revealed that the viscous variation component attains to 16 % of the friction coefficient on the case of high temperature ratio. Furthermore, the pressure work component occupies about 20 % of the overall Nusselt number for the all cases. The pressure work has a redistributive effect between mean kinetic energy and internal energy. The energy transfer from mean kinetic energy to internal energy appears on the low temperature side whereas that toward the opposite direction on the high temperature side at the present low Mach number flow.

A calculation method for steady compressible flows using a solution adaptive unstructured mesh is proposed and it's applicability is demonstrated for two-dimensional and axisymmetric engine intake flows. The mesh is generated by the rule-based Delaunay triangulation method. The governing equations are discretized by the finite volume method. The second order accurate Roe's upwind scheme is used for the inviscid flux calculation. The viscous fluxes are calculated by the transformation to the local general coordinates. The utility of the adaptive unstructured mesh method is shown in the 2D engine intake flow calculations. The present method using a low Reynolds number k-c model is valid for the velocity profile in the turbulent boundary layer. The calculated discharge coefficients with valve lifts in an axisymmetric engine agree well with the measurements.

圧縮性流れの解法において、非構造格子上で有限体積法による離散化と2次精度の空間差分スキームを用いる際、時間解法には前処理付きBi-CGSTAB法が当時一般的なRunge-Kutta法やガウスザイデル法に比べて定常解への収束が早くかつ安定であることを示した。