Search Results

The paper presents a new method based on neural networks to model the dynamic behavior of combustion pressure in SI engine cylinders, represented only by conventional input-output data. The approach is based on a functional representation of the pressure curve. The function parameters are adjusted by training a static neural network (SNN) for each working cycle. These parameters resp. “weights” are used in the following as reference pressure feature sequences. The sequences are simulated using time delay neural network (TDNN) as functions of engine speed, manifold pressure, ignition time and A/F ratio. The developed models can be used as stand alone models or as submodels within a global structure. It can be integrated as a real-time model in a HIL simulator to stimulate an ECU or implemented within an ECU for torque estimation. Performance of the proposed modeling strategy is verified by comparing experimental data from a test bench to real-time simulation results.

Real-time models, which reflect dynamic behavior of the SI engine, are needed for building up ECU testing devices like HIL simulators. In this paper the thermodynamic processes are reduced to some basic assumptions and combined with neural approximators of testbench data. So the parameters of the approximators can be easily adapted to similar new engines, while the principle structure describing interaction of the time- and angle-based processes remains unchanged. The model has been implemented and tested in a HIL-simulator. The performance of the proposed modeling strategy could be proved by comparing measurement data from a test bench to real-time simulation results.

SI engines are dynamic systems with highly nonlinear characteristics which are controlled by ECUs performing complex control algorithms. Hardware-in-the-Loop (HIL) simulation is an important tool to support test and verification during the development phase. The simulation model has to accurately reflect the dynamic behavior of the SI engine in the whole operating area. This paper describes a neural network approach to identify, i.e. to model a nonlinear dynamic system, the SI engine, represented only by I/O measurement data. The neural models have advantages with respect to robustness and measuring extent. They can be used as stand alone models or as sub-models integrated in a global model based on a physical structure. Measurements from a test bench compared to real-time simulation results prove the performance of the proposed modeling strategy.