Shuvom, M. and Haq, M., "Development and Analysis of Adaptive Neural Network Control for a Cybernetic Intelligent ‘iGDI’ Engine," SAE Technical Paper 2015-01-0157, 2015, doi:10.4271/2015-01-0157.
As combustion can vary widely between engine cycles if left uncontrolled, strict and robust control is required to meet optimum performance at different operating conditions. In this research, intelligent control techniques implemented on a Gasoline Direct Compression Injection (GDCI/GDI) engine. A research four cylinder 2.0 L GDI engine modeled with optimal control hardware that is frequently called as the conceptual Cybernetic intelligent GDI or ‘iGDI’ engine. The engine features Free Valve Actuation (FVA) hardware and precision fuel injector connected directly to the engine cylinder that found assistive for control flexibility by technical assessments. Then a mechatronic neural control approach is proposed and discussed with adaptive control techniques. Adaptive and predictive neural network control architectures developed for two distinct plant operation modes. The engine and the controllers are modeled and simulated with GT-SUITE and SIMULINK coupled simulation for control performance validation. High volumetric efficiency (∼97%) obtained overcoming pumping loss with throttleless operation. Transient simulation response at both conditions has been recorded and optimized with different neurocontrollers. The controller is trained for time varying plant dynamics with Nonlinear Autoregressive with eXogenous Input (NARX) neural network. AFR set-point tracking achieved with NARMA-L2 controller for minimum BSFC and NOx. This mode is termed as ‘ECO’ mode in this literature where 14% reduced BSFC obtained from the base model. Another application considered where Maximum Brake Torque (MBT) is required for the ‘POWER’ mode operation with NN predictive controller. The result obtained shows 15% improvement in BTQ from base model. The transient tracking response obtained satisfactory for iGDI engine control goals. Transient switching time between the modes obtained rapid on simulation timescale. Finally, plant performance evaluation is discussed in comparison with previous plant model upgrades and graphically shown in contours. It is clearly resulted that computational intelligence could effectively handle highly nonlinear dynamic real-time engine control problem with advantages of online optimization features and rapid prototyping.