Session: 04-01: Powertrain Diagnostics and Control Strategies

Paper Number: 148334

148334 - Machine Learning Assisted Modeling of Ignition Delay in a Light-Duty Gasoline Compression Ignition Engine 

Abstract: 

On the way to carbon neutrality of transport through the use of renewable and synthetic fuels, Gasoline Compression Ignition (GCI) combustion is considered a promising technology to achieve high engine efficiency and ultra-low pollutants emissions. As with other self-ignited combustions, a deep knowledge of the compression ignition dynamics represents the key for keeping the combustion process stable and controllable, especially for the first chemically driven combustion stage. The prediction of the ignition delay (ID) in different operating and environmental conditions is mandatory to guarantee engine performance and durability over time. Several approaches based on experimental evidence for ID estimation have been proposed but the tradeoff between costs and accuracy of the estimation is unfavorable. Machine Learning (ML) offers promising modelling tools to lower the cost for testing control strategies compared to traditional physical or empirical models. This work proposes an innovative Artificial Neural Network-based (ANN) approach to predict the ignition delay in a light-duty GCI engine using the information coming from standard sensors mounted on the engine. After an experimental characterization of the physical interactions between ID and the engine operating parameters, such as engine speed and load, intake and injection pressure, Exhaust Gas Recirculation (EGR), injections parameters, in a 1.3-liter light-duty turbocharged GCI engine, the ANN-based model was designed and calibrated. To define the best model configuration for the ID estimation using stock sensors, various training algorithms and Feedforward Neural Network (FNN) configurations were evaluated. Once identified the best FNN model, the results show that the ID estimation through the ANN-based model is characterized by relatively low NRMSE percentage of 2.16% on test dataset, demonstrating to be enough accurate for engine control and diagnosis purposes overcoming the limitations of the existing ID physical models when running far from the standard engine conditions used for the model calibration. Finally, the input sensitivity analysis reveals that the most important variables within the model are the thermodynamic, the engine speed, and the chemical potential of the in-cylinder charge, which further emphasizes the robustness of the model.

Presenting Author: Giacomo Silvagni Univerisity of Bologna

Presenting Author Biography: Dr. Giacomo Silvagni is a post-doc researcher at the University of Bologna (Italy). After his bachelor's and master’s in mechanical engineering at the University of Bologna, Giacomo focused his PhD and post-doc activities on studying and testing advanced combustion methodology in a compression-ignited engine. Besides that, his activity is focused on the development of control-oriented strategies and their implementation on a rapid control prototyping platform aimed at improving combustion controllability and engine performance and reducing pollutants for both conventional and advanced combustion methodologies. He was also involved in a research project with the University of Alabama as a research aide for studying alternative renewable fuels applied to dual-fuel heavy-duty engines. Besides combustion-related research tasks, His research mainly focuses on developing zero-carbon sustainable powertrains and propulsion systems, hydrogen-based energy plants, and sustainable aeronautical and space propulsion systems through experimental and simulation points of view.

Authors:

Giacomo Silvagni Univerisity of Bologna
Alessandro Rossi University of Bologna
Vittorio Ravaglioli University of Bologna
Enrico Corti University of Bologna