Enhancing Knock Detection and Combustion Diagnostics : A Machine Learning Approach with Ion Current Sensors
Björnsson, Ola LU
(2025)
- Abstract
- Achieving robust and efficient control of internal combustion engines is critical for meeting stringent environmental regulations and reducing greenhouse gas emissions. Natural gas engines, in particular, offer significant potential for cleaner energy but present unique challenges due to variations in fuel composition and combustion dynamics.
This thesis investigates novel methods for improving combustion diagnostics and control, focusing on ion current measurements as a cost-effective and versatile tool.
The research encompasses three main areas: (1) closed-loop control of the combustion phase using ion current-based diagnostics, (2) development of a virtual pressure sensor leveraging artificial neural networks, and (3) knock... (More) - Achieving robust and efficient control of internal combustion engines is critical for meeting stringent environmental regulations and reducing greenhouse gas emissions. Natural gas engines, in particular, offer significant potential for cleaner energy but present unique challenges due to variations in fuel composition and combustion dynamics.
This thesis investigates novel methods for improving combustion diagnostics and control, focusing on ion current measurements as a cost-effective and versatile tool.
The research encompasses three main areas: (1) closed-loop control of the combustion phase using ion current-based diagnostics, (2) development of a virtual pressure sensor leveraging artificial neural networks, and (3) knock detection using ion current as a standalone sensor and in combination with vibration sensors.
All experiments were conducted on a 13-liter, six-cylinder, heavy-duty natural gas engine.
The first part of this work examines ion current-based closed-loop control of the combustion phase.
By estimating the peak pressure location, the system dynamically adjusts spark timing to compensate for offsets caused by variations in fuel composition.
This approach effectively restored performance when the combustion phasing deviated from nominal conditions and improved combustion stability.
Next, a virtual pressure sensor was developed using artificial neural networks to predict in-cylinder pressure from ion current signals.
This method accurately predicted in-cylinder pressure, enabling precise estimation of combustion parameters such as peak pressure location, gross indicated mean effective pressure, and heat release.
These findings demonstrate the feasibility of replacing costly in-cylinder pressure sensors with robust and affordable ion current sensors.
The research also focused on knock classification using convolutional neural networks trained on ion current data.
These models achieved high accuracy in identifying knock events and were further improved by incorporating knock indicators from vibration sensors into a dual-input convolutional neural network architecture.
This combined approach reduced cylinder-to-cylinder variability and enhanced overall detection reliability.
The findings highlight the potential of ion current measurements, combined with machine learning, to provide cost-effective and robust solutions for combustion diagnostics and control.
By enabling engines to operate closer to the knock limit, these methods contribute to enhanced fuel efficiency and reduced emissions.
(Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/bce7f44f-9005-4e7c-b95e-6ecf8852facd
- author
- Björnsson, Ola
LU
- supervisor
-
- Per Tunestål LU
- Andreas Jakobsson LU
- Marcus Lundgren LU
- opponent
-
- Prof. Yamasaki, Yudai, University of Tokyo, Japan.
- organization
- publishing date
- 2025
- type
- Thesis
- publication status
- published
- subject
- keywords
- ion current, machine learning, knock, classification, spark-ignition
- pages
- 142 pages
- publisher
- Institutionen för Energivetenskaper Lunds Universitet - Lunds Tekniska Högskola
- defense location
- Lecture Hall M:B, building M, Ole Römers väg 1, Faculty of Engineering LTH, Lund University, Lund.
- defense date
- 2025-02-28 10:00:00
- ISBN
- 978-91-8104-372-3
- 978-91-8104-373-0
- language
- English
- LU publication?
- yes
- id
- bce7f44f-9005-4e7c-b95e-6ecf8852facd
- date added to LUP
- 2025-01-31 12:05:47
- date last changed
- 2025-04-04 14:47:20
@phdthesis{bce7f44f-9005-4e7c-b95e-6ecf8852facd,
abstract = {{Achieving robust and efficient control of internal combustion engines is critical for meeting stringent environmental regulations and reducing greenhouse gas emissions. Natural gas engines, in particular, offer significant potential for cleaner energy but present unique challenges due to variations in fuel composition and combustion dynamics.<br/>This thesis investigates novel methods for improving combustion diagnostics and control, focusing on ion current measurements as a cost-effective and versatile tool. <br/>The research encompasses three main areas: (1) closed-loop control of the combustion phase using ion current-based diagnostics, (2) development of a virtual pressure sensor leveraging artificial neural networks, and (3) knock detection using ion current as a standalone sensor and in combination with vibration sensors. <br/>All experiments were conducted on a 13-liter, six-cylinder, heavy-duty natural gas engine.<br/><br/>The first part of this work examines ion current-based closed-loop control of the combustion phase.<br/>By estimating the peak pressure location, the system dynamically adjusts spark timing to compensate for offsets caused by variations in fuel composition.<br/>This approach effectively restored performance when the combustion phasing deviated from nominal conditions and improved combustion stability.<br/>Next, a virtual pressure sensor was developed using artificial neural networks to predict in-cylinder pressure from ion current signals.<br/>This method accurately predicted in-cylinder pressure, enabling precise estimation of combustion parameters such as peak pressure location, gross indicated mean effective pressure, and heat release.<br/>These findings demonstrate the feasibility of replacing costly in-cylinder pressure sensors with robust and affordable ion current sensors.<br/>The research also focused on knock classification using convolutional neural networks trained on ion current data.<br/>These models achieved high accuracy in identifying knock events and were further improved by incorporating knock indicators from vibration sensors into a dual-input convolutional neural network architecture.<br/>This combined approach reduced cylinder-to-cylinder variability and enhanced overall detection reliability.<br/><br/>The findings highlight the potential of ion current measurements, combined with machine learning, to provide cost-effective and robust solutions for combustion diagnostics and control.<br/>By enabling engines to operate closer to the knock limit, these methods contribute to enhanced fuel efficiency and reduced emissions.<br/>}},
author = {{Björnsson, Ola}},
isbn = {{978-91-8104-372-3}},
keywords = {{ion current; machine learning; knock; classification; spark-ignition}},
language = {{eng}},
publisher = {{Institutionen för Energivetenskaper Lunds Universitet - Lunds Tekniska Högskola}},
school = {{Lund University}},
title = {{Enhancing Knock Detection and Combustion Diagnostics : A Machine Learning Approach with Ion Current Sensors}},
url = {{https://lup.lub.lu.se/search/files/207655661/Avhandling_Ola_Bj_rnsson_LUCRIS.pdf}},
year = {{2025}},
}