Virtual Sensor Algorithms for Estimating Climate Emissions
Basic Information
Students: Kalle Niskanen, Yejun Zhang, Anton West, Elias Silvola
Project manager: Kalle Niskanen
Instructor: Kai Zenger
Other advisors: Hoang Nguyen Khac, Amin Modabberian
Starting date:1.2.2022
Completion date: 3.6.2022
Abstract
Global warming and climate change are a growing concern in today's world. Greenhouse gasses (GHG) which contribute to this phenomenon are created during the combustion of fossil fuels. For instance, nitrogen oxide (NOx) and particulate matter (PM) emissions have been growing considerably in recent years. In particular, the shipping industry is responsible for a significant portion of GHGs, as one large cargo ship causes emissions equivalent to 50 million cars. These reasons have caused international organizations, such as the European Union (EU) and the International Maritime Organization (IMO), to take action and set new emission regulations to combat climate change. Therefore, engine manufacturers are forced to find novel solutions in order to comply with these new regulations. This has led to increases in both research and hardware costs, as conventional physical sensors for emissions are expensive and unreliable to use. Consequently, there is a growing demand for developing alternative ways to estimate GHG emissions. Virtual sensors could provide a plausible solution to this problem, if sufficient accuracy for the virtual sensor is reached.
In this project, data-driven virtual sensors for NOx and carbon dioxide estimation were successfully created. Principal component analysis (PCA) was used to extract features from cylinder pressure profiles. These features, along with additional engine data, were used as inputs to neural networks to train the virtual sensors.
Objectives
- Develop a functional virtual sensor for NOx
- Reach 5 % mean estimation error for NOx
- Create a virtual sensor for other emissions
Results
All of the project objectives were reached. A virtual sensor estimating NOx with a mean estimation error of 4.47 % was created. Additionally, a virtual sensor for estimating CO2 with a mean estimation error of 0.88 % was created. It was concluded that neural networks were suitable for training the virtual sensors that estimate climate emissions. As demonstrated previously in the literature, cylinder pressure played a key role in NOx and CO2 estimation. Because virtual sensors could be used to replace physical sensors, reductions in cost could potentially be made.
