Data-Driven Anomaly Detection and Virtual Sensing for Engine Control

This doctoral thesis treats data-driven methodologies to enhance the performance of anomaly detection and virtualization in the context of engine control. Anomalies, specifically knocking combustion pose the major obstacle towards optimal thermal efficiency of an engine and considerable risk of critical damages, thus their detection assumes a priority role for meeting future regulatory targets for engine performance and lifetime. Virtualization on the other hand offers the potential to replace expensive, invasive or inaccessible sensors with mathematical model which adequately approximate these signals. Throughout the thesis, concepts to tackle both of these challenges are presented, with a special emphasis on those incorporating Convolutional Neural Networks (CNN which in their nature as deep learning approaches offer extensive possibilities of customization to any underlying problem. With regard to these opportunities, the work done leverages theory-guided principles of engine research to maximize the learning efficiency of a given model. Most effectively, the neural networks employed in this dissertation?s underlying work exploit frequency patterns of the concerned signals to detect excited resonances, or use in-model processing steps to split the virtualization task into independent estimations of distinct frequency bands. The research presented comprises novel modeling techniques, which are build from principles of theory and previous work from literature, compared to state-of-the-art models and also analyzed in terms of their limitations. Data for training and testing all models was acquired predominantly from heavyduty engines, which will still remain in operation in many sectors for the foreseeable future and therefore demand optimization in accordance with the targets of climate-specific regulations.