High Emitting Vehicle Identification using Machine Learning (HEVI-ML)

• Applicant: Dr. Pinky Kumawat
• Subject Area: Measurement Systems ; Atmospheric Science
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 567393450

Air pollution is a major environmental cause of premature deaths worldwide, requiring further reduction of pollutant emissions from road vehicles. Success, however, hinges on two factors, at least: that the initially low emissions of new vehicles stay indeed low over their entire useful life and that failure of the emission control system leading to high emissions are quickly detected and removed. Remote sensing can efficiently scan on-road vehicle emissions across the fleet, but the snapshot nature of remote sensing makes it difficult to track individual high emitters. Additionally, photographing number plates raises privacy concerns, limiting wider remote sensing use. To bridge these gaps, this project aims to develop new clever algorithms: (i) to “sniff” to which fuel type and emission standard a vehicle belongs, thus avoiding number plate reading; (ii) to determine the average emission from remote sensing emission snapshots for each individual vehicle and subsequently to check whether it is operating within legal limits (“clean”) or outside (“high emitter”). The algorithms will be developed for light duty vehicles first, and then the learning is transferred to heavy duty vehicles. The key idea is to determine “emission profiles” from portable emission measurement systems and chassis dynamometer measurements for each vehicle layer first. Onto these profiles, the snapshot data from remote sensing measurements are then “projected” to determine the probability for a normal or irregular operation. This requires sophisticated machine learning algorithms, including deep learning architectures, as well as bagging and boosting approaches. Big data and multiple parameters need to be analysed simultaneously to identify characteristic profiles. The plausibility of these models will be assessed using interpretable machine learning methods. As outcome, the detection scheme will enable real-time identification of high emitting vehicles based on on-road remote sensing measurements, while avoiding privacy intrusion. This would greatly boost the applicability of remote sensing measurements for routine in-service conformity monitoring.

• DFG Programme: Position