Assessing air and noise pollution through acoustic classification of vehicles fuel types using deep learningHvastja, Andrej (Avtor) Ćirić, Dejan (Avtor) Milivojčević, Marko (Avtor) Prezelj, Jurij (Avtor) acoustic vehicle classificationpsychoacousticsHilbert-Huang transformationnoise pollutionair pollutionenvironmental acoustic analysisengine fuel typeMeasuring traffic emissions typically requires expensive analyzers, limiting the scalability of monitoring systems. In this study, we present a novel, cost-effective method for assessing noise and air pollution by classifying vehicles based on their acoustic signatures using artificial intelligence. We collected a dataset of 449 sound recordings of vehicles in an idle state within a real-world urban environment to minimize background noise and vehicle related variables. Psycho-acoustic features-loudness, sharpness, roughness, fluctuation strength, and tonality-and features derived from the Hilbert-Huang Transform (HHT) were extracted from the signals to capture the unique acoustic signatures of different engine types. Using these features, we developed a deep neural network (DNN) capable of distinguishing among petrol, diesel, and high-emission diesel vehicles with an accuracy of 92%. This approach demonstrates that acoustic emissions, i.e., traffic noise, can be linked to exhaust gas emissions. Our findings confirm that acoustic analysis provides an alternative to traditional methods for monitoring urban air quality. By enabling large-scale and dense sensor networks, this methodology offers substantial benefits for real-time environmental monitoring, urban planning, and regulatory enforcement. Additionally, the dual focus on air and noise pollution supports more comprehensive assessments of urban environmental impacts. This study highlights the potential of integrating advanced acoustic analysis with machine learning to create accessible, non-intrusive tools for pollution monitoring and mitigation.20252025-06-02 12:35:12Članek v reviji169530sl