Analysis of Unsupervised Learning Approach for Classification of Vehicle Fuel Type Using Psychoacoustic Features

Milivojčević, Marko; Ćirić, Dejan; Prezelj, Jurij; Murovec, Jure · 2023 · OpenAlex-citations

DOI: 10.2139/ssrn.4584703

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of classifying vehicle fuel types (gasoline vs. diesel) using acoustic signals, motivated by the need for privacy-preserving, real-time monitoring systems to improve urban air quality and manage traffic flow. Unlike video-based identification, acoustic classification avoids privacy violations and can be implemented in locations such as underground garage entrances or gas stations. The research specifically investigates an unsupervised learning approach to detect hidden patterns in audio data without relying on labeled training sets, aiming for a robust, portable, and energy-efficient system. The methodology involves two main components: a custom-built, stand-alone portable sound acquisition system and a classification algorithm based on Self-Organizing Maps (SOM). The acquisition system utilizes ultrasonic sensors to detect vehicle presence and initiate recording, ensuring that only engine idle sounds are captured. A microphone is positioned on the ground below the engine compartment to minimize environmental noise interference, with recordings filtered to ensure sound pressure levels exceed 74 dB. The collected audio samples are processed to extract psychoacoustic features, including loudness, tonality, sharpness, as well as crest factor and zero-crossing rate. These features form a 10-dimensional input space. The classification is performed using a 1-D SOM neural network, which employs competitive learning to cluster the input data into distinct groups representing different fuel types. The results demonstrate that effective classification of vehicle fuel types is achievable using unsupervised learning with a limited set of features. The analysis reveals that while a 10-dimensional input space was used, effective clustering was demonstrated with just five single-valued psychoacoustic features. The SOM algorithm successfully mapped the acoustic signatures of gasoline and diesel engines into separate clusters, validating the distinct acoustic differences between the two combustion types even at idle speeds. The study highlights the complexity of the input space, noting that the topological properties of the SOM allow for the visualization of these hidden connections. The robustness of the system was confirmed through real-world data collection in a controlled garage environment, where the acoustic cavity under the vehicle helped isolate engine noise from external disturbances. The significance of this work lies in its contribution to privacy-friendly, non-intrusive vehicle monitoring technologies. By proving that unsupervised learning can effectively classify fuel types using psychoacoustic features, the study provides a foundation for developing IoT-based devices that can be deployed in urban infrastructure without the need for extensive labeled datasets or complex supervised training. This approach supports applications such as directing high-emission vehicles to better-ventilated areas or preventing incorrect fueling at gas stations. The findings suggest that while the current system is effective, further improvements in feature selection and algorithm optimization are necessary to enhance classification accuracy and robustness for broader real-world deployment.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• auditory warnings