Master thesis subject

The identification of noise sources is an essential step in environmental acoustics. In a large faction of cases disturbance is caused by many sources contributing to the general noise level. The identification of these sources can be difficult or extremely time consuming and this leads to the impossibility of designing proper mitigation actions. This is achieved through modeling the noise sources and assessing the achievable noise reduction. In the case of road traffic, in addition to the knowledge of the speed and the number of vehicles per hour, it is necessary to establish the class each vehicle belongs to. This is done by a human observer that has to stand by the measuring position, or watch some video record. This thesis work is about automating vehicle class recognition through the use of Artificial Intelligence methods.

We have recorded video and sound levels in bands of 1/3 octave from thousands of vehicles passages and we would like to develop deep learning techniques to achieve automatic classification. We intend to use Convolutional Neural Networks (CNN) to extract characterizing features out of the data. Having been optimized for image analysis and classification, CNNs are a powerful tool to study complex time series and have been used in frontier research like identifying noise sources in gravitational wave detectors[1,2]. Following this approach, tests on public environmental noise datasets have shown promising performances of CNN in terms of accuracy[3]. The approach would consider the signal described by 1/3 octave levels as function of time, building a time-frequency image of a passage.

We will consider supervised and unsupervised training. In the first case the network is trained against some predetermined classes, already recognized in the data. In the second the CNN will be free to determine its own classes.

This project is well suited for a master thesis for a student that has some experience with programming languages and is willing to learn how to use frontier tools in machine learning. Adequate computing power with a GPU equipped server is available.

References
[1] Razzano M. & Cuoco, E., Image-based deep learning for classification of noise transients in gravitational wave detectors, Class. Quantum Grav., 2005, 35 095016
[2] Sophie Bini, Unsupervised classification of short transient noise to improve gravitational wave detection, Master thesis, 2020, University of Pisa
[3] Piczak, Environmental sound classification with convolutional neural networks, 2015 Conference: 2015 IEEE 25th International Workshop on Machine Learning for Signal Processing