Development of pass-by noise prediction models for road vehicles employing machine learning

This work applies machine learning techniques to real-world data for developing pass-by noise models capable of accurate short-term predictions. Several aspects surrounding vehicle noise and traffic noise modelling are investigated, including single vehicle pass-by noise prediction, traffic noise sound pressure level (SPL) prediction and traffic noise annoyance prediction.

The proposed approach combines object detection of video data with the development of traffic noise prediction models. The use of summary statistics is proposed to accurately represent traffic dynamic features based on individual vehicle trajectories extracted from raw video data. The traffic dynamic features are further defined as input variables of the machine-learning model, which can be continuously and flexibly adjusted if new input variables are demanded. Considering the temporal dependencies of the data samples, recurrent neural networks (RNN) are proposed for traffic noise modelling. A robust bi-directional RNN model is developed for accurately predicting continuous traffic noise SPL levels in short-term, using only traffic dynamic features as model input. Psycho-acoustic annoyance, which refers to the individual subjective perception, is defined as another output variable of the developed road traffic noise models. By taking into account annoyance as the additional output variable, the impact of excessive traffic noise can be better assessed.

The developed traffic noise SPL model and annoyance model are further interpreted by applying the method of explainable Artificial Intelligence, evaluating the influence of traffic dynamics on traffic noise SPL level and annoyance level from both local and global perspectives. The model interpretations significantly increase the transparency of the black-box machine-learning models, which can serve as an effective tool to support policy makers and urban planners in their decision-making process concerning traffic noise abatement measures. In short, the proposed system has the potential to support a continuous spatial monitoring of traffic noise and warning for excessive noise with real-time follow-up intervention measures, which can eventually contribute to smart mobility management under the framework of Smart City.