• Acoustics
• car tire
• Deep learning
• pavement distress
• tire cavity noise

The purpose of the following work is to discuss a new approach to the identification and classification of the various types of deterioration of the road pavement through deep learning techniques applied to tire cavity noise (TCN) signals. The work is mainly focused on the possibility of realizing, through neural networks, a digital system for real-time identification of various types of road pavement degradation.
This work starts from a research developed as part of the SurfAce project, which led to the construction of a specific sensor for the evaluation of noise inside a car tire and a system for the acquisition and recording of sound levels, as well as video data detected with a specific camera. The measurement system aims to assess the state of degradation of the road pavement under investigation with measurements along the infrastructure carried out with a specifically equipped mobile laboratory.
In order to define a basis for future developments, different, achievable objectives were proposed: primarily, a solid analysis procedure was prepared and its efficiency was tested in a real life measurement campaign, resulting in the solution of some synchronization issues between the different instruments of the apparatus and leading to the development of a video analysis software based on edge detection. The procedure was used to acquire a labelled data set that could be used for neural network training.
The second objective was to find a way to efficiently characterize acoustic noise data retrieved from the tire cavity. Different approaches were proposed and, where possible, features were further analysed through selection and reduction procedures such as the primary component analysis. The separability of the data sets in the resulting feature spaces was also addressed, primarily through K-Means clustering.
The third objective was to find the best deep learning based feasible approach in order to classify road pavement distresses from TCN acoustic signals. Three different neural networks were examined: a CNN operating on MFCCs, a CNN operating on the noise spectrograms and an LSTM operating on the time histories of previously selected signal descriptors.
The neural network was then implemented on a microcontroller to allow real time pavement condition monitoring. in the near future, the system may be perfected and similar instruments may be developed. The simultaneous use of TCN and video data may provide useful and precise information on road condition to competent authorities. The introduction of a phone app and the consequent use of geolocation may lead to the development of an efficient and cost effective method for road damage assessment.