• acoustic signal processing
• AI
• embedded system
• machine learning
• MFCC
• microcontroller
• neural network
• road surface monitoring

Technology is continuously improving our lives and finding new solutions to old problems. Road surface monitoring is a worldwide challenge in the field of roadway infrastructure, and it has a huge economic impact. The state of road network is ruined in many areas of the whole world due to bad weather and traffic volume that endlessly stress streets. The inadequate road condition results in many undesirable factors, e.g., car accidents, higher gasoline consumption, bad driving quality and longer driving times.
Therefore, it is essential a continuous monitoring and maintenance to prevent road from deteriorating.
Nowadays, most steps of the evaluation are done manually by an inspector who drives along the road, collects raw data, identifies the type of defects and their location, and calculates a specific index for road surface classification (RCI). The current practice is time-consuming and depends on the experience of inspectors and their perception of pavement anomalies. In the years, some road monitoring sensors have been developed but these are not widely used due to limitations in cost or their susceptibility to weather and light conditions. It is increasingly important the design of new low-cost and reliable devices for a systematic and continuous surveillance of road defects.

This work aims at providing a new system able to evaluate the framed distress using solely a tiny board and an analog microphone placed inside the tyre cavity. It is proposed a novel method for real-time monitoring of road surface health. It is done through an embedded system that records, processes and classifies the audios thanks to AI-based tools. Contrary to other applications, all these steps are executed on a board. The developed classifier is able to distinguish between silence, unknown that includes both dirty and grass road, good quality road and pothole_badroad, a last class to which pothole and bad quality road belong.
When the MCU identifies a category, a string corresponding to the detected label is sent via BLE to an external device from which it is possible to perform further processing, for example, to indicate the anomalies on a maps app. This is a point of strength that limits power-consumption and data loss respect to when it is sent the whole raw data.

For this application, two different neural networks are compared and evaluated on several groups of tests: Tiny model and Conv model. The convolutional neural network selected to be deployed on the board is the Tiny because it is characterized by the smallest size (~18 kB instead of over 30MB) but at the same time conserves high accuracy (of 90% instead of 93.8%). Then, the classifier is quantized and converted in a suitable form to be deployed on the embedded system. The functioning of the application is evaluated on a board purposely developed by DII for this project. The experiments show that the microcontroller processes, classifies and communicates through Bluetooth Low Energy with external devices in real-time.