Machine learning model for road asphalt monitoring system: vibration-based approach

Abstract

To achieve safe and correct driving, it is necessary to have a surveillance plan and the maintenance of highways and roads, in order to maintain a good infrastructure. Mexico has a paved and unpaved network of 780, 511 km, of which is paved 174, 779 km. According to statistics from the INEGI, in 2019, there were 9,318 accidents due to poor road conditions. There are several types of breakdowns on any paved surface, and they may differ depending on the country. For example, potholes, cracks, and patches are some road surface damages essential to assess in Mexico. In 2020, INEGI presents that 96.8\% of the population identified the issue of potholes in streets and avenues, as the problem with the highest frequency nationwide, above crime. Thus, the conditions of our roads are of deep concern for the population.

Different forms of road condition monitoring are proposed in the last years by specially designed instruments, using cameras, lasers, which require time and money and can only cover a limited proportion of the road network. Analogous to a video feed visually inspecting the asphalt's surface, a vibration-based system measures the ground conditions based on mechanical feedback from a vehicle. Different road anomalies, including potholes, cracks and ruts in the surface, create forces on the car, the frequency and magnitude of the forces will depend a lot on the type of anomaly. After we investigated different related works, this thesis is going to build on some of their aspects, and make a mix of others. The idea of dividing into three different categories for the classification of the roads, and the usage of supervised learning for road surface quality and anomaly detection. Regarding data collection, it was done through a phone with an Android system and an application created specifically for this job.

This thesis proposes a pothole detection model using a vibration base method, using built-in vibration sensors in smartphones. We collected road condition data in Mexico City using a dedicated vehicle and smartphones with a purpose-built mobile application designed for this study, splitting the data into: bump, bump, normal. A processing method was applied to the collected data, and features were extracted, then classified with a neural network. The results indicated that using only the subset of two of the three selected event types, together with their six characteristics, they outperformed other subsets in identifying potholes. Our neural network classifier showed classification performance, with an accuracy of 98\%.