Traffic Load Estimation from Structural Health Monitoring Sensors using Supervised Learning

In the last decades there has been an evolution in the transport system, both for carriage of passengers and goods. In fact, the use of vehicles has become essential in the life of most of the population. This evolution has led to an increasing number of vehicles in circulation with the consequent increase in traffic congestion in travel times, road accidents and environmental problems related to pollution, especially near the large cities where urban traffic is often related to goods transport. Various Traffic Load Estimation (TLE) approaches have been implemented, which are essential for analyzing and managing vehicular traffic in the critical nodes of the road infrastructure. The standard approaches of TLE are based on the installation of dedicated sensors such as cameras or infrared sensors or by examining data from sensors present in smartphones. These types of approaches have various limitations, such as the large expense to install and maintain dedicated sensors or the need for user collaboration for smartphone-based approaches. Approaches based on the analysis of data from sensors, such as accelerometers, already present in the critical points of the road infrastructure for Structural Health Monitoring (SHM) have recently been studied. In this category, the proposed solutions identify and count vehicles by detecting anomalies without using labeled datasets. In this thesis, a TLE system is proposed using a supervised learning approach based on the data collected by the sensors of a SHM system present in a bridge in Italy. The datasets used in this study are accelerations grouped into time windows of various sizes. For each time window there are two labels corresponding to the count of light vehicles (such as cars) and the count of heavy vehicles (such as trucks) that have crossed the section of the bridge examined. These datasets are used to train Machine Learning (ML) models trained as regressors whose task is to estimate the count of light vehicles and heavy vehicles corresponding to each input time window. In order to carry out this task, various models of both classic ML and Deep Learning were tested and compared. From the experiments, the model that obtained the highest accuracy in the prediction of heavy vehicles was the Support Vector Regressor (SVR), which obtained a Mean Absolut Percentage Error (MAE%) of 7.6% and an R2 score of 0.97 in predicting heavy vehicles. On the other hand, the highest accuracy in the prediction of light vehicles was obtained by the K-Nearest Neighbors (KNN), which obtained a MAE% of 9% and an R2 score of 0.92.