Characterization of the Spatiotemporal Behavior of a Sweeping System Using Supervised Machine Learning Enhanced with Feature Engineering

This paper focuses on geolocation data processing to infer the behavior of a mechanical sweeping system. A framework based on the feature engineering (FE) andmachine-learning (ML) tools for geolocation data processing is proposed. A supervised multi-classification machine learning using a large range of classifiers, input variables, training and data test sets is used to predict the sweeping system behavior. The results showed that Logistic Regression (LR) and Support Vector Machine (SVM) are the best classifiers for predicting the sweeping behavior and some simulated instances constituted the best training sets. The sweeping state prediction accuracy provided with LR and SVM classifiers, when trained with historical data, were in average 86.22% and 86.13%, respectively. These predictions using the same classifiers, when trained with simulated data, were in average 87.40% and 87.22%. These promising results illustrate the potential of integrating FE and simulation to enhance the performance the ML tools when studying the behavior of complex logistics systems.