Manutenção preditiva de superaquecimento em motores de ônibus urbanos com aprendizado de máquina

Abstract

Proper maintenance of urban public transport vehicles is essential to ensure service reliability, reduce operational costs, and mitigate environmental impacts, particularly in high-intensity operating contexts. In this scenario, machine-learning-based predictive maintenance approaches have been widely proposed as alternatives to traditional strategies, enabling the early identification of failures from operational data. This study investigated the feasibility of early prediction of engine overheating events in urban buses using real telemetry data aggregated at the trip level, obtained from a commercial fleet management system in real operation. Supervised models, including Random Forest, Support Vector Machines, and Artificial Neural Networks, were evaluated in combination with different feature-selection strategies and retrospective temporal windows defined from the date and time of maintenance work order openings associated with thermal alerts, which were used to label trips as critical or normal. Explanatory variables were obtained by aggregating telemetry events and operational data at the trip level. The models were trained on historical data and evaluated on a temporally independent test set using metrics appropriate for imbalanced scenarios, including AUC, F1-score, precision, recall, and MCC. The results indicated globally limited performance and high variability across scenarios. Although, in the best experiments, the AUC reached values up to 0.875, the metrics associated with critical-class detection remained low, with F1-scores below 0.36, precision below 0.25, and MCC below 0.34, reflecting prediction collapses to a single class and low discrimination between normal and critical trips. The analyses showed that these limitations are strongly associated with insufficient telemetry granularity, the scarcity and heterogeneity of critical events, and the quality of labels derived from maintenance processes. It is concluded that, in the current state of the dataset, early prediction of engine overheating presents significant restrictions for operational use, and structural advances in instrumentation, data collection, integration, and recording are required to enable future predictive maintenance applications in urban bus fleets.