Inovando a manutenção preditiva de caminhões pesados com aprendizado de máquina, dados de telemetria e de reparo

Abstract

This work addressed a central challenge in the Brazilian road freight transportation sector, which is responsible for approximately 60% of goods movement. The high maintenance costs, particularly in powertrain systems, directly impact operational efficiency and total cost of ownership (TCO). Using telemetry data from Euro 6 heavy-duty trucks, a data-driven approach was developed to proactively identify vehicles at higher risk of failure in a specific component, aiming to support predictive maintenance strategies that reduce warranty expenses, minimize workshop overload, and preserve the reputation of manufacturers. Supervised and unsupervised techniques were evaluated. Classification models, especially LightGBM, Random Forest, and XGBoost, showed superior performance in distinguishing between failing and non-failing vehicles. On the other hand, clustering methods such as FCM and FSOM, although less accurate, contributed to identifying operational patterns and vehicle similarity, offering valuable input for maintenance strategies and product development. FSOM, in particular, stood out for its flexibility in capturing relevant operational nuances, even though it performed slightly worse than FCM in some metrics. The feature importance analysis highlighted the relevance of both operational attributes and specific characteristics of the monitored components, emphasizing the need for an integrated approach in predictive modeling. Additionally, the simulated economic scenarios indicated that, with few exceptions, preventive component replacement was more advantageous than the estimated cost of avoided failures, even when considering penalties for unnecessary replacements. However, the choice of optimal thresholds should also take into account logistical factors and customer satisfaction, which are beyond the scope of this analysis. Among the limitations is the use of data from a single manufacturer and the focus on only one component.