Autonomia de meios operativos para apoio à decisão em sistemas de comando e controle: uma abordagem preditiva e conceitual utilizando aprendizado de máquina

Abstract

In complex military scenarios, such as those faced by the Brazilian Navy (MB) in naval op- erations, logistical planning demands rigor, especially regarding the autonomy of operational elements, understood as the ability to operate without external resupply. This autonomy depends on factors such as supply consumption, the number of embarked personnel, and the operational conditions of naval platforms, acting in contexts supported by Command and Control (C2) systems. Although Machine Learning (ML) techniques are used to anticipate logistical demands, the semantic data heterogeneity and the lack of contextual domain un derstanding limit the effectiveness of purely quantitative approaches. To address these chal lenges, this work proposes the ATOp (Autonomia de Meios Operativos) approach, which integrates Ontology-Driven Conceptual Modeling (ODCM) and ML to enhance autonomy prediction. Grounded in the Unified Foundational Ontology (UFO), ATOp makes domain knowledge explicit and enables semantic data segmentation through the ontological meta category Situation, which represents contexts relevant to operational behavior. This seg mentation supports a regression pipeline with an ensemble of models, coordinating multiple specialized algorithms. The research included the development of the ATOp-NavalOntology, a well-founded ontology, data engineering from heterogeneous MB sources, and experiments involving Bayesian hyperparameter optimization and k-fold cross-validation, aimed at con structing the ATOp-PredictiveModel. The evaluation used metrics (R2 and RMSE) and statistical tests (Friedman and Durbin-Conover) to ensure reliable comparisons. The results show that combining well-founded ontologies and ML structures the data in a way that re veals relevant patterns, improves the quality of predictions, and outperforms individual mod els such as Gradient Boosting when applied directly to raw data. The ATOp-PredictiveModel achieved the best performance, with higher accuracy and lower computational cost, present ing statistically significant differences. As a contribution, ATOp offers a replicable model that integrates semantic knowledge into decision support in operational domains sustained by C2 systems.