Robust Multi-Class Intrusion Detection Using Explainable Ensemble Intelligence for Heterogeneous UAV and Smart Manufacturing Networks

Abstract

The convergence of unmanned aerial vehicle (UAV) systems, industrial Internet-of-Things infrastructures, cyber-physical manufacturing platforms, and heterogeneous communication networks has created an increasingly complex intrusion detection environment. Conventional intrusion detection approaches frequently struggle with heterogeneous traffic distributions, evolving attack behaviors, class imbalance, overlapping network structures, and the requirement for interpretable security decisions. This paper develops a robust conceptual framework for multi-class intrusion detection based on explainable ensemble intelligence, with particular emphasis on heterogeneous UAV and smart manufacturing networks. The proposed framework integrates multi-objective representation learning, structural network analysis, ensemble classification, adaptive feature selection, and explainable artificial intelligence to distinguish multiple categories of malicious and benign activity. Community-structure research provides a theoretical basis for identifying latent relationships among network entities, while multi-objective evolutionary clustering supports the discovery of complex and dynamically changing traffic patterns. The framework further incorporates structural consistency, graph embedding, k-core analysis, label propagation, and objective-function selection as complementary mechanisms for representing heterogeneous attack surfaces. Explainability is incorporated at both global and instance levels to improve analyst confidence and facilitate operational response. The resulting analytical framework indicates that combining structural intelligence with ensemble decision-making can improve robustness against heterogeneous and evolving attacks while reducing the opacity associated with conventional high-capacity classifiers. The study positions explainable ensemble intelligence as a practical direction for security monitoring across UAV-assisted and smart manufacturing environments, while identifying computational overhead, concept drift, data imbalance, and cross-domain generalization as important limitations requiring further empirical investigation.