SEL-LGBM: A Lightweight Explainable Stacking Ensemble Learning Framework with LightGBM for Enhanced Industrial IoT Intrusion Detection

Widespread deployment of Industrial Internet of Things (IIoT) networks in modern production settings has enabled previously unheard-of levels of interconnectedness while also introducing significant cybersecurity risks. SEL-LGBM (Stacking Ensemble Learning with LightGBM), a novel hierarchical ensemble framework, is presented in this work. It is designed to address intrusion-detection challenges in resource-constrained IIoT ecosystems. The descriptive, multi-phase methodology provides a two-level stacking configuration, strategic class balancing using SMOTE, feature dimensionality reduction using Extra Trees, and preliminary data preprocessing. The system uses three base level classifiers namely DT, NB and LR in the first tier and LightGBM as the second tier meta learner to synthesize predictions in a more accurate manner. An important step forward is added by bringing in global feature importance appraisal along with local instance-level interpretation of decisions, all combined with deep SHAP based explainability procedures. In the descriptive experimental evaluation on the Edge-IIoTset benchmark it was observed that SEL-LGBM reached 89.37% identification accuracy and 89.65% weighted F1 score while also shrinking the feature space by 59%, going from 61 down to 25 attributes. The performance is validated through benchmarking studies against modern approaches that explicitly include descriptive decision transparency abilities. Overall this direction finds a sort of balanced mix between classification strength, resource thrift, and interpretability, and it aligns with those key industry application expectations.