A feature-centric decision-making framework for diagnosing and enhancing system efficiency in intelligent multi-agent manufacturing

This research offers a feature-centric hybrid predictive framework to forecast the efficiency of the system in intelligent multi-agent manufacturing environments. By combining operational, learning-based, and cyber-physical indicators, the suggested approach caters to the growing demand for interpretable, resilient, and high-performance analytics in Industry 4.0/5.0 contexts. The paper presents a structured pipeline that involves recursive feature elimination for a principled feature selection, ANOVA-based sensitivity assessment for statistical variance attribution, and SHAP-based global explainability for model-embedded interpretability. To boost the predictive accuracy, three tree-based baseline learners-decision trees, random forests, CatBoost, and extra trees-are combined with two recent meta-heuristic optimizers: prairie dog optimization (PDO) and electric eel foraging optimization. The experimental results indicate that the hybrids, especially the PDO-enhanced random forest and extra trees models, lead to a significant increase in accuracy, stability, and error reduction across all the test stages. Sensitivity analyses continuously point out production efficiency, machine usage, Q-value, and security event as the main predictors, which confirms the multi-modal nature of industrial performance dynamics. The results emphasize the viability of feature-driven modeling and biologically inspired optimization in producing robust and interpretable outcomes that are suitable for practical smart manufacturing applications. This research adds a novel, explainable, and deployable predictive intelligence paradigm for modern multi-agent industrial systems as its contribution.