This study addresses limitations in current ensemble model approaches for AI-driven manufacturing by proposing an explanation-driven framework to enhance quality control systems. We introduce ExpLusion ( Exp lanation-driven L ate F usion ), a novel approach that integrates explanation-based late fusion strategies. Unlike traditional ensembles that combine model outputs based solely on performance metrics or learned weights, ExpLusion uses the similarity of local explanations (e.g., SHAP vectors) to select which models to aggregate for each instance. This explanation-driven selection ensures that only models exhibiting coherent reasoning patterns contribute to the final prediction, filtering out predictions based on noise or spurious correlations. Findings: ExpLusion was evaluated across multiple datasets representative of different industrial scenarios, including the SECOM and Bosch production line datasets. In particular, on the Bosch dataset, ExpLusion achieves an MCC of 0.521, outperforming all individual baseline models as well as the traditional late fusion approach (0.504 MCC). At the same time, it maintains a balanced predictive performance with precision values of 0.685 (negative) and 0.881 (positive), and recall values of 0.925 (negative) and 0.572 (positive). These results highlight the effectiveness of the proposed explanation-driven ensemble strategy for quality control in complex manufacturing processes. Future studies could explore ExpLusion’s scalability and adaptability in varied manufacturing contexts. Additionally, the framework’s dependence on explanation quality suggests further investigation is required to ensure robustness in diverse operational scenarios. By implementing intelligent fusion strategies that consider model-specific insights, manufacturers can achieve more consistent and reliable production outcomes, improving quality control efficiency. This research introduces a novel explanation-driven fusion framework for manufacturing quality control that uses local interpretability consensus as a model selection criterion. Rather than relying on performance-based weights or meta-learned combinations, ExpLusion prioritizes agreement in reasoning patterns, enhancing both reliability and interpretability in industrial monitoring contexts. • Introduces ExpLusion, an XAI-based late fusion framework for quality control. • Combines model-specific explanations to optimize ensemble predictions. • Validated using Bosch Production Line Performance dataset; improves accuracy. • Reduces false positives, enhancing precision and defect detection in manufacturing. • Offers insights into explanation-driven strategies for proactive quality control.
Ferraro et al. (2026) studied this question.