- Article
Industry 5.0 calls for decision-support systems that are human-centric, resilient, and transparent. However, in bulk-material logistics, predictive models are often developed as static tools, with limited attention to process structure, interpretability, and changing operating conditions. This study proposes a decision-support framework that combines the DMAIC cycle, machine learning (ML), explainable artificial intelligence (XAI), and human-in-the-loop model management, and applies it to rail unloading at a phosphate-handling facility. DMAIC was first used to structure process knowledge and identify the main drivers of daily throughput, including train frequency, wagon filling conditions, unloading duration, and operational losses. ML models were then evaluated for early prediction of final daily throughput, using only information available after the first train, with and without recent historical context. Adding historical features reduced validation MAE across all four tabular algorithms, with V2-XGBoost reaching a validation MAE of 2043 T/J and an RMSE of 2368 T/J. On the final chronological test period, however, all fixed ML models returned negative R2 values. Between the training and test periods, mean throughput increased from 19,070 to 22,233 T/J, and the proportion of seven-train days rose from 16.5% to 68.1%, indicating a marked regime shift. SHAP analysis was used to explain model behavior at the global and local levels, including regime-dependent predictive relationships. Based on these findings, we outline a human-supervised architecture linking early prediction, explanation, performance monitoring, and model review, in which operators and managers retain final decision authority. The study highlights the importance of evaluating predictive reliability under changing operating conditions in industrial logistics.
20 November 2026









