A Multi-Task Learning Framework for Predictive Maintenance and Decision Support in Sustainable Manufacturing Systems

Abstract This study develops a data-driven decision-support framework for sustainable manufacturing by integrating multi-task learning with sensitivity and robustness analyses in a unified environment. It addresses production forecasting, maintenance detection, and efficiency classification within a shared feature space, providing a comprehensive view of manufacturing performance and decisions. Evaluated on an industrial dataset with 1,000 observations and 36 features, results show efficiency differences across tasks. The classification method achieved nearly perfect accuracy (98.5%) in estimating production for each task, but explained only 71% of the variability in forecasting volume. It failed to detect maintenance events, yielding no F1 score, due to feature limitations in identifying rare events, leading to a performance score of 56.4/100. Challenges from class overlap and feature-space limitations highlight the need for improved imbalance-aware models. Sensitivity analysis showed stable performance with more data and moderate declines under noise, indicating reliable learning. Robustness diagnostics identified predictable responses and sensitivity to random initialization, emphasizing the importance of stability. The framework enhances transparency, interpretability, and managerial relevance in manufacturing analytics, with practical implications for planning, maintenance, efficiency, and resilient supply chains. It advances the decision sciences and smart manufacturing fields by offering a scalable, interpretable tool for robust industrial analytics and a foundation for future integration with prescriptive optimization and digital twin systems.