This study presents a machine learning-based decision support system designed to predict welding operation durations in the galvanized steel pole production line of Mitaş Poligon, utilizing the geometric and operational characteristics of manufacturing components. Traditional manual scheduling approaches suffer from experience-dependent subjectivity, leading to production floor imbalances and capacity plan mismatches. To overcome these constraints, a 1,892-row raw enterprise dataset was extracted from the factory’s ERP system and comprehensively filtered to remove missing dimensions or non-standard entries, establishing a core baseline of 1,395 production records. To support model generalization and ensure structural completeness across all 54 product configurations (Part Type × Welding Type × Position), a 605-row synthetic dataset was generated via a physical arc-welding speed model coupled with a 6% Gaussian noise factor, forming a robust 2,000-row hybrid validation dataset. Following the structured CRISP-DM lifecycle, seven advanced machine learning architectures were cross-validated. The CatBoost regressor outperformed alternative models, yielding a Coefficient of Determination (R²) of 0.9584, a Mean Absolute Error (MAE) of 0.4490 min (~27 seconds), and a Root Mean Squared Error (RMSE) of 0.6421 min. The final optimization pipeline was compiled into a standalone Python/Tkinter desktop graphical user interface (.EXE) for localized, offline floor utilization without structural database dependencies. This data-driven model standardizes planning times, mitigates manufacturing bottlenecks, and provides an objective operational framework for sustainable capacity analysis. Furthermore, the system supports the United Nations Sustainable Development Goals by advancing Industry, Innovation and Infrastructure (SDG 9), Responsible Consumption and Production (SDG 12), and Climate Action (SDG 13).

Keywords

Machine Learning, CatBoost, Production Scheduling, Hybrid Dataset, Decision Support Systems, Sustainability