Abstract:

Objective The rapid proliferation of personalized customization and fast-response orders in the garment industry necessitates dynamic scheduling for mixed-order production. Existing studies predominantly focus on single-order optimization, lacking frameworks to address multi-dimensional heterogeneity in fabric properties, order sizes, and equipment compatibility. A three-dimensional coupling model is developed to optimize cutting schemes, enhance resource allocation efficiency, and reduce production cycles. It is critical for enabling flexible manufacturing systems to adapt to small-batch, high-variability environments while balancing efficiency and equilibrium.

Methods A data-driven framework integrated machine learning and multi-objective optimization. Order attributes and cutting parameters were quantified into feature vectors. A classification and regression tree (CART) algorithm, trained on historical data, classified spreading units and matched them to compatible devices. An improved non-dominated sorting genetic algorithm II (NSGA-Ⅱ) was adopted to optimize scheduling with dual objectives, i.e. minimizing total cutting time (F₁) and maximizing production balance rate (F₂). Constraints included device availability and fabric-device compatibility matrices. Validation was carried out using real-world data from HL Customization Enterprise, involving heterogeneous equipment and mixed orders.

Results The proposed framework demonstrated significant improvements in production efficiency, resource allocation, and decision-making robustness through comprehensive validation with real-world data. The CART model identified fabric width as the most influential factor in device matching, with a feature importance of 0.299, followed by color (0.245) and texture orientation (0.102), while attributes like fabric elasticity and thickness showed negligible impacts. These criteria enabled the classification of 1 065 spreading units into seven device categories. Narrow-width fabrics measuring 91.5 mm or less were systematically assigned to devices P1-P3 and P5-P7 under rules such as limiting layers to 11.0 and texture orientation thresholds to 0.5. Orders requiring precision for wider fabrics, such as 144 mm materials, were routed to laser cutters (P4).The N S G A II helped achieve the dual-objective optimization, reducing the maximum completion time by 4.2%, from 291.95 to 279.63 hours, while improving the production balance rate to 89.96%. Pareto solutions revealed a spectrum of trade-offs, where the most time-efficient solution minimized the completion time to 279.63 h but yielded a lower balance rate of 70.54%, whereas the equilibrium-focused solution achieved an 89.96% balance rate at a marginally higher completion time of 284.64 hours. Total cutting time was converged to 274 h after 200 generations, with 78% of non-dominated solutions at generation 100 remaining optimal in the final set, indicating algorithm stability. Device workload variation was decreased sharply, as evidenced by the reduction in the coefficient of variation from 0.15 to 0.05, ensuring near-uniform utilization across 83% of equipment. Material efficiency improved significantly, with fabric waste decreasing by 23% through optimized device-fabric pairing. Orders with complex patterns, such as striped fabrics, achieved 98.2% material utilization when assigned to laser cutters, compared to 89.5% with conventional methods. The algorithm-maintained solution diversity, with crowding distances exceeding 0.8, and minimized premature convergence through a population size of 200, crossover probability of 0.6, and mutation probability of 0.4. Cross-generational analysis revealed a 12.3% increase in non-dominated solutions between generations 50 and 200. These results highlight the framework's capacity to reconcile conflicting objectives in mixed-order production. The integration of data-driven classification and multi-objective optimization enhanced operational efficiency while introducing adaptability to dynamic order influx, as demonstrated by a 27% reduction in rescheduling frequency for new orders. Empirical relationships, such as the inverse correlation between fabric width and device compatibility, further validated the model's alignment with real-world constraints.

Conclusions This study proposes a dynamic decision-making framework integrating the CART and NSGA-II to address resource allocation and efficiency coordination in mixed-order garment customization. The framework shortens the maximum completion time by 4.2%, improves production balance rate to 89.96%, and decreases material waste by 23% through optimized device-fabric pairing. The CART model identifies fabric width, color, and texture orientation as critical factors in equipment compatibility, while NSGA-II generates Pareto solutions that balance efficiency and equilibrium. The innovation lies in dynamically mapping order attributes to process parameters, overcoming limitations of single-objective optimization. The framework is scalable to discrete manufacturing sectors such as automotive and electronics, particularly for small-batch, high-variability scenarios. Future work should integrate real-time IoT data for adaptive scheduling and explore multi-stage production chain coordination. Practical implementation requires establishing process rule libraries and adopting digital twin technology for constraint simulation. This research advances intelligent transformation in garment manufacturing, providing an important reference for sustainable and flexible production paradigms.

Key words: customized garment production, multi-dimensional cutting scheme, CART decision tree, NSGA-II algorithm, dynamic scheduling optimization, garment production efficiency