Attention-enhanced schedule data imputation with advanced generative model
Wenqiang Xu, Ying Song, Jing Chen, Baixi Xing · Journal of Engineering Design · 2025
Although incomplete routing tables caused by design variations, sensor faults, and record keeping lapses continue to disrupt production scheduling, most existing research still assumes data completeness. We propose a Focus-Driven Attention GAN to reconstruct missing processing times while strictly honouring precedence constraints and machine compatibility. A context preserving embedding method encodes each operation along with its neighbours, while a binary Mark–Hint strategy directs model attention explicitly towards missing values. Within its adversarial structure, a novel focus-driven discriminator provides location-specific feedback, enabling fine-grained gradient propagation. The precedence aware generator integrates forward only intra-sequence attention, cross-token attention, and machine-wise attention to effectively capture long-range data sparse dependencies. Extensive evaluations on standard job shop benchmarks and a real world engine assembly line show that our method consistently achieves superior reconstruction accuracy compared to contemporary generative imputation models, enhancing downstream schedule feasibility.