In the context of an increasingly automated and personalized manufacturing mode, efficient assembly sequence planning (ASP) has emerged as a critical factor for enhancing production efficiency, ensuring product quality, and satisfying diverse market demands. To address this need, our study first transforms the assembly topology and process into a weighted precedence graph, wherein parts represent nodes, and the assembly interconnections between parts constitute weighted edges. Then, we formulate the quantitative models of semantic knowledge, encompassing three facets: assembly direction changes, assembly stability, and part assembly interference, and thus constructs a heuristic function. We propose a novel dynamic graph learning algorithm, i.e., assembly-oriented graph attention sequence (A-GASeq), utilizing the heuristic information as edge weights of the assembly graph structure to incrementally direct the search towards optimal sequences. The performance of A-GASeq is first evaluated utilizing three key metrics: area under the receiver operation characteristic curve (AUC), precision score, and time consumption. The results reveal the superiority of our model over competing state-of-the-art graph learning models using a real-world dataset. Concurrently, we apply the algorithm to actual industrial products of diverse complexity, thereby demonstrating its broad utility across different complex products and its potential for addressing complex assembly sequence planning problems in the field of smart manufacturing.

在制造模式日益自动化和个性化的背景下，高效的装配序列规划（ASP）已成为提高生产效率、保证产品质量、满足多样化市场需求的关键因素。为了满足这一需求，我们的研究首先将装配拓扑和过程转换为加权优先图，其中零件代表节点，零件之间的装配互连构成加权边。然后，我们制定了语义知识的定量模型，涵盖装配方向变化、装配稳定性和零件装配干扰三个方面，从而构建启发式函数。我们提出了一种新颖的动态图学习算法，即面向组装的图注意序列（A-GASeq），利用启发式信息作为组装图结构的边权重，以增量方式将搜索引导至最佳序列。首先利用三个关键指标评估 A-GASeq 的性能：接收器操作特征曲线下面积 (AUC)、精度得分和时间消耗。结果揭示了我们的模型相对于使用真实数据集的最先进的图学习模型的优越性。同时，我们将该算法应用于不同复杂度的实际工业产品，从而展示了其在不同复杂产品中的广泛实用性以及解决智能制造领域复杂装配顺序规划问题的潜力。