Prior-guided dynamic heterogeneous spatio-temporal graph learning for geometric quality prediction in multistage assembly processes

Assembly geometric quality directly affects assembly reliability and delivery yield in complex products, and is central to quality prediction and control in multistage assembly processes (MAPs). In MAPs, multi-source errors propagate through heterogeneous relationships and accumulate across stages, resulting in spatio-temporal coupling in final geometric quality. Existing methods insufficiently characterize error propagation paths, inter-stage accumulation, and the relationship between subsystem local quality state and overall product quality, limiting prediction accuracy and interpretability. To address these limitations, this study proposes a prior-guided dynamic heterogeneous spatio-temporal graph network (PG-DHSTGN) for hierarchical multi-task prediction of geometric quality indicators (GQIs) before final inspection. Assembly-domain priors are encoded as relational semantics and edge features in dynamic heterogeneous graphs, providing graph-structured inputs with explicit assembly-engineering semantics. An edge-feature-enhanced heterogeneous graph transformer (HGT-EF) is designed, where edge features participate in heterogeneous attention computation and regulate message passing to characterize the differentiated effects of intra-stage heterogeneous topological relationships and propagation paths on quality evolution. A temporal attention gated recurrent unit (TAGRU) is further introduced to model inter-stage error propagation and accumulation. In addition, a hierarchical multi-task learning framework jointly predicts subsystem-level GQIs and product-level GQIs, while homoscedastic uncertainty weighting is used to adaptively balance multi-task losses. Results from an industrial case show that, compared with the average results of all baseline models, the proposed method improves MAE, RMSE, and R 2 by approximately 9.8%, 6.8%, and 1.7%, respectively. Ablation studies support the predictive effectiveness of the proposed components, while attribution analyses provide structured model-level predictive attribution.