Prior superpixel-driven multi-modal query hypergraph convolutional network for weld quality assessment

The quality of weld is critical to the reliability of structural connections and operational safety in applications such as automotive manufacturing, rail transportation, and aerospace. However, conventional single-modal welding quality assessment methods often suffer from limited information representation, weak robustness to interference, and poor adaptability to complex operating conditions. To address these limitations, this study proposes a prior superpixel-driven multi-modal query hypergraph convolutional network (PS-MQHGN) for high-precision weld quality evaluation. PS-MQHGN employs a cross-modal query hypergraph convolution module to explicitly model the information transfer pathways between different modalities, such as images and process signals, thereby achieving effective multimodal feature integration. In addition, a prior superpixel-driven dual-image interaction module is designed to actively focus on regions relevant to welding quality and to align up and down surface features based on label consistency. To further enhance performance, a dual-dimensional hypergraph feature aggregation mechanism is introduced to suppress the adverse influence of noise and weakly correlated hyperedges on model stability. In a representative resistance spot welding scenario, experimental results demonstrate that PS-MQHGN achieves high-precision weld quality prediction under complex welding conditions, with a mean absolute error of 0.1576 mm, corresponding to a 15.7% reduction compared with state-of-the-art models. PS-MQHGN provides reliable support for weld quality assessment and process control.