Digital-physical process fusion based assembly quality prediction for pilot solenoid valves

In continuous automated assembly lines, the quality of pilot solenoid valves is formed through multi-stage process coupling, multi-source disturbances, and latent operating-condition drift, making early quality assessment difficult when relying only on end-of-line inspection. This paper proposes a digital–physical process fusion-based method for assembly quality prediction and virtual debugging. First, a manufacturing-semantic structured fusion preprocessing method reorganizes heterogeneous assembly records into single-sample representations with explicit process–quality correspondence, process-stage affiliation, variable identity, and production-context information. Second, a Feature-aware Residual Transformer (FRT) is constructed for multi-output quality prediction. By integrating process-feature tokenization, feature and operation embeddings, gated state-token fusion, Transformer-based interaction modeling, residual local compensation, and attention pooling, FRT models multi-stage process relationships and latent production-context effects. Experiments on real production-line data show that FRT achieves competitive and balanced performance against linear models, tree-based ensembles, sequential models, Transformer-based models, and tabular deep learning baselines. It obtains the lowest validation MSE and lowest test MAE, while maintaining test MSE, RMSE, and R² close to the strongest tree-based baselines. Indicator-level analyses reveal heterogeneous predictability across quality indicators. Finally, conditional virtual inference through controllable process perturbations demonstrates the potential of the trained model for preliminary process window exploration and low-cost virtual debugging.