Machine node-enhanced graph contrastive learning with long-range prompt model for quality propagation in multistage manufacturing systems

More complex multistage manufacturing systems (MMSs) have become mainstream production processes. Deep learning can accurately predict product quality indicators and help improve product qualification rates. Existing single-stage models can only predict the quality of a single machine or a single stage, ignoring the propagation effects. Although some multistage models consider quality propagation effects, they simply aggregate the machine features within a stage. In addition, existing multistage models do not conform to the actual production process when utilizing machine space relationships. Traditional models with severe label data dependence are difficult to make full use of noise data, resulting in reduced prediction accuracy and lack of interpretability. To address the above problems, this paper proposes a machine node-enhanced graph contrastive learning method with long-range prompt (MNGCLP), extracting propagation effects in both pre-training and fine-tuning. Specifically, the graph data is first used to model the production relationship between machines and form a production process graph. Then, a machine-enhanced graph is designed in contrastive-based pre-training to better add production information to avoid violating the production process and reduce label dependency. Finally, the long-distance relationship between machine nodes captured in the original graph is used as prompts to fine-tune the pre-trained model. Experiments in public production data show that the proposed method outperforms traditional models and provides a reasonable explanation for the prediction results, validating the effectiveness of MNGCLP. Compared with supervised graph neural network, the proposed model improves the overall RMSE, MSE and MAE by 2.4 %, 4.8 % and 9.8 %, respectively. Compared with the contrastive graph neural network, the proposed model improves the overall RMSE, MSE and MAE by 2.5 %, 5.0 %, and 6.3 %, respectively.