Abstract: Aiming at the characteristics of strong continuous processes, complex variable coupling and significant raw material fluctuation in process production, this paper proposes a multi-process quality prediction method combining hierarchical graph construction and multi-level coding and decoding. Secondly, the macro chain-gated fusion module is used to depict the step-by-step transfer influence between processes. On this basis, the improved LSTM and self-attention module are combined to achieve high-precision prediction of time series feature extraction, and a unified representation across processes and scales is realized. Finally, the multi-process quality index prediction experiment was carried out for the three key processes on the silk production line of an enterprise after data preprocessing such as data cleaning, feature screening and normalization. The results show that the proposed model is superior to the comparison models in mean absolute error, root mean square error and goodness of fit, and the accuracy and stability of multi-process prediction are significantly improved. It can effectively capture the coupling within processes, the cumulative transfer fluctuation between processes and the time dynamic characteristics, which provides a reliable reference for the quality control and production optimization of the process industry.