Research on Data Analysis and Quality Prediction of Manufacturing Processes Based on Improved Deep Learning Algorithms

The smart manufacturing creates a vast volume of variously structured process data, but this quality prediction has difficulties in terms of non-uniformity of sampling rates, lack of labels, concept drift, and requirements to be sure about accuracy of confidence levels in decision-making. To address these challenges, we propose a Dual Path Time Transformer (DPTT) that combines a convolutional embedding branch with a masked self-attention branch, incorporating multi-task learning, uncertainty-sensitive loss weighting, and post-calibration. We also provide a one-stop solution for PLC/SCADA-MES-QMS data, including data alignment, robust interpolation, and process window creation. Results: For three representative industrial datasets, we achieve improvements of 2.8%–5.1% in classification F1 and reductions of 6.4%–11.7% in regression MAE. Compared to strong baselines, we demonstrate lower expected calibration error and greater robustness during distribution migration. Conclusion: The proposed technique is an effective method for process-level quality prediction. It considers uncertainty and is interpretable, making it applicable to large-scale applications in edge cloud environments.