Physics-Constrained Mixture Density Network for Welding Process Data Augmentation

Process parameters in manufacturing have a critical impact on product quality. However, the scarcity and difficulty of acquiring high-quality sensing data severely limit the application of data-driven manufacturing. To address this challenge, this study proposes a physics-constrained Mixture Density Network (MDN) model for process data augmentation under extremely small-sample conditions. Taking the welding process of metal plates as a case study, welding arc energy is introduced as a physical constraint to ensure the physical consistency and credibility of the augmented data. Experimental results show that the mean deviations between the generated and true arc energy values on the test dataset are 1.93%, 0.26%, and 0.27% under three different conditions. Compared with other mainstream data augmentation models, the proposed method exhibits superior accuracy and stability. The developed model provides an effective solution for small-sample data augmentation and contributes to the advancement of intelligent sensing and big data analytics in manufacturing systems.