Generative prototype augmentation with diffusion priors for cross-domain few-shot industrial process monitoring in complex manufacturing systems

In current industrial contexts, the scarcity of labelled sensor data from freshly installed production lines often limits the reliability of intelligent monitoring systems. Conventional data-driven approaches usually fail to generalize across domains owing to differences in equipment settings, process dynamics, and operational environments. To address these issues, this research introduces GPA-DP, a diffusion-guided prototype augmentation system for cross-domain, few-shot industrial monitoring. The proposed technique uses class-conditional diffusion models to create representative target-domain samples while maintaining the complicated temporal properties of industrial data. Furthermore, diffusion-based augmentation is combined with adversarial domain alignment and prototype-based learning to enhance feature consistency across diverse manufacturing domains. Experiments on four industrial datasets, including semiconductor manufacturing, additive manufacturing, aerospace production, and automotive assembly, show that GPA-DP consistently outperforms existing approaches under limited-label conditions, improving accuracy by up to 18% compared to competitive baselines. Additional evaluations confirm the proposed method’s resilience under different degrees of domain disagreement and sample scarcity. GPA-DP, with its scalable architecture and lightweight inference method, offers a useful framework for intelligent quality monitoring and defect detection in complicated industrial contexts.