Semi-supervised quality prediction for label-scarce industrial processes based on dynamic co-regression with multi-scale vision transformer

Abstract

In industrial process quality prediction, limited labeled samples and noise interference often degrade the application performance of data-driven models. To address these issues, a Dynamic Co-regression Multi-scale Vision Transformer (DCMSViT) semi-supervised model is proposed to improve quality prediction accuracy in industrial processes. First, in order to avoid selecting noisy pseudo labels and low-confidence samples, a dynamic co-regression (Coreg) mechanism is designed, which adjusts the confidence threshold during training to improve the overall reliability of pseudo labels. Second, a multi-scale attention mechanism is embedded in the ViT, which calculates the self-attention weights in parallel across different temporal windows to capture local and global features. Finally, adversarial smoothing regularization optimizes the overall loss by separately evaluating the expanded labeled set and the remaining unlabeled set, thereby mitigating overfitting and improving generalization. Experiments on two industrial processes shows that the proposed model effectively addresses label scarcity and noise interference, thereby improving quality prediction performance.