.png){kind=icon}
.png){kind=icon}
Return
A continuous learning framework for reliable defect segmentation with pixel-level uncertainty quantification
.png){kind=icon}
PRE
AI
DOI：10.1007/s00158-025-04173-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Constructing high-fidelity quality digital twins that mirror real-time production status is pivotal for realizing intelligent manufacturing. Nevertheless, the underlying deep learning-based perception models encounter severe challenges. One major limitation lies in the absence of prediction uncertainty estimation, which compromises the fidelity of the virtual mapping. Another critical issue is the scarcity of annotated data on production lines, which hinders the perception system from maintaining dynamic synchronization with continuously evolving physical processes. To overcome these challenges, a prediction risk self-awareness mechanism is proposed to enhance model fidelity. The mechanism is founded on a Bayesian U-Net and integrates a novel Uncertainty-Error Correlation Loss, explicitly driving the virtual model to learn the intrinsic correlation between prediction errors and uncertainty. This design enables autonomous reliability assessment of mapping results even in the absence of ground-truth labels. On this basis, a stream-based active learning system is developed to address the problem of dynamic synchronization, functioning as the engine for continual model evolution. By exploiting uncertainty obtained through self-awareness, the system establishes an intelligent feedback loop: when “low confidence” is detected in new samples, an alert is triggered, prompting engineers to provide annotations. The newly acquired high-value data are subsequently incorporated for model updating. Experimental evaluations on two industrial defect detection datasets demonstrate that the proposed framework surpasses existing methods in equipping perception models with prediction risk self-awareness and achieving efficient adaptive learning. This research provides an innovative methodology and enabling technologies for constructing high-fidelity and adaptive industrial quality digital twins, offering significant application value in real-world production scenarios where annotated data are scarce.
Keywords:
Interpretable machine learning
Pixel-level uncertainty quantification
Bayesian deep learning
Active learning of image dataset
Manufacturing quality assessment
Digital twin for quality assurance
Journal
IF:
4
Papers:
4.8K
Citations:
1.7W