Reproduce “Deep Vision in Smart Manufacturing: MODERN Framework for Intelligent Quality Monitoring and Diagnosis” on OpenResearch

About this paper

Yicheng Kang, Yuling Jiao, Xin Geng, Mahesh Nagarajan

Researchers from Miami University, Wuhan University, University of Miami, and The University of British Columbia developed the MODERN framework, a deep learning-based system for intelligent quality monitoring and diagnosis in smart manufacturing. It offers a generalizable solution for detecting persistent process shifts and accurately localizing defective regions in diverse industrial images, outperforming existing methods and providing theoretical guarantees on its efficiency.

Key insights

• The minimax optimal convergence rates for defect likelihood estimation and fault diagnosis establish strong theoretical guarantees for the MODERN framework's efficiency and statistical rigor.
• Counter-intuitively, increasing image resolution can slow down the convergence rates for both estimation and diagnosis, suggesting that higher resolution images may introduce more complexity for deep learning models with finite samples.
• The framework's performance is sensitive to image mislabeling, with a quantifiable asymptotic bound on the tolerable mislabeling rate relative to the sample size.
• The transfer monitoring technique, particularly the Augmented Bootstrap algorithm, significantly reduces the required sample size for adapting the framework to new settings, making deep learning-based quality control more practical.

Summary by alphaXiv.