ERI: Bridging Sensing, Learning, and Reasoning for Trustworthy Manufacturing Defect Inspection

This Engineering Research Initiation (ERI) award will focus on improving the reliability and interpretability of automated inspection systems in modern manufacturing. As manufacturing processes become increasingly varied and technologically advanced, products, materials, and production environments vary widely across applications. Differences in raw materials, processing parameters, equipment conditions, and environmental influences can produce defects that differ substantially in appearance, composition, and geometry. Such variability makes accurate detection and interpretation of defects increasingly challenging. Many current inspection systems rely on limited sensing information and are designed for relatively controlled conditions. They often do not explicitly account for uncertainty arising from measurement noise, surface variability, or changing production environments, and they provide limited insight into the underlying causes of detected defects. As a result, quality decisions may lack consistency and transparency under real-world operating conditions. This research addresses these challenges by developing analytical and computational approaches that combine complementary sensing information, explicitly consider uncertainty, and incorporate domain knowledge to improve how defects are detected, evaluated, and interpreted. By strengthening the scientific basis of automated inspection systems, the project contributes to advances in sensing and data-driven manufacturing research and supports the reliability of domestic production systems. The project also strengthens education by engaging students in research and integrating modern sensing and data-driven quality monitoring into engineering coursework, supporting workforce development and regional manufacturing capability. The specific goal of the research is to establish a theoretically grounded framework for manufacturing defect analysis. The research objectives include: (i) anomaly characterization through multi-modal sensing that integrates color imaging, hyperspectral imaging, and three-dimensional surface measurement to capture complementary physical attributes of defects across appearance, composition, and geometry; (ii) in-situ defect inspection under variable operating conditions through probabilistic learning models that explicitly represent both data-related and model-related uncertainty, enabling calibrated and reliability-aware inspection decisions; and (iii) root cause analysis through causal reasoning mechanisms that incorporate domain knowledge from structured and unstructured information sources to infer interpretable relationships between observed defect patterns and underlying material properties and processing conditions. By jointly modeling representation, uncertainty, and domain knowledge within a unified computational framework, the research contributes to the theoretical foundations of intelligent manufacturing inspection. The resulting framework supports improved reliability, interpretability, and generalization in automated quality monitoring systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.