Industrial artificial intelligence is undergoing a fundamental transition from traditional vision-based inspection systems toward multimodal reasoning frameworks that integrate vision, language, and structured decision-making. While modern deep learning systems achieve strong performance in anomaly detection and segmentation, they often fail to provide consistent, explainable, and spatially grounded reasoning required for deployment in real manufacturing environments.
In this talk, I present a unified perspective bridging over 20 years of semiconductor manufacturing and automation experience with recent advances in vision-language models (VLMs) for industrial anomaly reasoning. The focus is not only on improving detection accuracy, but on building trustworthy industrial AI systems that align perception, reasoning, and decision consistency.
We introduce the concept of evidence-grounded industrial intelligence, where AI systems are required to:
- Localize visual evidence of defects,
- Generate structured diagnostic reasoning aligned with visual cues,
- Ensure consistency between answers, spatial localization, and textual explanations.
A representative framework, EGVLR (Evidence-Grounded Vision–Language Reinforcement), is presented as a case study. EGVLR introduces a unified Evidence-Driven Diagnostic Protocol (EDDP) that enforces structured output alignment across reasoning traces, bounding boxes, and final decisions. The system further incorporates staged optimization including visual pre-alignment, instruction tuning, and geometry- and semantics-aware preference optimization to reduce hallucinated reasoning and spatial inconsistency in multimodal industrial tasks. Experimental studies on multimodal industrial anomaly benchmarks demonstrate that enforcing structured consistency significantly improves localization reliability, reduces false-positive spatial grounding on normal samples, and enhances reasoning stability compared to standard vision-language models. Beyond methodology, this talk emphasizes real-world deployment considerations in semiconductor manufacturing environments, including AOI (Automatic Optical Inspection) systems, AMHS logistics automation, and large-scale fab-level AI integration. We highlight practical constraints such as latency, robustness, and interpretability that are often overlooked in academic benchmarks.
The goal of this talk is to provide a practical roadmap for researchers and industry practitioners to design deployable, interpretable, and reliable multimodal AI systems for industrial inspection and beyond.
