Factory Digitalization Ver16. _ The Limits of AI Relying Solely on Numerical Models
Aug 27, 2026
Thank you for reading.
■ Introduction
In my previous role, I worked on implementing AI in the production of automotive cast components. Through that experience, I came to recognize a clear limitation in AI systems that rely solely on numerical models.
Casting processes involve factors such as turbulence in molten metal flow, surface reflections, and mold wetness—conditions that directly affect quality but cannot be fully expressed as numerical data.
■ Integrating Process Images into Production Data
This led me to consider a different approach: capturing images at key points in the production process and integrating those images as part of the production data.
Images can record “process variability” that numerical data simply cannot capture.
■ Transforming the Inspection Process into a “Data-Generation Process”
Equally important is transforming the post-production inspection process itself.
By introducing an AI-based visual inspection system and applying it to 100% of products, we can accumulate inspection images for every item produced.
Traditional inspection leaves only a pass/fail result. It does not preserve the shape, spread, or characteristics of defects.
Furthermore, human visual inspection suffers from:
missed defects
inconsistency between inspectors
multiple defects being present but only the first one being recorded
To eliminate these issues, we need a visual inspection line capable of automatically classifying defect location, shape, and type, and recording all images.
Inspection then evolves from a simple quality gate into a “data-generation process that extracts defect characteristics.”
■ How AI Begins to Understand the True Nature of Defects
By correlating:
process images that capture variability during production, and
inspection images that capture the actual defect shape,
AI can begin to understand:
“Which process conditions lead to which defect shapes.”
Once this structure is established, the accuracy of AI-generated production prescriptions—used to suppress defects—improves significantly. This leads to a substantial reduction in defect occurrence.
■Why This Approach Extends Naturally to Food and Cosmetics Manufacturing
I am convinced that this approach can be applied directly to process‑sensitive industries, such as food, pharmaceuticals, and cosmetics.
These industries also involve process conditions that strongly influence quality but cannot be fully expressed numerically—such as dough firmness, viscosity, fermentation progress, surface dryness, color variation, or ingredient distribution.
■ The Next Stage of Manufacturing DX
Integrating numerical data × process images × inspection images is, in my view, the next stage of manufacturing DX.
■ Contact
I can translate the concepts in this article into a concrete system design tailored to your factory environment. If needed, I can also prepare system architecture diagrams, network diagrams, and implementation steps.
Please feel free to reach out.
Email: info@metricjapan.com
■ TAGS
コメント