Factory Digitalization Ver17. _ The Huge Gap Between IoT and AI
Oct 11, 2026
Thank you for reading.
This time, I would like to discuss the prerequisites for implementing AI in manufacturing, particularly for quality improvement and defect reduction.
In recent years, expectations for AI applications in manufacturing have grown significantly. However, much of the discussion focuses on AI itself, while relatively little attention is paid to the data foundation that AI relies upon for learning.
I believe that the success of manufacturing AI is not determined primarily by the AI algorithm itself, but rather by the quality of the training data used to develop the model.
In this article, I would like to explore the "huge gap" that exists between IoT-based data collection and effective AI implementation.
■ Introduction
The AI discussed in this article is not Generative AI, document search AI, or chatbots.
Rather, I am referring to AI applications that utilize manufacturing data for:
Defect prediction
Root cause analysis
Optimization of manufacturing conditions
These are AI systems designed to learn from production data and contribute to defect reduction and quality improvement.
When manufacturing AI is discussed, many people seem to imagine a process like this:
IoT Data Collection
↓
AI-Driven Quality Improvement
The idea is straightforward: collect data from equipment, train an AI model, and improve quality. Conceptually, this is not wrong.
However, in actual manufacturing environments, there are several critical steps between the two.
In my view, the reality looks more like this:
IoT Data Collection
↓
Manufacturing Data Aggregation
(Batch/Lot Level)
↓
Quality Data Integration
↓
Training Data Construction
↓
AI-Driven Defect Reduction
■ AI Learns Relationships Between Process Conditions and Quality Results
With IoT technologies, it is possible to collect enormous amounts of data from manufacturing equipment.
Temperature, pressure, flow rates, vibration, electrical current, and many other variables can be captured at sub-second intervals.
However, simply having large volumes of data does not enable AI to learn.
What AI truly needs to know is:
Which product was being manufactured?
Which batch or lot was being processed?
Which raw materials were used?
What was the final quality outcome?
AI does not learn from temperature values or pressure readings alone.
It learns the relationship between manufacturing conditions and quality results.
Therefore, equipment data by itself cannot serve as effective training data.
■ IoT Is a Starting Point, Not the Destination
Advances in IoT technology have made it much easier to acquire manufacturing data than ever before.
However, from an AI perspective, data collection is only the beginning.
Most manufacturing facilities possess many different types of information, including:
Equipment data
Production records
Quality data
Raw material information
Maintenance records
The problem is not whether these data sets exist.
The real challenge is whether they are consolidated into the same production unit.
Only when equipment data, production records, quality results, and material information are organized within a common manufacturing context can they become usable training data for AI.
■ Align Quality Evaluation Units with Manufacturing Data Aggregation Units
Ideally, every manufactured product would have its own unique identifier, with complete linkage between:
Individual product ID
Individual inspection result
Individual manufacturing history
For example:
Product ID: A000001 → Good
Product ID: A000002 → Defective
Product ID: A000003 → Good
If all manufacturing conditions were recorded at the individual product level, the resulting training data would be extremely valuable for AI applications.
In reality, however, full traceability at the individual product level is often difficult to achieve due to cost and operational constraints.
As a result, many manufacturers perform quality evaluations at the batch or lot level.
For example:
Pattern Number: 100
Production Qty: 1,000 pcs
Production Time: October 10, 2026, 09:00:00 - 11:45:00
Inspection Results:
- Accepted: 970 pcs
- Rejected: 30 pcs
Defect Breakdown
- Sand Inclusion: 15 pcs
- Misrun: 8 pcs
- Shrinkage Cavity: 5 pcs
- Dimensional Error: 2 pcs
In this case, the following information must also be aggregated at the same batch level:
Temperature history
Pressure history
Cycle times
Equipment status
Alarm history
Raw material information
If quality evaluation is performed at the batch level, manufacturing data must likewise be organized at the batch level.
If quality evaluation is performed at the lot level, manufacturing data must also be organized at the lot level.
Only when inspection results and manufacturing data are managed within the same unit can they become effective training data for AI.
■ Data Without Quality Results Cannot Become Training Data
When discussing manufacturing AI, I consider the following question more important than:
"How much data do you have?"
Instead, I would ask:
"At what level do you evaluate quality?"
The reason is simple.
Data without corresponding quality outcomes cannot become training data.
For example, even if billions of equipment data points exist, AI cannot learn quality-related relationships if we do not know:
How many products were produced
How many products failed inspection
What types of defects occurred
How frequently each defect type occurred
AI learns relationships between:
Quality Results
↕
Manufacturing Conditions
Without this connection, large quantities of data have limited value.
■ Building Training Data Is the Real Challenge
Discussions about manufacturing AI often focus on models and algorithms.
In practice, however, the more important challenge lies beforehand:
Are the required manufacturing data available?
Are they aggregated at the correct batch or lot level?
Are they linked to quality results?
Are defect classifications properly maintained?
Are there missing or inaccurate data points?
Only after these conditions are satisfied, effective training data can be constructed.
And only after effective training data has been established can AI generate meaningful results.
In many manufacturing AI projects, training data construction is a larger challenge than AI development itself.
■ Conclusion
I am not skeptical about AI.
On the contrary, I believe AI will play an increasingly important role in quality improvement and defect reduction.
However, AI performance is not determined solely by the sophistication of the AI engine. It is heavily influenced by the quality of the training data it learns from.
As AI adoption accelerates, the importance of:
High-frequency IoT data collection
Batch/lot-level manufacturing data aggregation
Quality data integration
Training data construction
is becoming more critical than ever.
The success of manufacturing AI is not determined when the AI system is implemented.
It is determined by how manufacturing data are collected, aggregated, and connected to quality outcomes.
Bridging the huge gap between IoT and AI is the first step toward realizing effective AI-driven defect reduction in manufacturing environments.
■ Support
If you require support related to manufacturing digitalization, IoT implementation, data utilization, quality improvement, or AI-driven defect reduction, please feel free to contact us.
Contact: info@metricjapan.com
■ TAGs
コメント