The Cutting Edge of Generative AI in Manufacturing: How to Derive Insights Beyond Search Through data integration

What are the "data utilization barriers" that hinder the use of generative AI in manufacturing?

Japanese manufacturing sites have long faced common challenges related to data handling. While digitalization has progressed and various IT tools have been introduced, this has also created new difficulties in usability. We will delve into the factors hindering data utilization on the factory floor from three perspectives tailored to the realities of factories.

▼I want to know more about generative AI
⇒ Generative AI | Glossary

The existence of fragmented data across different departments

The first major obstacle is data fragmentation. Manufacturing operations are extremely broad, and optimal systems have been implemented for each purpose, such as design, development, manufacturing, and sales. Customer information is in the customer management system, order and sales figures are in the sales management system, line movements are in the production management system, and parts inflow and outflow are in the inventory management system, and so on.

While these systems are useful for improving efficiency in individual departments, the difficulty level increases dramatically when trying to optimize the entire factory as desired by management. To perform cross-functional analysis, data must be extracted separately from each system and manually combined using spreadsheet software. Data formats and item names often differ between systems, and this matching process alone consumes an enormous amount of time. As a result, despite possessing valuable data within the company, it remains too complex to be utilized effectively.

Reliance on expertise in data extraction

The second hurdle is the technical difficulty in accessing the data. Retrieving raw data stored in business systems generally requires specialized knowledge of database languages such as SQL. However, it is rare for factory workers to possess these IT skills, leading to a concentration of work in specific individuals or departments.

Therefore, even if someone wants to verify something on-site, they cannot look up the data themselves and in most cases have to request extraction from the IT department. However, the IT department also has a large workload, and it is not uncommon to have to wait several days, or even more than a week, from the time of request until the data arrives. The IT department is both the point of contact for data utilization and a bottleneck, which discourages on-site personnel from proactively exploring the data in depth.

Discrepancy between the speed of on-site decision-making and the provision of data.

The third obstacle is the time lag in obtaining information, which prevents on-site decision-making from keeping pace. Manufacturing processes are constantly in motion, and delays in decision-making directly translate into losses. For example, when a defective product occurs in a particular process, what the factory floor needs to know immediately is whether the cause lies in the raw materials, the equipment, or the worker's procedure.

If it takes several days to collect data to figure this out, even more defective products will continue to be produced during that time. By the time the analysis results are finally available, the information will have already become outdated, as production has moved on to another product. What the factory floor needs is an immediate answer to what is happening now, but traditional data utilization systems have made it extremely difficult to meet that speed. Reducing this time is a top priority for improving productivity.

Back to Top

Specific Examples of Generative AI Utilization in Manufacturing | On-site Data Utilization Use Cases

Recently, many companies have been considering the use of generative AI, and experimental proof-of-concept (PoC) projects are underway. However, most of these are limited to applications such as searching internal rulebooks or automatically generating meeting minutes. While these contribute to improving operational efficiency, the greater benefits for the manufacturing industry lie in using AI for more advanced data analysis.

The true strength of generative AI lies in its ability to interpret vast amounts of information across multiple systems and uncover relationships that humans might miss. It doesn't simply display data; it can instantly derive insights into what that data means and what actions should be taken next. Below are four specific examples of its application.

Visualization of priorities in equipment maintenance

In equipment maintenance, the conventional approach has been to conduct regular inspections or repairs only after a breakdown occurs. However, this often results in wasted resources, such as replacing still-usable parts or, conversely, having equipment fail just before an inspection, causing the production line to shut down. Future maintenance operations must embrace the concept of predictive maintenance, which involves anticipating failures and taking proactive measures.

Here, the AI compares the operational data from the Manufacturing Energy Management System (MES) with the maintenance records from the equipment management system.

• Example question: Identify which equipment has worn-out parts in relation to this week's production plan.
• Example Answer: We compared the cumulative number of shots in the MES with the parts replacement cycle in the equipment management system. Currently, the press machine on line C has reached 98% of its replacement threshold, and is assigned to manufacture product A, which is subject to high load. On the other hand, line D has equivalent equipment, but its wear level is 60%. We should either rearrange the production plan or prioritize inspecting line C.

In this way, by weighing the equipment's health against the expected workload, the AI can determine the optimal maintenance timing while ensuring safety.

Identifying supply risks in inventory management

In inventory management, it's not enough to simply track inventory levels; you also need to anticipate future shortage risks. By integrating and analyzing unshipped items from the order management system, inventory data from the ERP system, and work-in-progress information from the MES system using AI, significant cost reductions and inventory optimization become possible.

• Example question: For the backlog of orders, identify the items that are lacking in current inventory and work-in-progress.
• Example response: We compared the unshipped items in the order management system with the current inventory in the ERP system and the work-in-progress in the MES system. Even if all of the current work-in-progress of bolt B is completed, there is a shortage of 500 units to fulfill all orders. Since we cannot currently confirm the arrival schedule of raw materials, immediate procurement or a change in the order of commencement of the next process is necessary.

Because the AI checks even the contents of work-in-progress inventory, it can detect risks hidden behind seemingly sufficient stock levels early on, supporting an efficient supply chain.

Identifying the causes in anomaly detection

When the yield rate of good products declines, it is extremely difficult to pinpoint the cause. Whether it's equipment failure or human error, the answer requires combining multiple data points. Often, this involves the tacit knowledge and know-how of experienced workers, and many factories face challenges in transferring this technical expertise.

Here, the AI cross-references the MES inspection results with the HR system's skill map and the attendance management system's shift information.

• Example question: Regarding the recent decline in the yield rate over the past few hours, is there any trend in the combination of production line and personnel?
• Example Answer: We compared the MES inspection results with the skill map in the Human Resources and Training Management System. Currently, in the welding process on Line 2, failures are concentrated in the parts considered difficult, when workers with beginner-level proficiency in the relevant skills are assigned to those parts. Since there are no changes in the equipment parameters, we believe the cause is a mismatch in work assignments.

By enabling accurate situational assessment based on data, it prevents unnecessary guesswork on-site, contributing to smoother technology transfer and quality improvement.

Identifying the scope of impact in quality control

When a defect is found in a component, the most important thing is to immediately understand how far the impact has spread. By having AI instantly search MES traceability data, acceptance inspection records, and inventory management systems, highly accurate quality control becomes possible.

• Example question: Regarding the terminal continuity failure that occurred this morning, please check the impact on other products using the same component lot.
• Example response: We have reviewed the MES traceability data and the inspection records at the time of receiving the goods. The circuit board lot (Lot-99) used in the defective product is currently being used on lines 3 and 5. During the receiving inspection at the time of receiving this lot, the continuity resistance value tended to be close to the upper limit, although it was within the specifications. Strengthened inspections are required for all lines currently in operation that use this lot.

An investigation that would take hours manually can be completed in seconds with AI. This rapid response is the most effective way to prevent quality issues from escalating and protect the reputation of your brand.

Back to Top

The evolution of generative AI: From dialogue to autonomous notifications and actions.

The examples discussed so far have assumed a scenario where humans ask questions of the AI, but the range of applications is even wider. It is also possible to automate the process so that the AI constantly monitors data, detects any anomalies or signs, investigates the cause itself, and notifies humans.

For example, if the defect rate on a particular line starts to rise slightly, the AI autonomously collects relevant data and formulates a hypothesis about the cause. It then notifies the person in charge via email or chat tool. This eliminates the need for on-site personnel to constantly monitor data on screens, leading to improved work efficiency. By enabling proactive measures against problems, stress on the factory floor will be significantly reduced.

In this way, AI is evolving from a mere service or tool into a reliable partner that constantly monitors the factory's situation and provides appropriate advice when needed.

Back to Top

Key points for successfully utilizing generational AI in manufacturing (data integration and AI agents)

To achieve such advanced applications, there are two perspectives that are even more important than the AI model itself: data integration and the use of AI agents.

"data integration" - preparing the foundation for data.

To enable AI to make intelligent decisions, it must be provided with the right information. The first step is to organize the data scattered across various departments and systems into a format that AI can access.

• Standardize the names and units that differ across systems.
• AI organizes and stores information in a format that is easy to calculate and understand.
• We will build an environment that aggregates data from multiple systems in real time or periodically.

It may seem like a mundane task at first glance, but the more solid the foundation of this data, the more accurate the insights derived by the AI will be. The key to success is to proceed step by step, rather than trying to connect everything at once.

Autonomous "AI agents"

The next important aspect is the use of an AI agent function. This is a system where the AI itself thinks about how to solve a problem based on user instructions and takes action.

• The AI clearly determines which data to look at to understand the cause.
• It autonomously performs tasks from data extraction and comparison to identifying trends.
• Based on the analysis results, we will propose specific, practical actions that should be taken on-site.

In this way, by having AI take over the thinking and work processes, even people without specialized IT knowledge can equally benefit from advanced data analysis.

Back to Top

Finally

The challenges of data utilization in the manufacturing industry lay in structural barriers such as data fragmentation, reliance on specialized personnel, and time lags in decision-making.

As introduced in this article, by utilizing generative AI, it is possible to handle data from previously separate business systems across different systems, creating an environment where insights can be obtained immediately on-site, even without specialized knowledge.

This goes beyond simply improving operational efficiency; it's an approach that transforms the very system itself, enabling "anyone to utilize data immediately and across different departments."

Even starting with partial data integration is fine. Beginning small and gradually improving the speed and accuracy of on-site decision-making will be key to competitiveness in the manufacturing industry going forward.

If you're interested in creating a system like this, please feel free to contact Saison Technology.

Back to Top

The person who wrote the article

Affiliation: Data Integration Consulting Department, Data & AI Evangelist

Shinnosuke Yamamoto

After joining the company, he worked as a data engineer, designing and developing data infrastructure, primarily for major manufacturing clients. He then became involved in business planning for the standardization of data integration and the introduction of generative AI environments. From April 2023, he will be working as a pre-sales representative, proposing and planning services related to data infrastructure, while also giving lectures at seminars and acting as an evangelist in the "data x generative AI" field. His hobbies are traveling to remote islands and visiting open-air baths.
(Affiliations are as of the time of publication)

Recommended Content

Related Content