Hybrid Intelligence for Root-Cause Analysis in Manufacturing: An Adaptive Knowledge-Driven Decision-Support Framework
This study proposes and validates a hybrid knowledge-driven decision-support framework that integrates semantic retrieval, FMEA-based engineering knowledge, and Ishikawa-based context assessment with a LightGBM model to achieve high-accuracy root-cause identification in manufacturing, effectively reducing the number of hypotheses experts must investigate while maintaining robustness against noisy data.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the complex world of modern manufacturing, where machines assemble intricate parts with microscopic precision, things inevitably go wrong. A coil might not wind correctly, a solder joint might fail, or a sensor might misread a signal. When these defects occur, they are recorded as non-conformity reports, often written by human operators in free-flowing text that describes what they saw and heard. The critical task that follows is root-cause analysis: the rigorous process of figuring out exactly why the defect happened so it can be prevented from happening again. Traditionally, this investigation relies heavily on the experience of quality experts who must sift through scattered notes, recall past incidents, and mentally cross-reference technical manuals to find the answer. While human intuition is powerful, it can be slow and inconsistent, especially when information is incomplete or when a defect looks like many different problems at once. The goal for engineers is to move from simply finding defects to preventing them entirely, a concept known as zero-defect manufacturing, but achieving this requires a way to organize the chaos of daily production data into clear, actionable insights.
A team of researchers from the University of Sannio and the University of Padova has developed a new approach to help solve this puzzle. They created a hybrid system that acts as a decision-support tool for quality managers, combining the ability of computers to read text with the structured logic of engineering safety manuals and the broad categories of human and machine factors. Instead of trying to replace the human expert, the system is designed to do the heavy lifting of sorting through thousands of past cases and technical rules to present a short, ranked list of the most likely reasons for a defect. The researchers tested this method in a real factory that produces electromechanical components, feeding it thousands of historical reports and production records to see if it could learn to identify the true cause of a problem as accurately as a seasoned human would.
The system works by listening to three different voices at once. The first voice is the text of the defect report itself. The computer uses advanced language processing to understand the meaning behind the words, searching through a massive database of past reports to find cases that sound similar, even if they were written with different words. The second voice comes from a technical document called a Failure Mode and Effects Analysis, which is essentially a pre-written map of what can go wrong in a specific machine or process. This part of the system acts as a filter, checking if a suggested cause is even physically possible given the stage of production where the defect was found. For instance, if a part failed during a final electrical test, the system knows to ignore causes that could only happen during the initial material inspection. The third voice looks at the context of the event: who was working, which machine was running, what batch of materials was used, and what time of day it was. By organizing these details into standard categories like people, machines, and materials, the system can see if the current situation matches patterns that have led to specific problems in the past.
These three streams of information are then brought together by a learning engine that has been trained on thousands of examples where the correct answer was already known. This engine does not simply add up the scores from the three voices; it learns how to weigh them against each other. If the text description is very clear but the machine data is missing, the system knows to rely more on the text and the technical rules. If the text is vague but the machine has a history of breaking in a specific way, the system shifts its focus to the machine data. This adaptability is crucial because factory reports are often messy, with missing details or hurried handwriting. The system is designed to handle this uncertainty, ensuring that a lack of information does not lead to a wrong conclusion, but rather triggers a different way of reasoning that still points toward the most likely answers.
When the researchers tested this framework, the results were striking. In a factory setting where there were twenty-three possible reasons for any given defect, the system successfully narrowed the list down to just three top candidates in nearly all cases. Specifically, the true cause identified by human experts was found within the top three suggestions in 97.5% of the cases when all information was available. Even more impressively, the system maintained high accuracy even when parts of the information were missing or when the text was noisy and difficult to read. In scenarios where the operational context, such as the machine or operator details, was completely absent, the system did not collapse; instead, it relied more heavily on the text and technical rules, still managing to identify the correct cause in over 80% of cases. This robustness suggests that the system can function effectively in the real world, where perfect data is a rarity.
The practical value of this work was demonstrated when the system helped a quality team solve a recurring problem with a specific product. The factory had been seeing intermittent failures during final testing, described in reports with slightly different wording each time, which made it hard to see a pattern. The system recognized that these different descriptions all pointed to the same underlying issue and, by cross-referencing the machine and measurement data, identified that the problem was not with the product assembly but with the testing equipment itself. The equipment had a loose contact and was overdue for calibration. Once the team focused their investigation on the test bench rather than the product, they fixed the issue, and the number of recurring failures dropped by more than 60% in the following months. This outcome highlights how the system does not just guess; it provides a structured path to the truth, allowing experts to focus their time on verifying a few strong hypotheses rather than searching blindly through dozens of possibilities.
The researchers emphasize that this tool is not intended to take the place of human judgment. Instead, it serves as a powerful partner that organizes the vast amount of information available in a modern factory, making the investigation process faster, more consistent, and easier to trace. By learning from every new case that is solved, the system continuously improves, building a shared knowledge base that helps the entire organization learn from its mistakes. The study confirms that combining the ability to read text, the logic of engineering rules, and the context of daily operations creates a more reliable way to find the root of manufacturing problems. While the system was tested in one specific factory, the approach offers a promising path toward more efficient and higher-quality production, turning the often chaotic process of fixing defects into a structured, data-driven science.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.