Most anomaly detection systems in manufacturing operate on a single principle: a model is trained on historical data, launched, and is expected to perform well. However, real factories and production lines do not operate under ideal laboratory conditions. New types of defects appear, equipment changes, and experts accumulate experience that is difficult to formalize and even more difficult to transfer to a machine. The gap between what the model can do and what a human specialist knows only widens over time.
This is precisely the problem the LG AI Research team tackled. At the AAAI 2026 conference, they introduced a system called AEGIS – and its main idea is not to replace the expert, but to continuously learn from them directly during operation.
Why «Train Once» Doesn't Work
Industrial anomaly data is, by definition, rare. You are lucky to find a dozen defective items for every thousand normal ones. And new types of defects appear unexpectedly: when a batch of raw materials changes, a piece of equipment wears out, or the production process is altered.
Classical machine learning approaches quickly become obsolete under these conditions. A model trained on last year's data might not recognize an anomaly that was first encountered just last week. Retraining it from scratch every time is expensive and time-consuming.
At the same time, there is another problem: experts on the production floor can notice things that are not in the training data. But their knowledge typically remains in their heads – or, at best, in the form of instructions that the model cannot read.
A System That Doesn't Stand Still After Deployment
AEGIS is designed to continue evolving even after it has been deployed in a real production environment. Simply put, it doesn't just analyze incoming data – it includes a human in the learning loop.
When the system encounters something unusual, it doesn't just issue a verdict. It flags the uncertainty and asks an expert for labeling. The specialist looks at the object, says, «this is a defect» or «this is normal», and the system uses this information to adjust its performance. Gradually, the model becomes more accurate specifically on the data that a particular production facility deals with.
This approach is called active learning – where the model itself chooses what to ask the expert about, rather than receiving labeled data in random batches. This saves specialists' time: they answer truly important questions instead of manually sifting through thousands of examples.
Memory Without Forgetting – Harder Than It Seems
Herein lies one of the most difficult technical challenges faced by all continual learning systems. When a model learns from new data, it often «forgets» what it knew before. In the research community, this is called catastrophic forgetting.
Imagine a system has learned to recognize scratches on metal well. Then, it is further trained on data for a different type of defect. Now, it handles the new one well, but its performance on the old one has degraded. In a manufacturing setting, this is unacceptable.
In AEGIS, special attention was paid to this problem. The system is designed to integrate new knowledge into the existing base without destroying it. This is achieved through a special architecture and mechanisms that preserve the «memory» of previous defect classes when adding new ones.
Defects No One Has Seen Before
A separate challenge is the so-called zero-shot anomalies. These are situations where a new type of defect has just appeared, and the system has no samples for training.
Classical models simply don't know what to do in this case. AEGIS approaches the task differently: the system relies on text descriptions from experts. A specialist can say, «look for this type of deviation», and the system will try to find it, even without seeing any examples. This is made possible by using multimodal models that can correlate text descriptions with visual data.
Simply put, if an expert can describe a defect in words, the system will try to find it. This significantly reduces the time between «we've discovered a new problem» and «the system can detect it.»
How This Works in Practice
The LG AI Research team tested AEGIS not only on standard test datasets but also in conditions simulating real industrial operation: with a gradual influx of new defect classes, limited labeling, and expert involvement in the process.
The results showed that the system consistently maintains accuracy on known defect types while simultaneously learning new ones – which is non-trivial in itself. In traditional approaches, this is almost always a trade-off: either you remember the old well, or you learn the new quickly.
Moreover, the workload on experts was significantly lower than with manual labeling: active learning allows the system to «ask» only about cases where it is genuinely uncertain.
What's Behind This – Looking at the Bigger Picture
AEGIS is not just another defect detector. It is an attempt to build a different model of interaction between AI and experts in manufacturing.
Most industrial AI systems are designed to be static: train, deploy, use. Human operators are left out of the picture – they interact with the system but do not influence its development. AEGIS proposes a different principle: the system and the expert grow together, and the knowledge of one is gradually integrated into the knowledge of the other.
This also changes how we should think about implementing AI in manufacturing. Not «set it and forget it», but «set it and continually improve it together.» On the one hand, this requires more conscious participation from specialists. On the other hand, it makes the system significantly more dynamic and adaptive to the conditions of a specific facility.
The question of how well this approach scales to different types of production and how easily it can be adapted to conditions vastly different from AEGIS's testing environment remains open. But the direction seems logical – especially for industries where defects are costly, data is scarce, and experts still know more than any model.
