2024-07-03 | Quality Control

How the Anomaly Detector model was used to automate the quality control process using only normal data



  • Pouch-type batteries have flexible surface geometry, complicating defect identification. This makes it challenging to use existing rule-based inspections and deep learning model-based inspections like classification and segmentation
  • Finding skilled inspection workers similar to those in Korea is difficult when expanding globally, necessitating the introduction of inspection automation



  • LISA’s two-step approach
  • (1) ‘Anomaly Detector’ algorithm trains a model using only normal data (unsupervised learning) and detects defects → Minimizes false negatives
  • (2) Human reviewers label normal data within NG data using classification and segmentation algorithms → Minimize false positives



  • Introducing AI inspection to a process where traditional rule-based inspections were not feasible due to the difficulty of specifying the form of defect
  • Lowered labor costs through inspection automation
  • Inspection stabilization time reduced by 67% (from 3 months to 2 weeks)

Full Story

Lack of automated inspection tools for easily damaged pouch batteries

Among the batteries that go into electric vehicles are pouch batteries. They’re stacked battery materials and wrapped in a thin material like aluminum foil. At first glance, they look like an envelope. They have the advantage of being tightly packed inside, providing high energy density and allowing for easy design changes.

The challenge is that these batteries are difficult to inspect for defects automatically. As mentioned earlier, the thin surface material is easily creased or dented/scratched with the slightest amount of force, which means that there is a lot of freedom in surface geometry, making it difficult to specify the shape of the defect.

In this case, a traditional rule-based solution, where the inspection is performed based on a set of good/bad rules, is not feasible. You’d have to enter hundreds, thousands, maybe even tens of thousands of rules.

Defect detection AI based on deep learning models such as classification and segmentation is also difficult to utilize because, as mentioned above, it is difficult to define the exact category of defects.

Company K, a producer of pouch-type batteries, relied on three shifts of skilled workers for visual inspections. However, upon entering the overseas market, this approach was no longer feasible. The problem is that it’s hard to find the same skilled workers as in Korea, and labor costs are too high.

They couldn’t put off automating their inspections any longer.

Two-step approach (unsupervised learning -> review -> supervised learning) for countless defect cases

The Anomaly Detection algorithm in LISA, an industrial AI solution, allows you to train a model using only normal data and then perform defect inspection. You don’t need to create abnormal data.

Quality checking and model training is a two-step process. First, the anomaly detection model is trained using only normal data (unsupervised learning). Then, defect detection is performed by labeling all subset data that deviates slightly from the trained normal data as defective. In short, it minimizes the number of “false negatives” that are NG but not filtered out by NG.

However, there are some that fall into the normal category. These are called ‘false positives’. The model is still too sensitive.

Then we have a human reviewer to verify that it’s a real defect or not and label it as normal or abnormal. We do a second round of training on that data (Supervised Learning). This utilizes algorithms like classification and segmentation. As this process is repeated, the false positives are gradually reduced and the model becomes more accurate.


  • Introducing AI inspection to a process where traditional rule-based inspections were not feasible due to the difficulty of specifying the form of defect
  • Lowered labor costs through inspection automation
  • Inspection stabilization time reduced by 67% (from 3 months to 2 weeks)

Using AHHA Labs’ industrial AI platform LISA, inspection stabilization time can be reduced to one-sixth of the previous duration, particularly when applied to new production lines.

This is thanks to the aforementioned two-step approach.

Traditional supervised learning-based AI inspection tools require at least three months for the model to complete training and become proficient in defect detection, needing to learn a substantial amount of normal and abnormal data.

However, LISA’s anomaly detection algorithm can achieve zero false positives in just two weeks using only normal data. It then continues to learn more and more to reduce false positives.

In short, you’ll be able to automate defect detection early on.



We are building an MLOps platform that is truly AI-driven. 

The AHHA Labs team is working towards another goal. During our consulting work with Company K, we realized a pain point: too much human intervention is required to actually “run” an MLOps platform.

While many MLOps companies often claim to be “fully automated,” the reality is that the really important decisions still need to be made by experienced and skilled AI engineers, such as which datasets to train on and how much more to perfect the model.

Customers don’t have AI experts, so they utilize automated MLOps platforms, but they need AI experts to utilize them, which is a paradox.

AHHA Labs’ MLOps platform, coming in Q3 2024, will be truly automated.

For example, AI will make recommendations on which datasets and how much more to train to improve your model. The AI takes over key decisions that would have traditionally required a skilled AI engineer to make.

For our customers, this makes AI accessible to literally anyone. In the truest sense of the word.