top of page
  • Writer's picturedeevio

deevios AI improves root cause analysis in production at BASF

How can AI and existing image data be used to draw better conclusions from production and then, in a second step, optimise process parameters and the resulting product quality? The process experts at BASF Polyurethanes GmbH in Lemförde asked themselves this question and approached deevio.

Together, we retrofitted an existing image processing system, created an AI model, integrated it into the BASF machine environment, and developed a feature that allows BASF to create new AI models completely autonomously. And it has now done so on several production lines and across several continents.

The challenge: retrofit of an existing image processing system

Until now, products were analysed with a 3D measuring device in order to draw conclusions about product quality through rule-based measurements and to monitor the production process. However, due to the high variability of the features to be detected and the required accuracy of BASF, the rule-based image processing reached its limits and the assignment had to be done manually. Another challenge was the fast processing of large amounts of data, since BASF produces the product in very high quantities at its site in Lemförde.

In order to optimize the process of rule-based image processing, BASF and deevio started a project to develop a customized AI-based software solution.

Development AI model: A large number of product features to be recognized.

The AI uses sample images of real products from the existing camera system and learns from them independently what constitutes a good component and what constitutes a defective one. In doing so, the AI also learns to deal with the variability of the products, thus guaranteeing high accuracies during recognition while keeping pseudo-rejects low.

To train a successful AI model, the training process requires at least 50 to 100 images per defect category. An AI model can then be developed based on this data. For this purpose, BASF provided deevio with sufficient images per defect category from the existing 3D measuring device and consistently labeled them together with deevio. The labeling process is an important step in this process, because without consistent labeling, the AI model cannot make a correct decision either. Based on this, deevio has developed an AI model that now categorises the products into the correct class with an accuracy of > 98%.

The Confusion Matrix shows whether the AI model classifies the respective image into the correct class. Ideally, all data points on the confusion matrix lie on the diagonal from top left to bottom right, which is the case here in almost all cases. With more images and better labels, BASF can now independently improve the AI model and add more classes.

In addition to quality evaluation, this also makes it possible to draw conclusions about the causes of defects in production, which was previously possible in a limited form.

Advantages of the AI model in production at a glance

  1. Instead of using metrics such as component dimensions, quantified functional defect classes are created

  2. Enables correlation between process and quality based on external characteristics

  3. Automated information about the current production process

The collaboration - remote installation during the Corona pandemic

After the successful development of the AI model, the next step in the collaboration was approached. The next challenge was to integrate it into the production at BASF. In production, a high inspection frequency of the components is required, since an up to 5-digit number of products has to be evaluated in a very short time. Therefore, deevio has further developed the AI box so that this requirement can also be mapped. The images are now classified with a speed of up to 0.004s per image.

Due to the Covid pandemic, commissioning was carried out completely remotely, which went smoothly thanks to good cooperation with the experts from Lemförde. Deevio was also able to handle the integration into BASF's complex automation structure through an adaptive customer solution. The remote commissioning was also a good preparation to support the global BASF teams in the implementation and operation of the new AI model.


Working with deevio, BASF Polyurethanes was able to improve production in several categories.

  • Automated classification in high accuracy creates detailed basis for inference on production and root cause analysis

  • BASF can develop new AI models as well as new inspection criteria and products with the self-training feature developed in cooperation and programmed by deevio

  • Successful integration into BASF machine environment with PLC interface.

  • Already over 2 years of operation with AI software

  • Roll-out to different plants worldwide

If you are also looking to improve your product quality or gain more knowledge from the production process in your company, we would be happy to assist you. We can create new vision systems or, as in this case, retrofit existing systems.

81 views0 comments


bottom of page