Miba AG Creates Deep Learning System for Enhanced Quality Inspection

The System Employs a Data-Centric Approach to Continuously Retrain Models

“After switching from the rule-based approach to this more model-based, data-centric approach—with the backbone of MATLAB—we were able to quickly get the systems running. With this approach, it’s easier to reduce the false positive rate.”

Key Outcomes

  • Ten unique deep learning applications deployed across 40 inspection stations with over 25 trained networks
  • Development time reduced by 30–50%, with new models trained and deployed in under 1 hour
  • Increased productivity by reducing false positives and shortening downtime with deep learning models

Miba AG is an Austrian manufacturer of parts for the energy sector. To scale its production while maintaining the highest inspection and quality standards, Miba’s engineers developed a deep learning–based visual quality inspection system using MATLAB®.

The team used Image Acquisition Toolbox™ to capture images of manufactured parts on the production line. They preprocessed the images with Computer Vision Toolbox™ and Image Processing Toolbox™, isolating regions of interest, detecting poses, and extracting critical features. Using Deep Learning Toolbox™ and Statistics and Machine Learning Toolbox™, they developed deep learning models that detect defective parts based on the preprocessed images.

Based on the model’s predictions, the system sends commands to programmable logic controllers using Industrial Communication Toolbox™, ensuring immediate action within the production line. Simultaneously, the inspection system sends images and log data to databases and displays relevant information on dashboards to the production line operators.

Using MATLAB to train models that identify missing or wrongly labeled data, Miba’s engineers also created a data-centric development pipeline to continuously upgrade their models. As engineers gather new images, they annotate them and use Deep Learning Toolbox and Statistics and Machine Learning Toolbox to retrain, evaluate, test, and deploy the models.

The entire system is an executable file that runs on average-priced industrial PCs installed at the production plant without the need to send data to the cloud.