Anomaly detection for machine vision

Problem

Manufacturing companies, especially in the automotive supply industry, must meet high quality requirements. Above all, tolerance limits on produced components are very tight.

Due to instabilities in industrial processes, scrap parts are repeatedly produced. For this reason, each component must be visually inspected. Of course, this is very time and resource intensive

Solution

This product uses images from industrial cameras and detects defects on components. It processes images in real-time and classifies between good parts and bad parts. Furthermore, the system can also learn different types of defects to classify even more accurate between types.

USP

Three things that distinguishes this product from existing solutions:

Easy scalable - The system does not need to be configured with tolerance limits of parts. Instead it uses a couple labeled images (good parts / bad parts) and learns the tolerances itself. This way the solution can be adopted to new parts or different production lines very fast.
Few labeled images - The solution uses a semi-supervised learning approach, meaning that in the core it uses an anomaly detection approach to identify abnormalities. On top it uses some human labeled images to validate the anomalies.
Visualizing defects - Classifying in real-time is the core functionality but the product is also able to visualize the defect on the image. So process engineers can faster inspect the images to stabilize their process.

The solution supports Azure IoT Services and is currently only available in German.

Anomaly detection for machine vision

Autor: craftworks