Artificial Intelligence for Quality Assurance

© Hörmle GmbH

Controlling the quality of laser-cut edges


What's it good for?

In order to automatically assess the quality of laser-cut edges on sheet metal with regard to roughness and burrs in images, researchers at Fraunhofer IPA have implemented a "Convolutional Neuronal Network" (CNN) and deep learning. They have thus succeeded in raising quality assurance to a new level.

What's new?
Correlation rates of 75 percent and 85 percent respectively were achieved for the assessment of roughness and burrs. The quality of the cutting edge can then be improved in a next step by optimizing machining parameters.

Self-learning surface inspection

 

What's it good for?
Fraunhofer IPA has developed an adaptive, optical inspection method for reliably detecting surface defects and contamination, as well as fluctuations in product samples in serial production processes.

What's new?
Based on an unsupervised learning process, the optical inspection system automatically adapts to changes in surface structures. This enables reliable 100 percent in-line inspection that is successfully implemented, for example, to check for impact marks on sealing surfaces.

Classifying objects to recycle catalytic converters

 

What's it good for?
Until now, machines have found it difficult to correctly identify old equipment such as vehicle catalytic converters because they are often in poor condition due to corrosion, deformation or surface damage. Self-learning object classification attains significantly better results.

What's new?
The discarded devices are first measured three-dimensionally with a laser line sensor and then classified according to various features (e.g. object contour, bounding box, curvature, control geometry) in three steps based on neuronal networks. The researchers achieved a classification rate of over 90 percent.

Color segmentation

 

What's it good for?
Checking the assignment of multi-pin connector chambers is a challenging task because many different colored cables, which may be twisted, are used here.

What's new?
Researchers at Fraunhofer IPA have successfully trained and implemented a neuronal feed-forward network for color segmentation in camera images. This enables cables to be detected and separated optically and automatically, thus serving as a basis for cable tracking and inspecting chamber assignment.

Automated visual control of biological processes

 

What's it good for?
In biology and pharmacy, large quantities of fertilized zebrafish eggs are required. Knowing the number of cells contained in the eggs is essential in order to differentiate between fertilized and unfertilized fish eggs. Upwards of four cells, the fish egg is considered fertilized. Up to now, this distinction has been made manually.

What's new?
To automatically identify fertilized fish eggs, a camera image is taken of each egg and analyzed by a deep learning network, a multi-layered "Convolutional Neuronal Network" (CNN). Despite the fish egg lying in a random position during image acquisition, the result is a recognition rate of 99.8 percent.

Further Information

 

Department

Image and signal processing