Machine Learning for LWL - Sachsenkabel

Our daily project work already shows what a strong impact technologies such as artificial intelligence and machine learning will have on product development in the future. With our expertise in AI, we are constantly creating completely new approaches to thinking and solutions in everyday development.

The example of our Machine Learning Project for LWL-Sachsenkabel GmbH shows how these new approaches can look in concrete terms. Quality control plays a decisive role in the assembly of high-performance fibre optic cables. Even the slightest scratch or speck of dirt interferes with the optical signal path and impairs the transmission performance of the cables. 

Together with LWL-Sachsenkabel we have rethought the subject of quality assurance and created alternative ways to support the testing process by developing our own neural networks.

Intensive network training and data analysis

On the basis of microscope images and the corresponding information, we have developed two convolutionary neural networks based on division of labour within a very short period of time. Network 1 detects whether the product meets the quality criteria and network 2 determines the product type.

The networks were developed with Tensorflow and Keras. Our work included not only the training of the networks, but also an intensive analysis and processing of the available data sets. During the process both architecture and hyper parameters of the models had to be adapted to the properties of the data. The enormous level of detail and the high resolution of the images were a particular challenge, since these factors increase the input vector of the mesh and thus the required computing power.

The final deep learning networks developed by UXMA consist of several successive layers of convolutional and dense layers. The correct setting of the size of the kernel and stride played a major role in the hyperparameter tuning. We were able to successfully reduce the overfitting by using weight normalization and dropout layer.

Netztraining before

Left picture: The training course of a network designed to recognize small details, but focused on the recognition of larger image structures. It shows that the network has a poor generalization.

Netztraining after

Right picture: After adjusting filter size and stride, the network also achieves very good results for test data.

In addition to our own development work, we have evaluated the transfer learning approach which is often used in the computer vision environment. Nevertheless, we have achieved better results on this data set with models trained from scratch. This shows how important it is to adapt machine learning models to the individual circumstances of the data situation in order to achieve useful results.

Together with the integration of our neural networks into an API and the development of a corresponding web front end, the implemented AI solution offers our customer LWL-Sachsenkabel significant support in the quality assurance process.


  • Consulting in the field of artificial intelligence and machine learning
  • Analysis and preparation of data
  • Design and implementation of tailor-made deep learning solutions
  • Integration of the AI into a client-server architecture
  • Consulting for further development in the field of computer vision

Need support for AI-based software development? Contact our machine learning experts Jule Martensen and Frank Lindecke via software (at) for a non-binding exchange of information. A list of all our areas of expertise at UXMA can be found in the Services Overview


Presentation of the project results via Tech-Talk for the interdisciplinary Sachsenkabel and UXMA development team

Jule Martensen
Jule Martensen is a software developer at UXMA. Among other things, she deals with the representation of content based on object recognition by neural networks.
Say Hello to the New Home Connect­!

Say Hello to the New Home Connect­!