english  | Home | Impressum | Datenschutz | KIT

Deep-Learning Convoluted Neural Networks for Automated Optical Inspection of Power Electronic Systems

Deep-Learning Convoluted Neural Networks for Automated Optical Inspection of Power Electronic Systems
Forschungsthema:Deep-Learning Convoluted Neural Networks for Automated Optical Inspection of Power Electronic Systems
Typ:Bachelor-/Masterthesis
Datum:now
Betreuer:

Dr. Thomas Blank

Dipl.-Inform. Steffen Bähr

The thesis is mainly supvervised by IPE. A workplace at campus south can be arranged on request.

Job description
Power electronic systems using SiC-MOSFETS can be operated at much higher temperatures than systems comprising standard silicon IGBTs. This might cause high thermal and mechanical stress in the power module. Hence, new interconnect technology based on low temperature silver and copper sintering processes as well as innovative double-side cooled power modules are developed at the Institute for Data Processing and Electronics (IPE). However, during operation or already in production steps the modules might break. The assessment of the root failure is done through ultra-sound image inspection. During the past 5-10 years Deep Learning Convoluted Neural Networks (CNN) were developed, tailored for automated pattern recognition and feature extraction for medical and other purposes. The quality of the automated feature extraction is very high. Hence, CNN methods are likely to be successfully utilized in automated quality inspection for electronic systems as well.

Task
In the thesis, CNN methods implemented in MATLAB® shall be developed and evaluated for automated quality detection of power electronic substrates. The input data („pictures“) of the substrates is generated by an advanced ultra-sound Scanning Acoustic Microscope (SAM). Our industrial collaboration partners provide a picture sample library, which shall be augmented by new Images and, which can be acquired with the SAM of the institute. Specific power modules with well-defined defects shall be prepared and used to demonstrate to robustness of the parametrized CNN. 

You should:

Summarize the state of the art of CNN-algorithms

Propose a network suitable for the pattern recognition tasks

Implement the network under MATLAB

Train and validate the network

Develop methods to run the CNN using the GPU capabilities of MATLAB

Propose methods to run the algorithms natively on GPUs


Personal qualification 
You study electronic or mechanical engineering or computer science with a strong engineering and computational background. You have good programming skills in MATLAB and some background in neural networks. Additionally, you have strong analytical skills and the motivation to solve challenging technical tasks.


Contact
 Mr. Thomas Blank, E-Mail: Thomas Blank∂kit edu; phone: +49 (721) 608-25618