Enabling automation with deep learning at optical X-ray inspection station succeeding various soldering processes in electronic control units manufacturing

Abstract

This master thesis is thought to describe a general framework, environment and a step by step process, to facilitate implementation of deep learning in electronic control units manufacturing industry focusing on optical failure detection at automated optical inspections processes after soldering processes. The approach comes here from the desire to encourage and support any electronic manufacturing company to benefit from applying latest technologies in artificial intelligence, deep learning especially, because they are more than ever accessible, open source, could be based on cloud computing where also the needed environments is configured and ready to use at very attractive cost. Furthermore, this thesis will demonstrate the benefit of how automated decision making can be enabled by trained neuronal networks and prove that the knowledge bar is now inclined towards process experts users more then on practitioners having programming skills and mathematics and statistics knowledge. Within this master thesis, I will dedicate a special focus on image recognition making use of convolutional neural network. As a starting point I turned my studies towards Google and Facebook open source machine learning libraries, TensorFlow and PyTorch, developed by artificial intelligence research group of corresponding companies. This are offering incredibly powerful models for computer vision applications. Indeed, it was observed that convolutional neural network are the most suitable for analysing visual images. The case study will try to demonstrate that the ResNet50 of PyTorch trained convolutional neural network is able, with an accuracy of 0.93, to classify the optical images acquired by the X-ray system, enabling sorting them according to defined classes and in the end enabling automated decision according to pass fail criteria. Keywords: deep learning in manufacturing electronics, neural networks, convolutional neural networks