Neural Network-Based Inspection of Printed Circuit Boards

Abstract

The goal of this master thesis is to research the feasibility of creating a neural network-based automatic optical inspection system using off-the-shelf parts and open-source software. The purpose of the finished system is to be an easily implementable solution for quality control. As a significant part of the production at Hapro electronics is the assembly of printed circuit boards (PCB), the leading case of the project is the classification of components on circuit boards. However, the system created in this thesis can be applied to other product inspections.

As a proof of concept for this thesis, a prototype inspection cell has been built. Several experiments using data from the prototype with different neural network models have been conducted.

The experiments concern the models' depth, the model's base architecture, and the components on which the models were trained. Several models were created for each dataset and evaluated using a test dataset containing data not present during training.