Automatic Particle Classification Through Deep Learning Approaches for Increasing Productivity in the Technical Cleanliness Laboratory

Understanding the properties of particles plays a vital role in assessing the component cleanliness and its origin in the manufacturing process. We propose a classification method using deep convolutional neural networks. Using a dataset of 70,000 annotated images, we achieve a accuracy of 97.7% for a binary classification in metal and non-metal particles comparable to state-of-the-art polarized light microscopy according to VDA 19-1 and ISO 16232. Manual follow-up checks in a cleanliness laboratory are not required due to the robustness of the classification system.