Clustering of Image Data to Enhance Machine Learning Based Quality Control in THT Manufacturing

In this work, machine learning (ML) models are presented to identify false calls during quality control with automated optical inspection (AOI) in through hole technology (THT) manufacturing. While ML-based approaches with both, image data and numerical data, have already been investigated extensively in SMT manufacturing due to the higher market share, research for THT manufacturing does not have the same extend [1]. The presented models classify images into false calls and true defects, which were identified by the AOI as defects beforehand. Since the AOI uses different test routines to control pin, meniscus and foil of the different components and the board’s surface, multiple models for groups of images are developed to achieve better performance than a single one. To assign the images to the corresponding models, clustering of image data is done in two steps. First, the dataset is divided into subcategories based on the supplementary and descriptive data on the test routine. Second, the unsupervised machine learning algorithm k-means is used on images in each subgroup for further assignment to a dataset. A cumulative examination of the results of different convolutional neural networks (CNN) on the individual clusters leads to a relative improvement in false call detection of 6.8% while error slip can be reduced from 0.6% to 0% in an independent test data set, which is not used for model training.