Towards Self-adaptive Defect Classification in Industrial Monitoring

Andreas Margraf, Jörg Hähner, Philipp Braml, Steffen Geinitz


The configuration of monitoring applications is usually performed using annotations created by experts. Unlike many industrial products, carbon fiber textiles exhibit low rigity. Hence, surface anomalies vary to a great extend which poses challenges to quality monitoring and decision makers. This paper therefore proposes an unsupervised learning approach for carbon fiber production. The data consists of images continously acquired using a line scan camera. An image processing pipeline, generated by an evolutionary algorithm is applied to segement regions of interest. We then cluster the incoming defect data with stream clustering algorithms in order to identify structures, tendencies and anomalies. We compare well-known heuristics, based on k-means, hierarchical- and density based clustering and configure them to work best under the given circumstances. The clustering results are then compared to expert labels. A best-practice approach is presented to analyse the defects and their origin in the given image data. The experiments show promising results for classification of highly specialised production processes with low defect rates which do not allow reliable, repeatable manual identification of classes. We show that unsupervised learning enables quality managers to gain better insights into measurement data in the context of image classification without prior knowledge. In addition, our approach helps to reduce training effort of image based monitoring systems.


Paper Citation