Towards high-performance deep learning models in tool wear classification with generative adversarial networks

Abstract

Image-based deep learning (DL) applications have been increasingly applied in manufacturing processes in recent years and enable real-time decisions to be derived on the basis of raw image data. However, there are major obstacles that prevent a widespread application. The generation of extensively labeled datasets is associated with costly experiments, long downtimes and demanding measurement setups. Thus, users are usually faced with a trade-off problem, where they have to weigh between the profitability and the performance of their DL algorithms. For this reason, this paper demonstrates how this obstacle can be tackled through the application of data augmentation (DA) techniques. The investigations are based on an image dataset from a tool wear classification task in a blanking process, in which a pre-trained convolutional neural network (CNN) identifies correlations between cutting punch radii and workpiece images. It is shown how synthesized images in combination with transfer learning (TL) models affect classification accuracies using basic image manipulation, different types of generative adversarial networks (GAN) and their hybrid application, varying the number of available images in the training dataset between 160 and 4800. These augmentation techniques lead to accuracy improvements of up to 18%, although their effectiveness depends heavily on the amount of image data available.