Application of CNN networks for an automatic determination of critical loads in scratch tests on a-C:H:W coatings

To enable an automatic evaluation of the scratch test according to DIN EN ISO 20502:2016-11, a new approach using convolutional neural networks (CNN) is developed in this study. For this purpose, a dataset of 2860 image sections is created using high resolution microscopy images of macro- and micro-scratches in a-C:H(:W) coatings, deposited by physical vapour deposition (PVD). Through transfer learning, four different pre-trained CNN networks are qualified to automatically detect the damage patterns of the critical loads Lc1 to Lc3. The classification accuracy is checked with a dataset of separated test images. Verification is done by the evaluation of additional scratch tracks, for which the values of Lc1 to Lc3 in Newton (N) are automatically determined by addition of a suitable image processing. The CNN network models AlexNet, Inception-v3, ResNet-101 and VGG-19 are being investigated for their applicability. All models show an exact detection range of the damage of Lc2 and Lc3, but only the network VGG-19 also classifies Lc1 in good agreement to the human evaluation. The automatic evaluation is robust against image or coating defects. In addition, the method exhibits a high potential for a future implementation on other coating types and surface conditions.