An automated method for the classification of the adhesion strength of thin PVD coatings applied on hardened steel substrates is presented in this study using deep neural networks. For the determination of the adhesion strength Rockwell-indentation tests were carried out according to VDI 3198. For this approach, pre-trained convolutional neural networks are adapted to classify microscopic images into the expected adhesion classes HF 1 to HF 6 using transfer learning with a dataset of 1650 already evaluated indentation images. The classification performance of the Matlab implemented network models AlexNet, GoogLeNet and inception-v3 is compared with test and verification images of Rockwell indentations. The inception-v3 network shows good accuracy for polished (roughness Sa < 20 nm), hardened steel substrates with deposited thin coatings of a thickness up to 5 μm. The classifications of the implemented models exhibit an agreement of approximately 85–90% compared to human assessment. The evaluation is robust against disturbance variables such as different exposure times, brightness, image contrasting and magnifications. Different image capture devices can be used with no effect on the classification. The networks show promising results for automated industrial applications, such as in-line adhesion control in coating processes, as they do not require human operator support.

这项研究使用深层神经网络提出了一种自动方法，用于对涂在硬化钢基材上的薄PVD涂层的粘附强度进行分类。为了确定粘合强度，根据VDI 3198进行了Rockwell压痕测试。对于这种方法，采用预训练的卷积神经网络，使用转移学习方法将显微图像分类为预期的粘合等级HF 1至HF 6。 1650个已评估压痕图像的数据集。Matlab实现的网络模型AlexNet，GoogLeNet和Inception-v3的分类性能与Rockwell压痕的测试和验证图像进行了比较。Inception-v3网络显示出良好的抛光精度（粗糙度Sa <20 nm），硬化的钢基材，其沉积的薄涂层厚度可达5μm。与人类评估相比，已实施模型的分类显示出约85–90％的一致性。该评估对于干扰变量（如不同的曝光时间，亮度，图像对比度和放大倍数）具有鲁棒性。可以使用不同的图像捕获设备，而不会影响分类。该网络在自动化工业应用中显示出令人鼓舞的结果，例如在涂布过程中进行在线粘合控制，因为它们不需要人工支持。图像对比和放大。可以使用不同的图像捕获设备，而不会影响分类。该网络在自动化工业应用中显示出令人鼓舞的结果，例如在涂布过程中进行在线粘合控制，因为它们不需要人工支持。图像对比和放大。可以使用不同的图像捕获设备，而不会影响分类。该网络在自动化工业应用中显示出令人鼓舞的结果，例如在涂布过程中进行在线粘合控制，因为它们不需要人工支持。