A precise and repeatable classification of a fracture sample is necessary for a qualitative statement about the suitability of adhesives. For this purpose, we have developed a system that can record samples both as a color image and as a height map and determine the proportions of the fracture surface types using machine learning. The classification of fracture surfaces of adhesives has so far been a time-consuming, manual process characterized by subjectivity and experience. Initial publications show the application of deep learning for the classification of metal fractures. Due to the variety of substrates and adhesives, deep learning is not suitable for analyzing adhesive fracture samples due to the amount of data required. Instead, an algorithm was developed that extracts features from the image data and uses a random forest for classification. By combining feature extraction with the random forest, the algorithm can achieve good results even with minimal training data. One image with small areas marked per class is sufficient to generate a good classification. Classification can be further improved with additional images or reinforcement learning. Because the amount of data is small and the training is fast, a separate model can be trained quickly for each substrate-adhesive combination. When comparing the classification results of two experts, deviations in the two-digit percentage range were observed. Deviations are also possible with the algorithm due to the use of random classification of the data during training. By optimizing the parameters, the maximum influence could be reduced to 3% and the average influence to 0.8%. It was also shown that the inclusion of both sample sides and the height image leads to a significant improvement in classification. Together with an industrial partner, a demonstrator was built that is able to record the images in color and with height information and automatically evaluate the images.
Adam et al. (Thu,) studied this question.