An attempt to detect anomalies in car body parts using machine learning algorithms

Abstract Industries, which produce hundreds of terabyte of CT data per year, demand automated evaluation approaches. This work provides a first glance of an attempt to automatically detect and characterize possible defects and/or anomalies which formed during common joining processes. We investigated a standard riveting process with respect to the resulting final head height of steel selfpiercing half-hollow rivets. The methods include conventional image processing algorithms, like edge-detection, thresholding and principle component analysis (PCA) which were used to pre-process the CT data. In order to automatically evaluate the reconstructed volumes, which contained several of the aforementioned rivets, we compared the performance of different, publicly available, convolutional neural network (CNN) architectures. Furthermore, we investigated the impact of data augmentation and showed by means of a k-fold cross-validation that the training data causes no overfitting of the network. The obtained results suggest that an automated evaluation of the generated computed tomography scans, with regard to a rivet’s final head height, is feasible. However, in order to increase the network’s reliability and accuracy, the amount of training data needs to be further enlarged and diversified.