Automated detection of micrometer-cracks and delamination in CT volumes of previously stressed CFRP pressure rods

Abstract

Industrial tasks that produce hundreds of terabytes of data per year require efficient evaluation tools for the purpose of saving valuable resources such as time or highly trained personnel. Therefore, processes with high automation potential need to be identified and put into practice. Before deploying a new tool, however, thorough investigation is required. It needs to be tested whether the achievable degree of automation of such a new tool yields results which are on one hand reliable and on the other hand of sufficient quality. This work evaluates the applicability of artificial neural networks (ANNs) to detecting micrometer-cracks and delaminations in reconstructed computed tomography (CT) volumes of previously stressed carbon fibre reinforced polymer (CFRP) pressure rods. Common network architectures, varying network parameters and different training sets have been investigated and compared in order to determine the combination that performs best. The constellation (network, hyperparameters, training data) that performed best reached an average precision (AP) of 0.87. Based on the rather small data set of approx. 1.6E3 images and the unstructured nature and great diversity of the investigated features, this result can be regarded as very good. The detected feature sizes ranged from approx. 100 micrometer to a centimeter in length and from tens of microns to a few hundred microns in width. The results suggest that artificial neural networks have the potential to be used reliably for the automatic detection of micrometer-cracks and delaminations in CT volumes of CFRP pressure rods, provided the data set used for training is large and diverse enough and the network is being updated when new data is available.