Overcoming Data Scarcity in the Quality Control of Safety-Critical Fibre-Reinforced Composites by means of Transfer and Curriculum Learning

Abstract

Fibre-reinforced composites are one promising material class to provide a response to the increasing environmental awareness within society. Due to their excellent lightweight potential, fibre-reinforced composites are preferably employed in safety-critical applications, requiring extensive quality control (QC). However, commercially available QC systems are only able to measure fibre deviations, not directly detecting the error itself. In consequence, a worker is required to perform a manual inspection. Artificial intelligence and especially convolutional neural networks (CNN) offer the opportunity to directly detect and classify defects. However, to train the corresponding algorithms large amounts of data are required, which are often inaccessible in production. Artificial augmentation of the available data is a popular approach to tackle this problem, yet, resulting most of the time in undesired overfitting of the CNN. Therefore, in this contribution we examine the transfer of human learning behaviour elements to algorithms in form of transfer learning (TL) and curriculum learning (CL). The overall aim is to research, whether CL and TL are appropriate approaches to address data scarcity in e.g. production environments. Therefore, we perform our research on the error detection of three-dimensional shaped fibre-reinforced textiles.