In the context of industrial production, the utilisation of data recording and processing techniques is becoming increasingly prevalent across the manufacturing sector. The solutions integrate sensors, facilitate the transmission of data, and enable data-driven decision-making, thereby reducing downtime and optimising quality. However, challenges emerge due to the limited non-transferable data or models between processes. Alterations to the production process can render collected data invalid, resulting in restricted datasets and potential overfitting. To address these issues, techniques such as data augmentation are employed. This study aims to develop a data augmentation methodology applicable in dynamic, data-scarce production environments, enhancing the robustness of regressor predictions.
The data augmentation method that has been developed is based on a sampling space that has been constructed on the grounds of the inherent uncertainty in the measured data on material properties. It is assumed that augmenting the data within the confines of its own uncertainty will ensure that the labels of the augmented data points can be preserved regarding their original counterparts.
The present study demonstrates, using data on car body part production in press shops, that the application of data augmentation techniques enhances the robustness of the models in question. Furthermore, the methodology developed has demonstrated superior performance compared to a standard augmentation technique, namely jittering.
The existing literature lacks data augmentation methods for regression tasks in manufacturing. This is because the majority of existing approaches were developed for classification purposes, given that data augmentation originated in image and sound processing.
