Digital sensing of production areas and subsequent automated analysis of the captured data can significantly improve the efficiency of factory planning processes. The segmentation of class-specific regions in the captured data using deep neural networks shows great potential for such analysis. However, previous approaches are based on supervised learning and, therefore, require comprehensive, annotated datasets that are costly to generate. The use of large foundation models such as DINOv2 or SAM, combined with few-shot learning approaches, could reduce these efforts in the future. In this work, we first present a method that implements such a combination and subsequently evaluate its performance on an exemplary dataset. The obtained results confirm the method’s potential, especially in scenarios with limited availability of labeled data.

Remanufacturing describes a process where worn products are restored to like-new condition and it offers vast ecological and economic potentials. A key step is the quality inspection of disassembled components, which is mostly done manually due to the high variety of parts and defect patterns. Deep neural networks show great potential to automate such visual inspection tasks but struggle to generalize to new product variants, components, or defect patterns. To tackle this challenge, we propose a novel image dataset depicting typical gearbox components in good and defective condition from two automotive transmissions. Depending on the train-test split of the data, different distribution shifts are generated to benchmark the generalization ability of a classification model. We evaluate different models using the dataset and propose a contrastive regularization loss to enhance model robustness. The results obtained demonstrate the ability of the loss to improve generalisation to unseen types of components.

Vacuum grippers are widely used for handling tasks in manufacturing, due to their simple design and robustness. Unfortunately, their dimensioning often results in oversizing due to limited knowledge about the suction cups, stemming from insufficient insight into how they perform in practice. Recent studies have shown the potential for energy savings through precise dimensioning, which allows for more sustainable production. However, research in this field has been dominated by investigating the behavior of single suction cups, although industrial applications typically involve gripper systems with multiple suction cups. When multiple cups are considered, optimization algorithms can be applied to improve the positioning of individual suction cups to ensure more stable grasps. Here, heuristics decide which parameters must be varied to achieve an optimized suction cup distribution. Subsequently, there is a lack of systematic investigations on the influence of different positions of the suction cups. In addition, the influence of the handling task, including the robot’s acceleration, has received limited attention. To address this gap, we conducted experiments to evaluate the influence of key design parameters. These include the offset between the workpiece’s center of gravity and the point of force application, as well as varying suction cup sizes and their geometric arrangement.

In order to quickly adapt the factory layout to changed product variants or quantities, fast re-planning cycles are crucial for manufacturing companies. A promising approach to speed-up such processes is the acquisition of 3D scans of the factory's shopfloor. These can be used to correctly assess its current state and generate an up-to-date database for layout planning. However, the manual analysis of these 3D scans still constitutes a time-consuming task and the terrestrial LiDAR sensors commonly used for data acquisition are associated with high investment costs. We therefore present an approach for automated analysis of factory layouts based on data captured by a low-cost RGB-D sensor. Semantic segmentation is performed using the acquired color and depth images to classify the different visible areas automatically. On the one hand, the potential of multi- and uni-modal deep learning models is assessed. On the other hand, the use of model confidence calibration approaches is evaluated to improve the reliability of the predicted segmentation masks, avoid false predictions, and hence increase users' trust in the results.

Deep learning based classification models, such as convolutional neural networks (CNN), show high potential for the automation of visual quality monitoring tasks. However, the performance of these models depends on the availability of a suitable database for their training. In the manufacturing domain the generation of such data comes with significant expenses. Especially in high-mix low-volume production, where processes and occurring defect patterns frequently change, the time required for data generation currently hinders the application of deep learning methods. Therefore, we propose a concept for continuous adaptation of deep learning models during their application. By combining an automatic detection of predictions with a high failure probability and continual learning approaches, overall efforts for data generation can be reduced. Initial investigation of the concept for image-based quality monitoring in friction stir welding (FSW) shows promising results.