Deep learning-based error recognition in manual cable assembly using synthetic training data

• DOI: 10.1016/j.procir.2024.04.005
• Publication Date: 2024-10-15T15:52:13Z
• Authors: Jonas Conrad; Tobias Stauffer; Xuanting Meng; Julian Ferchow; Mirko Meboldt
• Keywords: Manual assembly; Machine learning; Synthetic training data; Computer vision; Machine learning; Deep learning; Synthetic training data; Manual assembly; Quality control; Quality control; Computer vision; Deep learning

ISSN:2212-8271<br/>34th CIRP Design Conference<br/>Procedia CIRP, 128<br/>Assembly states can be automatically recognized using deep learning models for computer vision. However, the requisite extensive effort and specialized know-how for training data preparation pose significant hurdles for industrial application. To overcome this issue, methods to generate training data synthetically have been implemented. Yet, they do not close the gap from state- to error-recognition. Timely recognition of not only states but errors is crucial to mitigate high error costs in manual assembly processes. We propose the usage of an adaptable object detection-based error recognition that is evaluated in a real-world industrial scenario, where multiple components must be assembled onto a cable in a specified sequence. The deployed object detection model is purely based on synthetic images rendered from CAD files. In contrast to existing literature, the proposed solution allows to define various error states. In evaluation, a mAP0.5 of 99.2 % for object detection could be reached and the correct and various erroneous assembly states could be recognized with 96.25 % accuracy. These results highlight the potential of this solution for automated error recognition, which can ultimately reduce assembly errors and the resulting failure related costs in industrial applications. Future projects will transfer these learnings to more complex manual assembly scenarios.<br/>