Learning-Based Methods for Visual Understanding in Engineering and Production Workflows

Artificial intelligence (AI) has advanced rapidly across engineering and industrial domains, yet its adoption in production environments is often constrained by the need for higher system reliability, limited availability of high-quality data, and the challenge of embedding tacit engineering knowledge into learning-based models. These limitations hinder the broader industrial push toward flexible, data-driven automation capable of handling increasing product variability and shorter production cycles. This thesis investigates how learning-based methods can be designed and integrated as re-liable perception components within production workflows. Through four complementary case studies, the work demonstrates how visual representations—ranging from Computer-Aided Design (CAD)-derived images and engineering drawings to point clouds and RGB-D images—can be leveraged to address concrete industrial challenges across multiple stages of the manufacturing pipeline. The first case study predicts fixturing clamp configurations for welding operations in automotive manufacturing by learning geometric patterns from CAD-derived representations. The second applies optical character recognition to engineering drawings to accelerate quality-control and documentation workflows. The third examines scene reconstruction and 3D object detection from point clouds, using synthetic data generation to mitigate data scarcity. The fourth develops a fast, zero-shot pose estimation approach for robotic manipulation, enabling reliable object localization in dynamic industrial environments. Taken together, these studies show how AI methods informed by structured engineering knowledge can increase process efficiency, reduce manual workload, and help resolve persistent automation bottlenecks in modern manufacturing.