FSP3: Produktion
Refine
Year of publication
Document Type
- conference proceeding (article) (106)
- Article (91)
- conference proceeding (presentation) (38)
- Part of a Book (6)
- conference proceeding (volume) (5)
- Preprint (5)
- Report (4)
- Doctoral Thesis (3)
- conference proceeding (summary, abstract) (2)
- Master's Thesis (2)
Language
- English (152)
- German (113)
- Multiple languages (2)
Publication reviewed
- peer review (112)
- begutachtet (101)
- nicht begutachtet (54)
Keywords
- Produktion (10)
- Digitalisierung (8)
- KLEVERTEC / Vorlaufforschung (7)
- LILIPUT (7)
- Automatisierungstechnik (6)
- MEDAL (6)
- Produktionstechnik (6)
- ARCHIVE (5)
- Digitaler Zwilling (5)
- DrGaN (5)
Institute
- Fakultät Maschinenbau (131)
- Fakultät Elektrotechnik (65)
- EPT – Institut für Effiziente Produktionstechnik (43)
- Fakultät Informatik (42)
- IPI – Institut für Produktion und Informatik (38)
- IMS - Institut für Maschinelles Sehen (30)
- KLEVERTEC (21)
- IDF – Institut für Datenoptimierte Fertigung (20)
- IKR – Institut für angewandte Künstliche Intelligenz und Robotik (11)
- IEAT - Institut für Energie- und Antriebstechnik (6)
Synthetic Data Generation for AI-Based Quality Inspection of Laser Welds in Lithium-Ion Batteries
(2025)
Manufacturing companies are increasingly confronted with critical challenges such as a shortage of skilled labor, rising production costs, and ever-stricter quality requirements. These challenges become particularly acute when defect types exhibit high visual variance, making consistent and accurate inspection difficult. Traditionally, visual inspection of high variance errors is performed manually by human operators—a process that is both costly and prone to errors. Consequently, there is a growing interest in replacing human inspection with AI-based visual quality control systems. However, the adoption of such systems is often hindered by limited access to training data, labor-intensive labeling processes, or the absence of real production data during early development stages. To address these challenges, this paper presents a methodology for training AI models using synthetically generated image data. The synthetic images are created using Physically Based Rendering, which enables precise control over rendering parameters and facilitates automated labeling. This approach allows for a systematic analysis of parameter importance and bypasses the need for large real training datasets. As a case study, the focus is on the inspection of laser welds in battery connectors for fully electric vehicles—a particularly demanding application due to the criticality of each weld. The results demonstrates the effectiveness of synthetic data in training robust AI models, thereby providing a scalable and efficient alternative to traditional data acquisition and labeling methods. The trained binary classifier reaches a precision of 0.94 with a recall of 0.98 solely trained on synthetic data and tested on real image data.
In aging societies across western industrialized nations, the loss of expertise due to retiring skilled workers presents a critical challenge for industry. That is especially true on the shop floor, where much of the knowledge is tacitly gained through years of hands-on experience rather than formal documentation. This study explores current knowledge management (KM) challenges and systematically identifies high-potential applications for large language models (LLMs) as part of a broader research initiative aiming to develop human-centered KM solutions supported by generative artificial intelligence (GenAI). We conducted two structured workshops with 23 participants from 14 German manufacturing companies. Three core barriers and two prioritized LLM use cases were identified, contributing specific design recommendations for LLM-supported KM systems for companies. The results advance the understanding of GenAI-assisted knowledge retention in industrial settings and provide a practical foundation for addressing the demographic shift through intelligent, technology-driven solutions.
Der demografische Wandel verstärkt den Arbeits- bzw. Fachkräftemangel in der produzierenden Industrie, wodurch das Thema Wissensmanagement für viele Unternehmen an Bedeutung gewinnt. Insbesondere die Erhebung und Konservierung impliziten Wissens stellt eine Herausforderung dar. Die vorliegende Studie beleuchtet, inwieweit Large Language Models (LLMs) im Bereich der Wissenserhebung mit Experteninterviews sinnstiftend unterstützen können. Drei Experten testen und bewerten den entwickelten Chatbot. Die Ergebnisse des Interviews sind vielversprechend, die Zusammenfassung hingegen zeigt Verbesserungspotenziale
This paper presents a novel integration concept for mid-voltage power semiconductor modules. The concept is based on a dielectric liquid that is used for both, cooling and insulation. The semiconductor chips are attached to Cu inlays in a PCB substrate. Dual side cooling is achieved jet by impingement on the bottom side and immersion in the dielectric liquid on the top side. A first laboratory demonstrator was designed and build-up with 2.0 kV rated SiC MOSFETs in a 140×100 mm² module package. Shell immersion cooling fluid S5 LV was applied as coolant. Test results indicate a junction to fluid thermal resistance of 0.54 K/W at a flow rate of 5 l/min and with pressure drop of 12 kPa. The results obtained are on the same level compared with a state- -the-art automotive power of modules with a pin-fin array integrated in the baseplate and water-glycol as coolant.
In event-based control systems, actuator operations are typically triggered by physical sensor signals to ensure process safety. However, this can lead to systematic delays, especially in pneumatic systems, due to system-related reaction times. This paper presents a method to compensate such delays by statistically advancing control signals without the need for additional sensors.
A bidirectionally coupled digital twin, based on virtual commissioning models, continuously evaluates live data and derives optimized advance signals during operation. The method integrates event-based control with time-based triggering and employs a hybrid approach combining empirical quantiles with analytical confidence intervals to maintain safe operation. Validated in a testbed under realistic industrial conditions, the approach achieves an average time saving of 315 ms per compensated transition. The system architecture decouples real-time PLC execution from non-real-time analysis, enabling scalable integration with diverse control and simulation platforms. The results demonstrate the potential of digital twins for active, data driven optimization in discrete manufacturing systems.
In der Fertigung werden neben CAD-Dateien technische Zeichnungen verwendet, aus denen etwa Maßhaltigkeiten manuell ausgelesen werden. In einem Forschungsprojekt sollen daher durch eine KI-gestützte Zusammenführung der 2D- und 3D-Informationen Kostenvorteile durch durchgängige Automatisierungslösungen erzielt werden.
Fehlende Interoperabilität proprietärer Digitaler-Zwilling-Lösungen hemmt die Skalierung in der Industrie. Um dieses Problem zu lösen, analysiert diese Studie erstmals Deutschlands Asset Administration Shell (AAS), auch bekannt als Verwaltungsschale, durch Befragung von über 50 Manufacturing-X und KoPa35c Projekten, an denen mehr als 200 Unternehmen beteiligt sind. Der internationale Vergleich mit den USA, China und Japan zeigt dabei: Nur Deutschland führt sowohl bei Frameworks als auch Umsetzung. Dies resultiert in der Etablierung der AAS als IEC-Standard mit praktischer Implementierung in Forschungsprojekten.
Embedding represents a game-changing packaging strategy for wide-bandgap semiconductors, slashing parasitic impedances to enable faster, cleaner switching, lower losses, and higher frequencies. Yet questions about reliability, scalability, and heat management persist. Here, we use multiphysics finite-element simulations to explore an embedded half-bridge of two 1.2 kV SiC MOSFETs across a range of commercial insulated substrates - alumina, Si 3 N4, AlN, and IMS with varying layer thicknesses. A Pareto-based thermomechanical optimization pinpoints aluminum nitride as the best configuration, delivering 0.27K/W thermal resistance, 0.2 % plastic strain, and 1.9 % creep strain during sintering. Creep concentrates in the silver sinter layer, matching experimental observations, underscoring the need to address time-dependent deformation in reliability assessments. A major improvement is achieved by redesigning the top copper interconnect from a solid block to a pillar like geometry, which reduces creep strain in the sintered layer by four times. We also identify a critical sintering cool-down rate above which creep vanishes and only plastic strain remains providing a new lever for process control. Finally, we demonstrate scalability by paralleling four optimized prepackages into a power module with just 3 nH of stray inductance, ready for high-frequency, high-efficiency conversion.
Following the recent breakthrough of generative artificial intelligence, humanoid robots have shown significant developments within recent years. Unlike traditional robotics and automation solutions, humanoid robots promise increased flexibility. This development has positioned them as a possible solution to fill the automation gap between large companies and many small and medium-sized enterprises (SMEs), as caused by small batch sizes. This study aims to evaluate the application potential of humanoid robots for SMEs via a qualitative research approach. Within this study, six experts were interviewed based on six categories and hypotheses derived from an extensive literature review. Their responses were systematically analyzed via a qualitative content analysis. The results of this study have shown a large application potential for simple logistics and manufacturing tasks, as well as a significant need for complex construction tasks. Most SMEs have reported an increased need for mobile and flexible automation systems, perceiving humanoid robots as a potential solution within the next five years. This study presents industry sector specific automation needs and underlines the requirement of mobile, versatile, and safe systems with a low-code approach, to decrease entry barriers for SMEs and provide value to their specific needs.
We present an efficient algorithm for monitoring milling tool wear using data from a spindle integrated Cutting Force Sensor. In a near-real-world milling experiment, we collected quality data of the manufactured part, tool wear, and cutting forces. A preprocessing pipeline transforms the data from time-do-main to angle-domain and ensures its integrity. Using Singular Value Decomposition (SVD) for dimensionality reduction, we achieve a compact encoding of key sensor data and avoid manual intervention in the feature engineering. Notably, the SVD reconstruction error proves to be a reliable indicator for tool wear.