Kategorie: ‘Theses’
SMART-LLM: Sensor-based Maintenance bot for Analysis and Retrieval of Time Series data using LLMs
The goal of this thesis is to design, implement, and evaluate a sensor-based maintenance bot that uses Large Language Models (LLMs) to support predictive maintenance and decision-making. The bot should be capable of retrieving, analyzing, and reasoning over time series sensor data as well as unstructured maintenance-related documentation (e.g., technical manuals, incident reports). The result is a unified system that assists technicians and engineers in diagnosing issues, suggesting preventive actions, and retrieving relevant information in natural language.
This work is conducted in collaboration with the Fraunhofer Institute for Production Technology (IPT) as part of the research initiative “Generative AI for Production and Business Operations“, aiming to explore practical applications of generative models in manufacturing and industrial operations.
FAIRification of Data Models in Manufacturing
In modern manufacturing, data plays a crucial role in optimizing processes, enhancing efficiency, and enabling interoperability across different systems. However, data models in manufacturing are often heterogeneous, proprietary, and lack standardization, making data sharing and integration challenging. The FAIR principles [1] – Findability, Accessibility, Interoperability, and Reusability – provide a structured framework to improve data management and usage across various domains. Applying these principles to manufacturing data models can enhance data exchange, facilitate automation, and support decision-making in Industry 4.0. This thesis aims to investigate methods for FAIRifying manufacturing data models (especially with focus on the Digital Shadow Reference Model from the Cluster of Excellence Internet of Production [2]), addressing key challenges such as semantic alignment, metadata enrichment, and interoperability with existing standards.
Aligning Regulatory Requirements with Industry Standards: Creating Transferable Compliance Guidelines
TEAMS: Tailoring Engagement and Alignment for MLOps Stakeholders
The goal of this thesis is to investigate and define stakeholder engagement and involvement within the lifecycle of sensor-based MLOps. While MLOps principles streamline the deployment and maintenance of machine learning (ML) models, ensuring proper stakeholder involvement remains a challenge. Different stakeholders must collaborate effectively across various lifecycle stages to ensure model reliability, fairness, and operational success.
Enhancing Fake News Detection using Multi-Agent LLM Frameworks
The proliferation of misinformation and fake news on social media platforms poses a significant challenge in today’s digital age. Traditional automated fake news detection systems often struggle with the complexity of the task, lacking the ability to provide detailed explanations and interpret nuanced contextual information.
This thesis explores the enhancement of fake news detection using Multi-Agent LLM frameworks. By emulating human expert behavior, the proposed system employs specialized LLM agents, each focusing on a distinct aspect of the problem.
This research investigates the effectiveness of using Multi-Agent LLM frameworks to enhance fake news detection, in comparison to traditional machine learning models and single-agent LLM systems. It also examines the impact of integrating real-time web knowledge and the optimal design and composition of such frameworks. This innovative approach aims to provide comprehensive, human-like explanations for its assessments, enhancing user trust and interpretability, while maintaining the speed and scalability of automated solutions.
Creating the Digital Nervous System for the Internet of Sustainable Production
Small-Scale Agent-Aided AutoML with pre-Existing Code and Models
AutoML automates machine learning pipelines, making model training accessible without deep expertise. Recent advancements use LLM-based agents to optimize pipeline steps, but existing solutions often require large-scale models with high computational costs. Smaller, open-source models provide a more accessible alternative, especially when combined with domain-specific pre-trained models. However, integrating these models into AutoML via agent frameworks remains unexplored. This thesis aims to exten an agent-based AutoML framework using small LLMs and integrating existing classification models into the pipeline using an ontology
OGRETA: Ontologie-Entwurf für das GRETA-Kompetenzmodell
Das Ziel dieser Bachelorarbeit “OGRETA – Ontologie-gestützter Rahmen für das Entwicklungstracking von Arbeitskompetenzen” ist die Überführung des GRETA-Kompetenzmodells in eine Ontologie, die als Grundlage für ein Empfehlungssystem dienen kann. Das GRETA-Modell definiert die beruflichen Kompetenzen von Lehrkräften in der Erwachsenenbildung unabhängig von Fachgebiet, Erfahrung oder Beschäftigungsart. Durch die Strukturierung dieses Modells als Ontologie soll die Basis für ein intelligentes System entwickelt werden, das individuelle Empfehlungen für Weiterbildungsmaßnahmen im Rahmen des Forschungsprojekts TrainSpot – Train the Trainer HotSpot ermöglichen wird.
An Ontology-Based Agent for Explainable Unstructured IoT Data
Sensor data is often unstructured and while available datasets show some clear use cases, for example calculating energy consumption over time, relationships between measurements can often go unnoticed without a thorough examination of the data. While exploratory data analysis can reveal connections, without a clear analytical direction the results may be limited to general information, such as clustering or embeddings. In many cases, stakeholders or key decision makers may however lack knowledge to go beyond such analysis. Using Graph Retrieval-Augmented Generation (Graph-RAG), LLMs can infer connections between entities within a given knowledge graph, potentially providing more accurate and meaningful outputs. In general, data based on Ontologies can be represented as such graphs and has already been used to enhance LLM Agents with domain-specific knowledge. Therefore, if an LLM agent would be able to infer and explain important characteristics of given data with the help of a data-focused IoT ontology and convey them to a stakeholder, one could directly go on to more expedient data analysis.
Knowledge Graph-Based Chinese Tourism Recommendation System with Large Language Models
With the increasing popularity of personalized tourism experiences, recommendation systems play a crucial role in helping travelers discover destinations, activities, and itineraries that match their preferences. Traditional recommendation models often rely on collaborative filtering or content-based filtering, which may struggle with cold-start issues, lack of contextual awareness, and limited adaptability to dynamic tourism trends.
Knowledge graphs (KGs) provide structured representations of entities and their relationships, offering a powerful way to enhance recommendation systems by integrating domain knowledge, user-generated content, and external tourism data. Additionally, Large Language Models (LLMs) can be used to enrich the KG with semantic understanding, improving recommendations by leveraging context-aware reasoning and natural language interactions.
This thesis explores the development of a knowledge graph-based tourism recommendation system for Chinese travelers by constructing a time-sensitive ontology, integrating tourism data from Xiaohongshu (Red Notes), and leveraging LLMs for enhanced personalization and recommendation generation.