Die Integration von Large Language Models (LLMs) in die Bildung von Modellkonfigurationen für Co-Simulationen eröffnet neue Perspektiven in der Forschung und Entwicklung. Diese Arbeit untersucht, wie LLMs, gestützt durch Knowledge Graphen (KGs), zur Verbesserung und Automatisierung von Co-Simulationsprozessen eingesetzt werden können. Durch die Modellierung und Nutzung des Wissens in KGs sollen LLMs in der Lage sein, präzise und kontextgerechte Konfigurationen zu generieren, die den Anforderungen komplexer Simulationsumgebungen entsprechen. Der Fokus liegt dabei auf der Entwicklung von Methoden zur effizienten Interaktion zwischen Nutzern, Co-Simulationstools und LLMs, um eine nahtlose und effektive Wissensintegration und Modellkonfiguration zu gewährleisten. Diese Arbeit zielt darauf ab, die Brücke zwischen theoretischem Wissen und praktischen Anwendungen zu schlagen, um die Effizienz und Genauigkeit von Co-Simulationen zu steigern. Es wird erwartet, dass die Ergebnisse dieser Forschung nicht nur die Qualität der Simulationen verbessern, sondern auch den Zeitaufwand und die Komplexität bei der Erstellung und Anpassung von Modellkonfigurationen erheblich reduzieren.
Diese Forschungsarbeit verfolgt das Ziel, einen innovativen Ansatz zur Nutzung von LLMs für die Konfiguration von Co-Simulationen mittels Knowledge Graphen zu entwickeln. Hierzu werden zunächst bestehende Ansätze wie Prompt Engineering, Retrieval-Augmented Generation (RAG) und Tool-Using analysiert und bewertet, um deren Potenzial und Grenzen zu identifizieren. Eine umfassende Literaturrecherche wird durchgeführt, um die neuesten Entwicklungen und bewährten Verfahren in diesen Bereichen zu ermitteln. Aufbauend auf dieser umfassenden Analyse wird ein Konzept entwickelt, das die effiziente Interaktion zwischen Nutzern, Co-Simulationstools und Knowledge Graphen durch LLMs ermöglicht. Abschließend erfolgt eine prototypische Implementierung des entwickelten Ansatzes, um die Praxistauglichkeit und Effektivität zu demonstrieren. Der Prototyp wird anschließend evaluiert, um die Leistungsfähigkeit und Anwendbarkeit in realen Szenarien zu prüfen und gegebenenfalls Optimierungen vorzunehmen.
- Sehr gute konzeptionelle Fähigkeiten
- Grundlegende Kenntnisse im Software Engineering und bei IT-Systemen
- Sehr gute Deutsch- und Englischkenntnisse
Daniel Dittler
♦
Following the increasing popularity of Deep Learning (DL) methods, many experts are investigating the applicability of DL techniques for anomaly detection.
However, in most cases, these techniques are applied in a rather brute-force manner.
They ignore essential and, in many cases, apriori known information about the system structure and behavior.
Especially, they assume to take every possible connecting point into consideration, which is not efficient while modern cyber-physical systems could have thousands of potential assess points to be connected to.
The student will investigate several Simulink models as examples, and research a general solution for getting the best optimized access points given existing models.
The expected deliverable is a program that takes simulation models as input, analyzes the system architecture, and outputs the identification of suitable access points for anomaly detection.
The result can be verified with fault injection experiments.
1. literature research of related mbsa methods, such as dynamic fault trees, bayesian network, reachability analysis, importance metrics
2. Scenario study, get familiar with different case study models and corresponding simulations.
3. Test different approaches to the models
4. Compare the results of different approaches
Matlab/Simulink
Sheng Ding
♦
Following the increasing popularity of Deep Learning (DL) methods, many experts are investigating the applicability of DL techniques for anomaly detection.
However, in most cases, these techniques are applied in a rather brute-force manner.
They ignore essential and, in many cases, apriori known information about the system structure and behavior.
Especially, they assume to take every possible connecting point into consideration, which is not efficient while modern cyber-physical systems could have thousands of potential assess points to be connected to.
The student will investigate several Simulink models as examples, and research a general solution for getting the best optimized access points given existing models.
The expected deliverable is a program that takes simulation models as input, analyzes the system architecture, and outputs the identification of suitable access points for anomaly detection.
The result can be verified with fault injection experiments.
1. literature research of related mbsa methods, such as dynamic fault trees, bayesian network, reachability analysis, importance metrics
2. Scenario study, get familiar with different case study models and corresponding simulations.
3. Test different approaches to the models
4. Compare the results of different approaches
Matlab/Simulink
Sheng Ding
♦
Deep learning performs remarkably well on many time series analysis tasks recently. The superior performance of deep neural networks relies heavily on a large number of training data to avoid overfitting. However, the labeled data of real-world time series applications may be limited, especially anomaly detection. As an effective way to enhance the size and quality of the training data, data augmentation is crucial to the successful application of deep learning models on time series data. In this thesis, the student will systematically review different data augmentation methods for time series anomaly detection.
The student will implement data augmentation on different CPS anomaly detection datasets and evaluate the results.
1) Literature research of existing anomaly detection datasets in Cyber-Physical Systems.
2) Literature research of state-of-the-art data augmentation method.
3) Categorization of the datasets according to the application field and characteristics.
4) Extract normal and abnormal patterns from real-world data.
5) Transform the extracted patterns into synthetic data.
6) Evaluation of the augmented dataset.
python
Sheng Ding
♦
Nowadays, deep-learning-based methods for anomaly detection and diagnosis are attracting more and more attention. To make research results into the application, a good way of demonstration is to develop a service for a certain kind of task. In this thesis, the student will train neural networks on multivariate time series data from CPS systems for anomaly detection using existing code. The highlighted part is to deploy it as a containerized REST API. The use case is where externally collected sensor data is streamed to the API for near real-time anomaly detection analysis.
1) Train different models for the dataset using existing code.
2) Put the existing code for processing data in a REST API using the Flask framework.
3) Visualization of the data and detection.
4) Deploy it within a Docker container.
5) Optional: Use Kubernetes for exposing the API as a service.
python
Sheng Ding
♦