Using LLMs to generate modular anomaly detection solutions in automation

• Applicants: Professor Dr.-Ing. Alexander Fay; Professor Dr. Oliver Niggemann
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 559936667

Engineering science has specific requirements for Artificial Intelligence (AI): In addition to requirements for reliability and specific development processes, the main difference is that usually only a small amount of data (Small Data instead of Big Data) is available regarding the interesting system states, e.g. anomalies. In AI, the established solution for this is the use of prior knowledge. This proposal explores the use of prior knowledge with a focus on the improvement of anomaly detection in production plants. Prior knowledge can be divided into knowledge about the structure and knowledge about the dynamics of the system under investigation. Previous work, particularly by the applicants, has shown that neural networks work more efficiently when they reflect the system’s structure. Although the system structure is often roughly known and there are formal models to describe it, the structure is rarely available in sufficiently formalized form in practice. Various recent studies, as well as investigations by both applicants, have shown that Large Language Models (LLMs) such as ChatGPT or Claude are capable of generating structural models and modularizations from little to no formalized system descriptions or documentation. However, previous work in this area has led to results of very different quality. The reasons for this are still unknown. This leads to the first research question of the proposed project: how to ensure sufficiently good structural models with the help of LLMs. Recent research has shown that neural networks for anomaly detection within individual modules of the system are more robust in terms of defined robustness criteria when prior knowledge about the dynamics of the process is used. Physics-informed neural networks (PINNs) are an established approach for incorporating prior knowledge if the ordinary differential equations (ODEs) are known. Unfortunately, these ODEs are usually not included in the system description or the existing simulation models of the modules. However, recent research suggests that such ODEs can also also be generated in many cases using LLMs. This leads to another research question of the proposed project: the creation of sufficiently good dynamic models with the help of LLMs. Through the modularization described above, the problem becomes manageable, as modules (unlike entire systems) are more generic and therefore more familiar to the LLM. In this proposal, LLMs are first used to generate a structural model (including modularization) and then, for each module, to generate the equations necessary for PINNs that describe the dynamics. The approach will be evaluated using the anomaly detection application in production systems. This evaluation approach will proceed in phases with growing system complexity.

• DFG Programme: Research Grants