Can large language models (LLMs) reliably and safely generate industrial control applications?
This is the question addressed by the MBO-KISS funding project. Led by Fraunhofer IKS, six partners from research and industry are developing methods to evaluate and optimize AI-generated control applications. The aim is to develop a methodological approach that enables the safe use of LLMs for industrial control applications.
IEC 61131-3 is the established standard for programming controllers in industrial production. However, specialists with knowledge of IEC 61131-3 languages are rare, and the same is true for current LLMs: Publicly available LLMs have been trained primarily on high-level languages such as Python, Java, or C/C++, but not on IEC 61131-3 languages such as Structured Text (ST) or Sequential Function Chart (SFC). These languages are mainly used internally within companies, meaning that there is hardly any freely accessible training data. Furthermore, real-time requirements, communication interfaces to actuators and sensors, and safety-critical aspects are rarely part of the training data.
In addition to the language barrier, the complexity of industrial control applications is also an obstacle to code generation via LLMs. Therefore, suitable decomposition, breaking down applications into less complex components, and a clean interface design are essential for successful code generation and subsequent orchestration. They ultimately enable the individual generated code components to be assembled into a functioning application.
The MBO-KISS project uses existing LLMs and develops methods to use them for generating control applications according to IEC 61131-3, despite the limitations described above. LLMs are used throughout the entire development process, from requirements specification to test case generation.
To improve the available LLM models, they are retrained with proprietary data sets. Techniques such as retrieval-augmented generation (RAG) are also used in prompt creation. RAG improves LLM results and helps avoid AI hallucinations by allowing the LLM to access external, trusted data sources.
The generated control applications then undergo a three-stage test workflow to validate the results:
The results from these static and dynamic tests are incorporated into subsequent generation cycles. This means that the code is iteratively improved with each test run.
The project result is a comprehensively tested, methodical approach for the safe use of LLMs in industrial control applications. Initial interim results are already promising. The generated control applications showed few errors when compared with the code style guide. After several rounds of the testing workflow, we have already reduced the number of errors to one tenth.
Once the project is complete, manufacturing companies will be able to use the developed approach to safely and efficiently deploy AI-based solutions for code generation in industrial automation. This will reduce engineering costs, minimize the impact of the shortage of skilled workers, and enable an optimized development process that can be easily adapted to their own systems.
As part of the MBO-KISS project, Fraunhofer IKS is also investigating which LLMs are best suited for code generation in Structured Text (ST). To this end, we are developing a comprehensive benchmarking system, which we will continue to expand after the project and provide companies with guidance on using LLMs in industry.
The methods and approaches developed by Fraunhofer IKS within the project are available for further use in future projects and for industrial customers. This provides a valuable foundation for innovations in industrial automation and AI in industry.
Would you like to benefit from our expertise? Contact us using the contact form:
Contact us now
Das Projekt wird gefördert durch das bayerische Staatsministerium für Wirtschaft, Landesentwicklung und Energie im Programm BayVFP Förderlinie Digitalisierung.
Fraunhofer Institute for Cognitive Systems IKS