idw – Informationsdienst Wissenschaft
Nachrichten, Termine, Experten
Nachrichten, Termine, Experten
AI language models are considered a key technology for industry, but they often face obstacles due to privacy concerns and their lack of expertise. Researchers at the University of Stuttgart and research partners at WRS Energie + Druckluft GmbH have developed and successfully tested a fully locally operated AI system. This AI uses a digital twin to provide concrete operational recommendations for compressed air systems without hallucinating. The new on-premises system is more secure and faster than comparable cloud solutions.
Smart compressed air system: Bridging the gap between complex data and usability
“Those looking to implement artificial intelligence in industry often face a dilemma: Cloud solutions are scalable, but they pose a security risk to sensitive production data. We have now resolved this conflict with our new AI assistant,” says Christian Wolf, a researcher at the Institute for Energy Efficiency in Production at the University of Stuttgart (EEP). As part of the “AI4Air” research project, EEP and WRS Energie + Druckluft GmbH have jointly developed and successfully tested, at an early stage of development, an on-premise large language model (LLM) that bridges the gap between complex plant data and user-friendly operation. The goal is to develop an intelligent compressed air system that can be used, for example, in robotics, automotive body construction, or the food industry. At its core is a proprietary AI that makes targeted use of relevant system information, enabling secure, high-performance, and application-specific interaction with compressed air systems through a direct connection to the plant’s digital twin.
Local language model instead of the cloud: security, speed, and control
While many industrial AI applications rely on general-purpose language models and external cloud services, “AI4Air” takes a different approach: The AI is fed with specific plant-related knowledge, operates entirely on-site, and is directly integrated into the technical system. This means that the specialized language models do not run on external servers, but in a secure local environment. The evaluation of the pilot phase shows that the system is superior to conventional cloud solutions. It responds to complex questions in an average of ten seconds. For questions specific to compressed air, the model achieves an accuracy of over 90 %, and sensitive production data never leaves the company. “This approach could provide a decisive competitive advantage, particularly for manufacturing companies with high standards for IT security and availability,” says Lennard Schwidurski, CEO of WRS.
AI data access: retrieving knowledge in real time right where it is generated
A well-known problem with generative AI models is “hallucinating” – making up facts. A key distinguishing feature of “AI4Air” is a reference guide developed in-house. It is based on a specialized information processing technique called Retrieval-Augmented Generation (RAG). Using RAG, the language model can specifically access a curated repository of sensor data, information from technical documentation, and maintenance logs. The result: Clear, context-based answers instead of generic AI responses. The hallucination rate drops to 26% - compared to up to 40% for standard models. Since the system always uses the latest system and status data, the information is updated in real time. “Our AI doesn’t just answer questions – it backs up its statements with real plant and digital twin data,” explains Wolf, the AI4Air project manager. “This is a crucial step from mere assistance to reliable decision support.”
Large language model meets digital twin: from analysis to operational recommendations
The core technology that sets “AI4Air” apart from standard chatbot solutions is its integration with the digital twin. While the twin provides a real-time physical and energy-based representation of the compressed air system, the LLM acts as an “interpreter.” It translates abstract data trends into operational recommendations. In this way, the system does not merely report an abstract drop in pressure, but prioritizes specific actions. For example, it can explain the current efficiency levels of individual compressed air components, identify the causes of rising energy consumption, prioritize specific measures such as fixing leaks or scheduling maintenance, and assess the impact of decisions. “The combination of a digital twin and conversational AI makes complex relationships intuitively accessible for the first time,” emphasizes Schwidurski. For example, it can make in-depth expert knowledge accessible even to non-specialists within the company. Following the successful evaluation of the on-premises architecture, “AI4Air” is now moving into the next phase: expanding the pilot applications with industry partners. The AI-powered decision-making tool will be further refined to unlock additional savings potential in energy-intensive compressed air networks – fully localized, explainable, and directly connected to real-world conditions.
About the “AI4Air” project
The “AI4Air” project (duration: 2 years) is funded by the InvestBW innovation support program. The project is a research collaboration between the Institute for Energy Efficiency in Production (EEP) and WRS Energie + Druckluft GmbH. Through InvestBW, the Baden-Württemberg Ministry of Economic Affairs provides financial support – particularly to SMEs and startups – for the development of new products, services, and business models in future-oriented fields such as artificial intelligence and digitalization.
University of Stuttgart, Institute for Energy Efficiency in Production (EEP)
M.Sc. Christian Wolf
christian.wolf@eep.uni-stuttgart.de
WRS Energie + Druckluft GmbH
M.Eng. Lennard Schwidurski
l.schwidurski@wrs-energie.de
https://www.uni-stuttgart.de/en/university/news/all/New-AI-assistant-for-the-saf... Press Release University of Stuttgart
Real-time data analysis: The foundation for the AI-powered compressed air assistant for automatic le ...
Quelle: Rainer Bez
Copyright: EEP
Together, they are bringing intelligence to compressed air technology: from left to right, Christian ...
Quelle: Rainer Bez
Copyright: EEP
Merkmale dieser Pressemitteilung:
Journalisten, Wirtschaftsvertreter, Wissenschaftler
Elektrotechnik, Informationstechnik, Maschinenbau
überregional
Forschungs- / Wissenstransfer, Forschungsergebnisse
Englisch
Sie können Suchbegriffe mit und, oder und / oder nicht verknüpfen, z. B. Philo nicht logie.
Verknüpfungen können Sie mit Klammern voneinander trennen, z. B. (Philo nicht logie) oder (Psycho und logie).
Zusammenhängende Worte werden als Wortgruppe gesucht, wenn Sie sie in Anführungsstriche setzen, z. B. „Bundesrepublik Deutschland“.
Die Erweiterte Suche können Sie auch nutzen, ohne Suchbegriffe einzugeben. Sie orientiert sich dann an den Kriterien, die Sie ausgewählt haben (z. B. nach dem Land oder dem Sachgebiet).
Haben Sie in einer Kategorie kein Kriterium ausgewählt, wird die gesamte Kategorie durchsucht (z.B. alle Sachgebiete oder alle Länder).
