New CD Laboratory at TU Graz: Data-Controlled Condition Monitoring in Steel Production

- EN - DE
In the new CD laboratory sensor data from rough surroundings is being used. Imag
In the new CD laboratory sensor data from rough surroundings is being used. Image Source: RHI Magnesita
By Susanne Filzwieser

In the new Christian Doppler Laboratory for Reliable Systems in Harsh Environments, researchers at TU Graz, supported by the refractories group RHI Magnesita, are focusing on data-driven condition monitoring in the steel production process.

Things are heating up in steel production. Inside blast furnaces, temperatures can rise up to 1700 degrees Celsius. They are thus the epitome of a "harsh environment". Refractory materials that can withstand these conditions are therefore indispensable - and not only in the steel industry, but wherever high-temperature processes are used. Refractory products, such as those manufactured by RHI Magnesita, not only safeguard materials during high-temperature processes, but also protect firing and furnace systems from thermal, mechanical and chemical stress. But how can the condition of the refractory lining of the plants be monitored as efficiently and precisely as possible in a data-controlled manner?


It is this question that Franz Pernkopf and his team from the Institute of Signal Processing and Speech Communication at Graz University of Technology (TU Graz) are addressing: "We want to further advance the data-based condition monitoring of machines in harsh, industrial applications and have deliberately chosen the steel production application example for this purpose." Pernkopf heads the Christian Doppler Laboratory for Reliable Intelligent Systems in Harsh Environments, which was opened today, 17 April 2023, together with corporate partner RHI Magnesita. Research will now continue for seven years with financial and scientific support from RHI Magnesita. The largest public funding body for the CD lab is the Federal Ministry of Labour and Economic Affairs (BMAW).

Machine learning for process optimisation

"Today, most application areas of machine learning are in the virtual world, keyword recommender systems, stock market forecasts or social media," says Franz Pernkopf. Intelligent systems that model complex dependencies and draw conclusions from the huge amounts of data collected are, however, also urgently needed for industrial applications. Franz Pernkopf: "We are currently seeing the transition of machine learning systems from the virtual to the real world, for example in applications for autonomous navigation, the Internet of Things and Industry 4.0. It is clear that this transition is ushering in various challenges. These have to be overcome."

"We are excited about the collaboration with the University and the opportunities it presents. Thanks to this cooperation, we can optimise the refractory products in use by our customers with the latest findings in the field of machine learning. We are convinced that this cooperation will not only enable us to make our customers’ processes more efficient and our own products more sustainable, but also to offer new digital solutions and products to support our customers even better," explains Constantin Beelitz, Regional President Europe, & Turkey of RHI Magnesita.

Closing the gap to industrial applications

Current machine learning methods are particularly effective when large amounts of data and sufficient computing resources are available. In many real-world application fields, however, data is not always available in abundance and in sufficient quality. In addition, the infrastructure itself, the data source, so to speak, is usually difficult to access during the operating phase - in blast furnaces, for example. "In this lab, we want to pave the way for the application of machine learning for condition monitoring in industrial environments. The aim is to close the gap between basic research in ML and industrial applications in harsh environments," says Franz Pernkopf.
In order to make significant progress in data-driven condition monitoring, research is being conducted in the following three areas:
  • Robust representation: Deep neural networks and hybrid models are being used to learn representations in order to avoid manual feature engineering. In addition, these are used for outlier detection, data augmentation and semi-supervised learning to counteract the limited amount of data or inferior data quality and to improve the generalisation capability of the models.
  • Learning and uncertainty estimation: Here Bayesian models are being adapted to provide uncertainty estimates for the predictions. Furthermore, methods are being developed to overcome domain shifts and transfer learning used for knowledge utilisation across related applications.
  • Explainability and process optimisation: Methods are being developed to understand the causes of the prediction of the black-box models. In addition, methods for model adaptation during the model use phase are being researched.

The fundamental innovation is the improvement of data-driven condition monitoring. The application under consideration is the modelling of refractory materials during production and their use in steel production. However, the results of the lab should be applicable to many other industrial processes.

Would you like to receive the latest stories, news, research stories, interviews or blog posts from TU Graz directly on your smartphone or in your email inbox? Subscribe to the TU Graz Telegram newsletter for free.