To explain the principle of real-time sensor monitoring, Trinitis, Professor of Computer Architecture and Operating Systems at the TUM School of Computation, Information and Technology (CIT) in Heilbronn, likes to reference Johann, the batty senior machine operator in the 1981 German box-office hit “Das Boot”, who constantly monitors the submarine’s diesel engines with an ear trumpet to detect unusual noises and potential damage early on.
Real-time sensor monitoring used in diverse critical infrastructures works in a similar way: Sensors listen to power plants, IT infrastructures, or steel production plants by monitoring a number of measuring data continuously. In this way, they ensure no damage or other irregularities occur and operations run reliably.
Advantages of processing data directly
During the process, the sensors generate a vast amount of data that, for now, is intended to be processed in clouds. The problem is that only a portion of the data reaches the clouds because, depending on the infrastructure, such transfers can be very laborious. Trinitis uses power plants as an example: “While they operate all over the world, fast grid connection is not available in some regions. Thus, it makes more sense to transfer less data, but the relevant kind.” The solution is to filter the mass of unremarkable data at the data source, referred to as the edge, and to send only critical mavericks to the cloud. Edge computing – processing data at the source – requires less energy and is more sustainable.
Combining edge and cloud computing and processing sensor data for artificial intelligence (AI) applications were the focal points of the SensE (Sensors on the Edge) project conducted by TUM in cooperation with Ingenieurbüro für Thermoakustik GmbH (IFTA; engineering office for thermoacoustics) in Puchheim in Upper Bavaria. The project, with a volume of approximately one million euros, ran from 2021 through the summer of 2024. Half of the project was financed by the Bavarian Research Foundation. Trinitis headed SensE jointly with Dr. Martin Schulz, Professor of Computer Architecture and Parallel Systems at TUM in Garching, and Dr. Roman Karlstetter, Technical Lead Software at IFTA.
Each problem causes unique vibrations
The project addressed one type of application of real-time sensor monitoring: a two-turbine gas power plant located in Germany. The sensors at the plant monitor the turbines’ acoustic vibrations in real time to ensure smooth operations – just like machine operator Johann did.
Trinitis reports on an incident detected in the recorded sensor data: “Once, a combustion chamber in the gas turbine broke down. This caused the characteristic vibrations in the entire device that point to this type of incident precisely. A tear in the shaft, for example, would trigger different vibrations.”
Combining edge and cloud computing also benefitted the training of an AI model for detecting anomalies: The time-consuming training was conducted in the cloud that was fed data from diverse edge sources. Thereafter, the model was returned to, and applied on, the edge.
Win-win for all
Trinitis summarizes the main project milestones: “We inspected different computer architectures for the edge to identify the ones that are best suited for data processing. In doing so, we analyzed a number of processor models and machine learning algorithms including the ‘Transformer’ everyone is talking about. Our doctoral student Dai Liu has dedicated a large portion of her work to applying these modern machine learning algorithms specifically to the processing of time-series data on the edge.”
Karlstetter of IFTA GmbH adds: “With SensE, we conducted research on issues with processing sensor data, particularly for use in AI scenarios, that had been unsolvable up to that point.” In his opinion, the most important outcome has been the creation of a demonstrator that continuously analyzes sensor data from high-performance gas turbines and is able to detect slowly developing damage several days in advance. “The demonstrator uses large amounts of historical sensor data to train an AI-based method,” he explains. “This allowed us to demonstrate that the methods developed as part of the research project are useful for solving actual problems in the industrial sector. Without the results gained from the project, implementing a comparable development technically would be difficult.”