Machhamer, R., Dziubany, M., Czenkusch, L., Laux, H., Schmeink, A., Gollmer, K.-U., Naumann, S. and Dartmann, G.: Online Offline Learning for Sound-based Indoor Localization Using Low-cost Hardware, in IEEE Access, Oct. 2019, 10.1109/ACCESS.2019.2947581 ©2019 IEEE
Peine, A., Hallawa, A., Schöffski, O., Dartmann, G., Begic Fazlic, L., Schmeink, A., Marx, G. and Martin, L.: A Deep Learning Approach for Managing Medical Consumable Materials in Intensive Care Units via Convolutional Neural Networks: Technical Proof-of-Concept Study, in JMIR Med Inform, Vol. 7, No. 4, p. e14806, Oct. 2019, ISSN: 2291-9694, 10.2196/14806
Ayad, A., Zamani, A., Schmeink, A. and Dartmann, G.: Design and Implementation of a Hybrid Anomaly Detection System for IoT, in Sixth International Conference on Internet of Things: Systems, Management and Security (IOTSMS), IEEE, Oct. 2019, accepted for publication ©2019 IEEE
Hauck, M., Machhamer, R., Czenkusch, L., Gollmer, K.-U. and Dartmann, G.: Node and Block-based Development Tools for Distributed Systems with AI Applications, in IEEE Access, Sep. 2019, 10.1109/ACCESS.2019.2940113 ©2019 IEEE
My research is at the interface between electrical engineering and computer science and interdisciplinary with projects mechanical engineering (e.g. BMVI - APEROL) or medicine (e.g. BMBF - IMEDALytics).
My research group works on distributed systems, application specific hardware, hardware-software-co-design, data analysis, signal processing & machine learning and security in communication systems and has a special focus on the application of these technologies in different disciplines (medicine, logistics, mobility, middle class).
The Internet of Things as the basis for industry 4.0 is ubiquitous and poses new challenges to science and society. The fields of mechanical engineering, electrical engineering and computer science will therefore increasingly converge in the future. Therefore, the term cyber-physical systems is often used to make it clear that technology will merge with our environment in the future. This networking and digitalization can open up major opportunities for meeting global challenges, in addition to risks such as threats to information security, which are also a central topic of my research group.
A central aspect of distributed systems is the distributed implementation of classical concepts such as signal processing and machine learning. Earlier signal processing algorithms were optimized for low-dimensional signals and implementation in DSPs was correspondingly simple. Today's applications generate high-dimensional data sets and require the implementation of more complex data processing systems and algorithms. Processing steps of signal processing and machine learning are increasingly implemented in complex distributed processing chains. Classical concepts of computer science and information technology must therefore be improved for future intelligent cyberphysical systems (CPS). Due to their distributed and decentralized nature, new distributed and efficient optimization methods, algorithms for data analysis and knowledge extraction and hardware platforms must be researched for CPS.
In the following, the three central research areas of my research group are described exemplarily:
Research area 1 - Platforms for digital services:
IoT & AI platforms: The most valuable Internet platforms in the world are in the USA and Asia. European IT platforms are far behind. In the new wave of digitalization, the physical world is now being connected to the Internet and smart services are being developed. For the German economy with its strong industry, the development of suitable platforms is enormously important so that it does not lose its leading position in the areas of industry, mobility and medicine. In particular SMEs needs simple and easily manageable platforms with which new technologies such as Blockchain & AI can be used for new services and business ideas. In order to accelerate the development of new Smart Services and business ideas for the Internet of Things, we want to develop tools and platforms specifically for SMEs.
- Development of an IoT platform (hardware/software) for medium-sized companies
- Development tools for hardware-software co-design
- Development Tools for data engineering
Smart Services & Contracts: The development of intelligent services is a central aspect of my research group. In the BMVI project APEROL, we develop novel distributed and block-based services for intelligent mobility. Expemplary these services will be implemented as an indoor post service in an autonomous vehicle. The research here includes both the development of a suitable front-end (app) and the development of suitable distributed communication platforms in the backend, which can be implemented decentrally and distributed.
Research Field 2 - Data Analysis and Signal Processing
In this field of research, we distinguish between algorithms for real-time systems, algorithms for Big Data applications and model-based learning. We are especially concerned with the application of these algorithms. In our research we always see the real system with the actual data first. These must first be analyzed and understood. The aspect of modelling is therefore still important. Future research approaches should deal with the fusion of data-based and model-based approaches, because model-based approaches allow the integration of expert knowledge and work with smaller amounts of data. Some exemplary research projects are outlined below:
Distributed signal processing and control for multi-agent systems: Many cyber-physical systems are so-called multi-agent systems, i.e. they consist of several independent agents that together have a larger goal. These agents are usually not controlled by a central unit. Typical examples are the new micro-satellites or swarms of drones. For example, in order to maintain a formation, the agents must pass on their local information (e.g. state vector, distance) to their neighbors. The agents in turn have to communicate via radio and thus use the available resources very efficiently, i.e. only communicate when necessary. This is why learning methods are interesting that can learn and predict the system behavior of the entire agent system. Furthermore, the algorithms for localization and for learning and estimating the future system states must be implemented in a distributed manner and must be extremely efficient, since most agent systems have only a limited energy budget available.
- D. Hinkelmann, A. Schmeink and G. Dartmann. Distributed learning-based state prediction for multi-agent systems with reduced communication effort Workshop on Sensor Data Fusion and Machine Learning for next Generation of Cyber-Physical-Systems in conjunction with ACM International Conference on Computing Frontiers 2018
- S. Tedik Basaran, G. Karabulut Kurt, G. Oke Günel, A. Schmeink, G. Ascheid and G. Dartmann, The Safety Analysis: Disagreement of Wireless Communication-based Consensus. IEEE Wireless Communication Letters, 2018
Machine Learning for low-cost hardware: For many applications in the area of warehouse and production technology, cost-effective, reliable localization in the indoor area is interesting. We are currently investigating methods that detect fingerprints of sound signals and can thus continuously learn a position in space. With these methods, very cost-effective microphones could be used for localization. The microphones could, for example, be installed in the IoT Octopus developed by the IoT expert group. Another area is the development of a low-cost artificial "nose" that can classify a wide variety of substances with the help of trained gas sensors.
- H. Laux, A. Bytyn, G. Ascheid, A. Schmeink, G. Karabulut Kurt and G. Dartmann. Learning-Based Indoor Localization for Industrial Applications. Workshop on Sensor Data Fusion and Machine Learning for next Generation of Cyber-Physical-Systems in conjunction with ACM International Conference on Computing Frontiers 2018
- M. Dziubany, M. Garling, A. Schmeink, G. Burger, G. Dartmann, S. Naumann, K. Gollmer, Machine Learning Based Artificial Nose on a Low-Cost IoT-Hardware, Chapter in Big Data Analytics for Cyber-Physical Systems, Machine Learning for the Internet of Things, 1st Edition, Elsevier, Paperback ISBN: 9780128166376, Publication date planed July 2019
Big Data Analytics in Industry: In addition to hardware and software, the integration of an analytics tool chain for process data analysis is an important aspect for applications with many sensors and data, in particular for industry 4.0. This includes, for example, classification, regression and logistic regression of measurement data. For this purpose, machine learning concepts are to be provided in software blocks and integrated into the hardware-software platform of F3. Additional data can be collected through the gateways, e.g. in production. Based on this data, processes can be optimized and new business models can be generated. Together with the IoT expert group, we have developed an edge cloud system that can be positioned locally in the production hall, for example. The process data can be collected and processed on site. Algorithms must be developed for this. In addition to optimizing production processes, data analysis is important for anomaly detection in order to detect IT attacks at an early stage. Anomaly detection requires suitable models of the overall system and continuous recording of all system states.
Associated third-party funded project: COSY (BMBF-funded): Cognitive Tools for Cyber-physical Systems
- Development of algorithms for Machine Learning for IoT and Industry 4.0
- Localization using Machine Learning and IoT
- Development of an artificial nose
- Development of algorithms for system identification
- Data-based models for environmental sensor data
Big Data Analytics in Medicine: In this project we want to develop adaptive models for the diagnosis of various diseases in intensive care units. These models will help young physicians to better understand the complex processes involved in blood poisoning (sepsis). The system should be able to predict different disease progressions based on adaptive models and current measurement data. The system continuously records medical measurement data from the telemedicine system in a large database. On the basis of this constantly growing database, models can be trained for certain disease courses. These stocastic models should be valid for a large number of patients as well as individually optimized for certain classes of patients. Both neuronal networks and stochastic Petri nets are suitable models.
- P. Vieting, R. C. de Lamare, L. Martin, G. Dartmann* and A. Schmeink*, Likelihood-Based Adaptive Learning in Stochastic State-Based Models, IEEE Signal Processing Letters, (accepted for publication, *Corresponding Authors), 2019
Associated third-party funded project: IMEDALytics (BMBF-funded)
- Machine Learning Methods and New Algorithms for Medical Data Analysis
- Decision support systems
- New stochastic models for medical data
- Integration of expert knowledge
Model-based approaches & digital twin: A major challenge is learning with small amounts of data (low-data problem), because the generation of data, for example for research into medical drugs, is very expensive. For this reason, we investigate model-based teaching methods in combination with data-based methods for computer-aided drug optimization. Such methods reduce the number of necessary laboratory experiments and could make drug development more cost-effective. Continuous Petri nets, which are already widely used for modelling systems biology processes, can be used as models. We investigate both algorithms that learn the parameters of given Petri nets (known structure) and concepts that learn not only the parameters but also the entire structure of these Petri nets. In contrast to neural networks, models based on Petri nets are not black box models. Such white-box models are more accepted in medicine and also allow expert knowledge to be taken into account.
- E. Zechendorf, P. Vaßen, J. Zhang, A. Hallawa, A. Martincuks, O. Krenkel, G. Müller-Newen, T. Schuerholz, T.-P. Simon, G. Marx, G. Ascheid, A. Schmeink, G. Dartmann, C. Thiemermann and L. Martin. Heparan sulfate induces necroptosis in murine cardiomyocytes - a Medical-In-Silico approach using machine learning, Frontiers in Immunology, 2018.
Research Field 3 - Technology for Distributed Systems:
Low-Cost Application specific Hardware: Together with the IoT Expert Group of the National Digital Summit, we have developed a generic IoT device. This device is characterized by low costs and flexibility. An important aspect of this gateway is the development of a generic hardware interface to various components in industrial applications. Another aspect is the connection of additional sensors. Radio-based sensors are interesting here because of their flexibility. The device should also be a basis for industrial projects.
- M. Dziubany, R. Machhamer, H. Laux, A. Schmeink, K. Gollmer, G. Burger and G. Dartmann. Machine Learning Based Indoor Localization Using a Representative k-Nearest-Neighbor Classifier on a Low-Cost IoT Hardware, 26th European Signal Processing Conference (EUSIPCO), 2018
IoT-CERT (Hardware and Security): Together with the Internet of Things expert group, my research group is working on a monitoring system for M2M cyber security. The aim here is to develop a certified IoT hardware platform that can serve as a SIEM sensor (SIEM: Security Information and Event Management). This platform serves as a honeypot in IoT systems and is intended to record data traffic and forward it for analysis and anomaly detection.
Hardware-software interfaces: In addition to the hardware, a semantic software interface is another important aspect of this project. Due to the diversity of different connected components, aspects of the respective hardware have to be abstracted. Self-learning software can learn new components and automatically generate test code for interface adaptation in the future. These tools from the BMEL-project IoT-Pilot should partially demonstrate these skills.
High-Performance Hardware Platforms: One challenge in the automotive sector is the development of machine learning procedures for autonomous driving. The range of these methods is wide and there are now powerful hardware platforms that can implement such machine learning methods in real time. The optimization of these algorithms is elementary to achieve real-time capability. At the RWTH Aachen University, I won the BMBF project PARIS before being appointed to the Trier University of Applied Sciences. In this project parallelization methods for the optimization of these algorithms are developed.
associated third-party funded project: APEROL (BMVI-funded)
- Development of Deep Learning methods for autonomous fleets
- Distributed AI systems and implementation in hardware
IT security: Cyber-physical systems are not accepted in industry without a guarantee of security. A future cyber-physical system must be able to detect possible attacks with the help of e.g. data analytics methods. Radio-based systems have become more and more interesting for industrial applications due to their flexibility. In traffic (autonomous driving), communication is only possible via radio. Many such systems are embedded and do not have complex encryption mechanisms. Due to the distributed nature of the radio medium, it is an easy target for attackers. Therefore I have been researching concepts to achieve IT security in layers 1 and 2 for several years. Here the working group cooperates with the colleague Prof. Gunes Karabulut Kurt of the ITU Istanbul.
- O. Cepheli, G. Dartmann, G. Karabulut Kurt, and G. Ascheid A Joint Optimization Scheme for Artificial Noise and Transmit Filter for Half and Full Duplex Wireless Cyber Physical Systems. IEEE Transactions on Sustainable Computing, 2017.
More details of my research group at University of Applied Sciences Trier.