Projects

Christian Schuster, Morten Schierholz

Nov 6, 2024

Electromagnetic compatibility (EMC) deals with the suppression of unwanted electromagnetic interference between electronic devices, systems and components. Statutory EMC guidelines must be complied with by all products before they receive market approval. Typically, complex simulations and measurement series during the design phase ensure that this can be achieved.
Increasing requirements in the field of EMC - think for example of the advancing wireless communication at ever higher frequencies - require a continuous development of engineering methods in order to make the right decisions early and cost-effectively during development. For this reason, various machine learning methods are being investigated in this project with regard to their applicability to the synthesis and analysis of electrical systems, as well as their impact on the environment. The fields of application of

• Signal integrity of wired channels,
• Stability of the power supply (power integrity ) of digital systems,
• radiation from printed circuit boards and cable harnesses,
• Bioelectromagnetic compatibility and
• Electromagnetic compatibility in the drivetrain of electric cars

is of particular importance. Our focus is currently on artificial neural networks, which are used to analyze circuit board structures and in relation to the optimization of data transmission as well as the decoupling of the power supply.
In order to increase data availability in the research fields, the SI/PI database has been developed and is currently being transferred to version 2.0.

Selected publications:

• Efficient Iterative Data Generation Using Evaluation of Prioritized Input Parameters in ANNs for SAR Prediction in Human Head Models at 13.56 MH
• Generation and Application of a Very Large Dataset for Signal Integrity Via Array and Link Analysis
• Machine Learning Based Data Validation for Signal Integrity and Power Integrity Using Supervised Time Series Classification
• SI/PI-Database of PCB-Based Interconnects for Machine Learning Applications

Volker Turau, Marcus Venzke

Sep 20, 2024

The project investigates, how machine learning (ML) is efficiently used on low-performance microcontrollers. Typical applications of embedded systems are considered processing sensor data. The microcontrollers used have 8 or 32 bits, between 2 kB and 512 kB RAM and only floating-point arithmetic in hardware (FPU) in some cases. Frequently, devices operated on batteries must have long operating times. Hence, for ML, major challenges exist in adapting energy, computing, and storage resources to the demands of such systems.

For ML, supervised and unsupervised learning is considered based on artificial neural networks (ANN), decision trees/forests and other techniques. Resource requirements are reduced by combining different approaches. The ML problem is reduced to the minimum required for the application, e.g. inputs for features are pre-processed and unnecessary ones are removed. Applied ANNs must not be chosen too large or must be compressed. Sub-problems not required to be executed in the embedded system (e.g. training and optimization runs) are performed on other, more powerful systems.

Two applications have been realized so far. It is a module for optical hand-gesture recognition and another for localizing transport boxes in industrial processes. For hand-gesture recognition, ANNs and random forests are trained with supervised learning, executed in a module with 9 light sensors (a kind of compound eye) and microcontroller Atmega4809 (8-Bit, 20 MHz, 6 kB RAM). For localization of transport boxes, a so-called threshold tree is trained with unsupervised learning, used in a battery-operated module with several sensors (e.g. acceleration, gyration, air pressure, light) and microcontroller ESP32 (32-Bit, 240 MHz, 512 kB RAM).

Alexander Klemps

Sept 16, 2024

The European XFEL at DESY is a world-leading research infrastructure in Hamburg (Germany), enabling scientists to observe and investigate microstructural processes and phenomena with spatial and temporal resolutions on the atomic and femtosecond scale. To improve the performance of the accelerator and also to ensure its competitiveness with other facilities, it is essential to optimize the EuXFEL for operation in continuous-wave (CW) mode in near future. Despite its advantages, an operation in CW mode requires a reduction of the beam energy and is associated with an increase in the geometric beam emittance. To still ensure the delivery of high quality beams, we pursue within the OPAL-FEL project the optimization of crucial beam properties such as the beam emittance by applying data driven methods.
Central to our approach is the implementation of a so called inverse model, predicting the optimal state of the photoinjector corresponding to a desired optimal emittance. The data used to learn such a model will be generated by forward processes like simulations or experiments. Next to the pure implementation of the inverse model itself, we also put a strong focus on the mathematical justification of our approach. Here our interest lies in the general invertibility of the involved forward maps and how effectively neural networks can approximate the corresponding inverse maps, leading to research in the field of inverse problems.

Julian Hoth

Sep 19, 2023

The data collection approach has been validated in various research projects and is now considered cutting-edge for cloud-based and centralized big data systems, with commercial implementation as the next step. This allows data analysts to enhance airline business processes. Currently, aircraft systems focus on connectivity, wireless communication, and data collection for condition diagnostics. However, the concept of distributed intelligence and decentralized learning, similar to natural systems, remains unexplored in aircraft systems.

The TUHH sub-project aims to design the first architecture for an AI-powered cabin system within the aircraft, focusing on catering processes and the catering value-added network through digital communication and machine learning. This architecture will be developed, tested, and examined with project partners. TUHH employs model-based systems engineering tools, specifically the SysML language, which has been proven in previous projects. Additionally, it will explore and expand technologies for cyber-physical cabin systems, available in the Institute of Aircraft Cabin Systems' laboratory, to create AI-driven systems for aviation digitalization.

Benedikt Kriegesmann

Mar 28, 2023

The SuCoHS project is concerned with the development of new materials, designs and processes for the production of fiber composite structures. It is investigating the extent to which weight and costs can be reduced if such structures are also used in areas with high temperatures. In this project, the TUHH is working on the numerical simulation of thermally and structurally loaded fiber composite components. In particular, the influence of stochastically occurring manufacturing defects and scattering temperature distribution on the load-bearing capacity of such components is to be quantified. This requires such a large number of non-linear simulations that it could not be mapped in the project. Therefore, artificial neural networks (ANN) are trained to replace the time-consuming simulations. Once a neural network has been trained, the structural response (load-bearing capacity) can be determined within seconds for a large number of combinations of scattering input variables (material parameters, manufacturing defects, temperature distributions).

Benedikt Kriegesmann, Benedikt Hamann

Mar 28, 2023

The overall objective of the current proposal is the development of a process for efficient topology optimization of fail-safe structures. The basic idea for this process is summarized in the figure below. For any new optimization problem, a “plain” topology optimization without consideration of fail-safety is carried out. Then, a convolutional neural network (CNN) derives a redundant version of the optimized topology. In a final step, the design suggested by the CNN is fine-tuned by a shape optimization in the density field.

Before the process can be applied, it is of course necessary to train the CNN for the task. Therefore, in the proposed project numerous combinations of design spaces, boundary conditions and loads will be considered, and for each combination both, a plain topology optimization and an expansive fail-safe topology optimization will be conducted. The CNN will be trained to map the density field originating from the plain topology optimization to the fail-safe topology.

Merten Stender, Norbert Hoffmann

Feb 22, 2023

This project employs advanced machine learning (ML) and deep learning techniques to analyze brake dust and noise emissions, which are significant environmental and noise pollutants. Vehicle brakes operate through mechanical friction between a brake pad and a friction surface, converting kinetic energy into other forms of energy, leading to wear and noise. ML models are used to simulate emissions under varying conditions, surpassing traditional methods. These models help create virtual twins to identify sensitivities in mechanical systems and optimize brake designs, such as by choosing different materials. Additionally, ML techniques like image processing, object recognition, deep recurrent networks, and clustering are applied to detect brake noise, predict emissions, and analyze complex data patterns.

Claus Emmelmann, Katharina Bartsch, Bastian Bossen

Jan 10, 2022

Additive manufacturing (AM) is gaining importance due to its resource efficiency and design flexibility, particularly in aerospace and medical industries. However, the limited material options, especially for metallic powder bed processes, hinder the use of optimal materials for specific products. This is due to the costly and time-consuming empirical testing required to develop and qualify new AM materials.

The project aims to digitize material development using data-supported simulations and artificial intelligence (AI), focusing on new aluminum alloys in laser beam melting processes. Key objectives include testing alloy suitability, predicting parameter windows for powder atomization and processing, and evaluating recyclability through alloy-specific aging effects. An artificial neural network (ANN) will be developed to predict material behavior, using synthetic and empirical process data generated through simulations.

By reducing the costs and time for material development, the project aims to make AM materials more economical and enable tailored material selection. The AI and simulation approaches could also be extended to other material types and production methods, promoting efficient, resource-saving manufacturing of high-performance products.

Robert Meißner, Benjamin Boll

Nov 22, 2021

In this project, state-of-the-art machine learning and deep learning methods are used to analyze the quality of bonded fiber composite samples.

An important aspect here is the occurrence of so-called “weak-bonds”, i.e. areas in which adhesion exists but is significantly reduced. These areas cannot be detected using traditional measurement methods such as pulse-echo ultrasound. However, they lead to a significant reduction in strength. For this reason, bonded joints, for example in aviation, must always be secured by a second load path or an additional mechanical connection.

In this project, the samples are evaluated using two modulating vibrations based on the principle of vibroacoustic modulation (VAM). A low-frequency vibration with a high voltage amplitude and a high-frequency vibration in the ultrasonic range are introduced and the change in the modulation of both vibrations is observed.

The evaluation of the measurement data using neural networks has shown that patterns exist in the data that enable the weak bonds to be detected with a precision of over 90%. The initial results are presented in the previously published paper.

Robert Meißner, Christian Feiler, Tim Würger, Mikhail Zheludkevich

Apr 8, 2021

Magnesium has versatile properties and significant potential in industries such as aerospace, automotive, medical implants, and batteries. However, controlling its corrosion behavior is a major challenge. For transportation, preventing material degradation is essential to avoid critical failures, and environmentally friendly alternatives to toxic Cr(VI)-based corrosion protection are needed due to upcoming bans. In medical applications, magnesium implants must degrade at a controlled rate to allow healing before dissolution, while in batteries, continuous anode dissolution is required for maintaining constant voltage.

The project focuses on using small organic compounds as modulators to control magnesium degradation, introduced either through coatings or electrolytes. Due to the vast number of potential compounds, a purely experimental approach is inefficient. Therefore, the project aims to develop predictive models based on quantitative structure-property relationships and atomistic simulations to pre-select suitable additives more efficiently. This work builds on previous research from the Helmholtz Center Hereon, specifically the "Interface, Design, Engineering and Assessment" (IDEA) project, which laid the foundation for data-driven models. Collaboration with the Helmholtz-Zentrum Hereon allows for experimental validation of machine learning predictions, ensuring more sustainable and effective corrosion control strategies for magnesium.

Kathrin Fischer, Nicolas Rückert

Aug 11, 2020

"Business analytics" involves the systematic analysis of operational data, including past data review, forecasting, and data-driven decision-making using mathematical methods. In maritime logistics, which manages the flow of goods and information in sea-bound transport, business analytics offers significant potential. This sector generates vast amounts of data, such as ship movements and weather patterns, which can be leveraged to optimize operations like fleet and personnel deployment and enable innovations like autonomous shipping.

This project aims to explore the opportunities that business analytics offers in maritime logistics, as well as the operational and strategic risks involved. It emphasizes the importance of critically assessing the data analysis methods for potential ethical and strategic concerns. By applying advanced data analysis and optimization techniques, the project seeks to improve maritime logistics strategies while ensuring ethical considerations are addressed.

You can find more information on the ORIS Institute website.

Görschwin Fey

Jun 24, 2020

Embedded systems are used in a wide variety of areas, including safety-relevant areas (both safety and security). In these cases, not only the design of a correct system is crucial. In addition, it must also be guaranteed at runtime that no misbehavior occurs, which can be caused by external influences through “malicious attacks” or internal malfunctions. One solution for this is the automatic detection of anomalies that occur during use. A classic method for this is “monitoring” of the system, i.e. observation of the system by an independent unit. This independent unit decides whether the observed behavior is nominal or unexpected. In the latter case, active troubleshooting will be triggered.

Depending on the field of application, however, the design of a monitor is time-consuming, as the nominal behaviour must first be well understood, then formalized and then the monitor itself designed.

The project is therefore investigating how suitable machine learning methods are for first learning nominal behavior automatically and then using it to detect errors. In initial experiments, the approach was applied to a simulated robot vacuum cleaner, for example. Various types of internal errors were successfully detected.

Görschwin Fey

Jun 24, 2020

When using machine learning, not only the accuracy of the information learned but also the underlying computing system is of central importance for reliability. If errors occur in the computing infrastructure, the correctness of the application is jeopardized.

This project investigates precisely this relationship. In an initial analysis, a specific hardware architecture was modeled and errors were injected into the computing system. To this end, artificial neural networks were first trained using the Keras and Tensorflow frameworks. These neural networks are then exported and broken down into the underlying operations. The subsequent execution of classifications on the simulated hardware architecture allows errors to be injected and their effects to be determined at application level.

Volker Turau, Florian Meyer

Jun 24, 2020

Wireless transmission technologies, part of the Industrial Internet of Things (IIoT), are becoming widely adopted in industrial applications due to their lower costs and ease of use compared to wired networks. IEEE 802.15.4 DSME, a MAC protocol designed for IIoT, supports energy-efficient and scalable communication through time (TDMA) and frequency (FDMA) multiplexing, accommodating thousands of sensor nodes.

This project explores how machine learning can enhance data transmission in IEEE 802.15.4 DSME, particularly by optimizing transmission channels under interference and allocating time slots for data packets. Since time slot allocation is decentralized, finding the optimal solution to minimize transmission latency is complex. Reinforcement learning is a promising method, using artificial neural networks (ANNs) to generate and improve slot allocations based on real-time feedback.

Additionally, the project tackles the challenge of applying machine learning on resource-constrained IIoT devices, which often have limited RAM, battery power, and processing capabilities. The goal is to identify time-, energy-, and memory-efficient algorithms, like compressed KNNs, decision trees, or support vector machines, to make machine learning feasible for these devices.

Leo Dostal, Helge Grossert, Daniel Duecker, Robert Seifried

Jun 24, 2020

The design of mechanical systems that experience vibration requires accounting for both load amplitudes and frequencies to ensure fatigue strength. This project aims to develop simulation models that accurately represent these load cases, focusing on electric vehicles. Unlike traditional cars, electric vehicles have heavy batteries that significantly impact driving dynamics, particularly in vertical motion. For example, driving over uneven surfaces creates vibration frequencies much higher than the actual driving frequency, which conventional vehicle models struggle to simulate.

The project explores using Long Short Term Memory (LSTM) neural networks to model these dynamic behaviors. LSTMs are well-suited for predicting time series by calculating new values based on previous data. These models can function as independent "black box" models or be combined with multi-body models to enhance predictions, such as calculating forces and accelerations more accurately.

Additionally, neural networks can be integrated directly into mechanical model structures, acting as function approximators for specific system dynamics. This approach results in parameter optimization problems, which can be addressed using conventional methods, improving the accuracy of dynamic system models for electric vehicle design and other applications.

Thomas Wrona, Pauline Reinecke

Jun 24, 2020

Big data and analytics (BDA) involves extracting valuable insights from large datasets using advanced algorithms, aiming to enhance efficiency and create new value. However, its role in achieving competitive advantages and corporate success is not clear-cut. Key strategic questions remain, such as how BDA can support strategic decision-making without stifling creativity, how it might lead to disruptive innovations, and how it affects competitive dynamics, especially if it results in a race to the bottom on price due to similar strategies.

The TUHH lighthouse project, "Business Analytics - Optimization Potentials and Strategic Risks for Maritime Logistics Systems," seeks to address these issues. It aims to explore both the optimization potential and strategic risks of BDA from various academic perspectives, including computer science, mathematics, logistics, and corporate management. The project focuses on the maritime logistics industry, which is undergoing significant technological change and facing intense competition from both established tech giants and new startups with advanced BDA capabilities.

Christian Becker

Jun 24, 2020

Due to an ever increasing penetration of the electrical power system with power electronics coupled generation and transmission devices as well as loads, their dynamic behaviour will dominate the overall system dynamics in the future. In contrast to the dynamic behaviour of synchronous machines which still make up for most of the power generation today, this scheme does not supply any inertia and corresponding frequency response characteristics. Inverters equipped with novel control schemes like the virtual synchronous machine that seek to improve system dynamics are yet to gain any significance in the power generation portfolio. It is thus suspected that the system dynamics in general might develop negatively if plant controllers are not properly co-ordinated, and that the small signal stability in particular could deteriorate.

A monitoring setup for a power system’s small signal stability must work efficiently in terms of computational resources to operate in real time. They also need to suffice with sets of input data limited by the measurement and communication setup in the entire system. Finally, they must cope with a strongly nonlinear behavior of the power system. Given these requirements, computational intelligence is a viable approach. This project proposes a power system small signal stability monitoring method using artificial neural networks (ANNs). Once trained they can cope with nonlinearities and function computationally efficiently.