Fremtidens forskere inden for ingeniørvidenskab støbes hos os. Vores ph.d.-studerende har høje akademiske ambitioner og leverer resultater af høj kvalitet til både den private og den offentlige sektor. Vores hovedfokus er anvendt forskning, og vi har et stærkt samarbejde med branchen for elektro- og computerteknologi, fordi vi forstår deres kerneudfordringer og bidrager til at udvikle løsningerne.
Her på siden kan du møde nogle af vores ph.d.-studerende og læse om deres projekter.
Regningen for korrosion løber årligt op i ca. 3 pct. af bruttoverdensproduktet, BVP, og der er derfor et stigende internationalt fokus på monitorering af infrastrukturprojekter. Institut for Elektro- og Computerteknologi, Aarhus Universitet, er i færd med at udvikle et smart plaster, der kan skære en stor del af rust-regningen.
I august 2018 kollapsede den 1.182 meter lange og 45 meter høje bro Ponte Morandi i Norditalien. 43 mennesker døde, 600 mistede deres hjem og Italiens stolte ingeniørvidenskabelige renommé led et eftertrykkeligt knæk. Og selv om den præcise årsag til kollapset endnu ikke er kendt, har efterforskere fundet beviser for, at korrosion og manglende strukturelt vedligehold var årsagen til den tragiske begivenhed.
Formålet er at udvikle en form for sensorplaster, der påsættes armeringen og støbes med ind i betonkonstruktionen. Sensoren og grænsefladeelektronikken drives ved hjælp af energihøstende teknologier for at sikre kontinuerlig overvågning af stålets tilstand,” siger Jaamac Hassan Hire, der er erhvervs-ph.d.-studerende på projektet.
Increases in computational power and storage within devices in Internet-of-Things(IoT) settings allows devices to utilize Machine Learning(ML). The advanced models from ML allows for more accurate decision-making, classification and inference. However, training these models is resource expensive and time consuming. An approach to train the model on is Federated Learning(FL). FL allows the training of a single model by multiple devices in a distributed fashion using only local information. Unfortunately IoT devices often require individual and specialized models due to their location and inherently different perceptions of the world. This project will examine object tracking and action recognition, where the individual devices are subject to changes in viewing angles, background and brightness.
The project will examine if it is possible to retain and combine the aspects of a globally trained model with the individual specialised models. Using this approach we can benefit from both the global and specialised models, exploring if inference is improved by combining the specialized model with the globally trained model. Furthermore I will explore the capability of individual devices to specialise their learning while enabling them to collaborate on different tasks.
This project is carried out with close collaboration with the Danish companies Vestas, Velux, Millpart, Hydrospecma, and Landia, and the Ringkøbing-Skjern municipality.
Each company focuses on different products and industries. The goal of this project intends to contribute to digital technologies of the involved companies for improving product design and production using virtualization and simulation methods, considering strategies to enhance optimization and efficiency of the processes.
This will be achieved through the combination and integration of already existing tools for digital twin representation and software for process and product management. This PhD project is associated with the Digital Transformation Lab (DTL) in Ringkøbing-Skjern Municipality targeting a digital transformation of the 5 companies in the area. The DTL will be used for the experimental parts of this PhD project. The DTL will be key in ensuring the best exchange of research-based knowledge from the university with practice-based knowledge at local companies.
Project title: Shortening time to market for product design using simulation and traceability in a digital twin context
PhD student: Santiago Gil Arboleda
Project period: 01.06.2021 – 31.05.2024
Main supervisor: Peter Gorm Larsen
Co-supervisor: Alexandros Iosifidis
Auditory Attention Decoding (AAD) is envisioned to become an important technology for the next generations of hearing aids. With an AAD component, a hearing device will be able to detect which sound sources the user is attending, and thereby the audio signal processing in the hearing device can be adapted accordingly.
In this project, we propose a new approach in which the AAD is based on cognitive processing of the audio event stimuli. This is expected to better separate attended and non-attended sound sources. The approach try to extract event related potentials and analyze using state-of-the-art deep learning model. We will validate the algorithm on Ear-EEG signal which is minimalistic and will allow the technology to be implemented in future hearing aid devices.
The ReMaRo ETN project aims to develop Artificial Intelligence methods for submarine robotics with quantified reliability, correctness specifications, models, tests, and analysis & verification methods.
Within this project, I am involved in the development of deep learning algorithms for vision-based navigation for underwater safety critical applications. One of the underwater contexts contemplated within the projects is pipeline inspection. That is, I will develop visual (camera-based) localization algorithms for pipeline inspection, supported by other sensors such as sonars. Moreover, I will work on the reliability assessment of these methods with the help of other PhD students from the ReMaRo consortium to identify whether if the localization measurements are reliable or not depending on the quality of the measurements taken.
My PhD project is part of MADE FAST in close collaboration with Vestas. In the last 15 years, the wind turbine has become bigger and bigger, and continues to do so. This means that the value chain is becoming increasingly challenged and transportations cost are rabidly increasing. An option to cope with this challenge is to move the manufacturing closer to the installation site. To this extent Vestas would like to create a movable factory that can be assembled/disassembled on site and transported in containers.
The vision of this project is to build a digital twin that enables establishing a movable factory solution that can be moved around the world and configured so the assembly processes can be conducted in a safe way. The digital twin will be able to demonstrate the assembly and disassembly of the movable factory for selected cases. In addition, it will be able to compare the actual assembly process sequence against the digital model and in case of larger discrepancies warn the appropriate operators.
The academic goal of the project is to provide new applied methods and results demonstrated in industry where we can document significant improvements with digital twins.
The Internet of Things (IoT) refers to a paradigm in which internet connectivity is ubiquitous among all kinds of devices everywhere. Associated with this is a massive increase in data collection and use, which has the potential to deliver huge long-term value to people and society. For example, reducing downtime and maintenance costs for production systems, improving autonomous vehicle safety and reducing environmental impact through efficiency improvements.
However, current technology is not prepared to deal with so many devices transmitting and using data simultaneously. To realise the benefits of IoT, an end-to-end framework that considers data compressing and data analysis wholistically is critical.
The goal of this project is to investigate the synergies and trade-offs between data compression and analysis. Specifically, this will involve developing algorithms for doing analytics directly on compressed data and optimising compression for both storage and analytics concurrently.
I am researching in optical nonlinearities to design a robust and scalable laser device with narrow linewidth for emission at various wavelengths. Nonlinear optics allows amongst other interesting effects, for wavelength conversion. This can be used for realizing lasers, at wavelengths otherwise hard to come by and have been used to achieve narrow linewidth. Implementing nonlinear optics in photonic integrated circuits (PIC’s) can enhance the nonlinear effects, while making it robust and small. This could for example allow for handheld instruments to measure and identify gasses in exhausted breaths for medical analysis.
Money laundering is a serious financial crime with devastating consequences; it enables drug dealing, corruption, and terrorism.
To combat money laundering, banks are required to monitor and report suspicious customer behavior to authorities. In practice, all banks address this task by use of electronic anti-money laundering (AML) systems. Traditional systems raise alarms based on pre-defined and fixed rules, effectively “if this, then that” statements. These systems exhibit poor performance, with up to 98% of all alarms being false positives.
The academic literature on statistical and machine learning methods for anti-money laundering (AML) is relatively limited. This is undoubtedly connected to the lack of publicly available data sets.
The Machine Learning for Anti-Money Laundering project aims to develop advanced machine learning models to raise and qualify AML alarms – Freeing up inquiry recourses for actual money laundering cases. The project is funded by Spar Nord bank, also making all its data available for the project. The project focuses on the application of recurrent neural networks and attempts to address three fundamental challenges of machine learning for anti-money laundering: extreme class imbalance, concept drift, and alert-feedback interaction.
Integrating manufacturing robots in the industry is currently a complex task, where system integrators spend multiple hours on selecting the appropriate parts and integrating them. The aim of this project is to decrease the integration time of manufacturing robots and lower the barrier between the untrained end-users and robotic technology. This will be done using digital twins, enabling simulation of the combination of different robotics parts. An example of this could be combining the necessary parts for a Pick and Place application, consisting of a base, a robotic arm, a gripper and sensors. This can then be simulated and displayed to the interested client.
The main outcome of this will be the ability to easily set up different robotic manufacturing applications and simulate them, depending on the clients needs. The clients can easily access the simulations through a web-platform, allowing them to better understand the potential robotic solutions they can adopt. This will create a state-of-the-art hardware and software one-stop-shop for automation, achieving the next big leap in democratizing robotic technologies.
People suffering from hearing loss can benefit from the use of hearing aids. To work properly, it is crucial that the hearing aids are fitted in close accordance with the hearing abilities of the individual hearing aid’s user. In some cases, hearing can deteriorate relatively quickly, especially with increasing age. Thus, it is important to re-fit the hearing aid recurrently.
Traditionally, hearing aid fitting is carried out in the clinic, where different behavioral tests are used to determine hearing thresholds. Alternatively, hearing loss can be characterized based on electrophysiological measures. This is typically based on the auditory steady-state responses (ASSR) recorded from a few electroencephalography (EEG) channels placed on the scalp.
Ear-EEG is a novel EEG recording method in which EEG signals are recorded from electrodes located on an earpiece placed in the ear. Ear-EEG can potentially enable integration of EEG recording into hearing aids and performing of ASSR based hearing threshold estimation in daily life.
Traditionally, ASSR based hearing threshold estimation has been performed using amplitude modulated continuous signals. Hearing tests based on monotonous stimuli of this kind make the user tired and unmotivated and can be inconvenient for the user, especially when the test must carry out for a long period of time.
Natural sounds such as speech is much more pleasant to listen to, thus speech-based hearing tests are more appealing to say yes to and can easier be implemented in daily life. Speech-based hearing test can be performed while the user, for instance, listen to the audio book.
The aim of this Ph.D. project is to investigate the possibilities of using the natural sounds, and in particular speech signals, to estimate hearing thresholds based on the ear-EEG.
Physicians are typically concerned with sounds originating from the body, which are traditionally listened to through an acoustic stethoscope. However, interpretation of sounds through such instruments is subject to the training of the individual physician. To overcome this challenge and further advance the method, an electronic version of the stethoscope has been invented, allowing for more objective interpretation of such sounds.
Ear-EEG is a technique specifically designed to monitor brain activity during everyday activities. Through an ear-fitted device, physiological signals reflecting the subject’s brain activity are measured.
This project aims to extend the ear-EEG platform with a body-coupled microphone, i.e. an electronic stethoscope in the ear, enabling recording in real-life environments with a discrete and unobtrusive wearable device. Furthermore, we seek to explore methods for joint processing of ear-EEG signals and sounds picked up from the body-coupled microphone . This would allow for a broader understanding of the current state of the body and is highly applicable in both research and in medical devices for health monitoring.
This project is carried out at the newly founded Center for Ear-EEG
Lasers are fundamental for modern telecommunication thanks to the reduced spectral distribution of their optical emission as compared to other light sources. This laser ’linewidth’ relates to the statistical notion of 'coherence' and it is inversely proportional to the distance (or time) over which the electromagnetic wave remains with a deterministic phase. However, this wonderful and unique property of lasers can be greatly influenced by the environmental conditions. In addition, advanced modulation formats employed in coherent communication require lasers with increasingly smaller linewidths, power consumption and footprint.
The goal of this project is to explore both theoretically and experimentally a new type of diode laser with narrow and stable linewidth. It will be designed as a tunable laser with high output power, operating in the telecom C-band. Beside telecommunication, possible applications that will benefit from this work include high-resolution spectroscopy, on-chip nonlinear optics and LIDAR. This laser design intends to reduce significantly the size, the weight, the cost and the complexity of its packaging without affecting its performance. It will be fabricated with mature material systems, thus giving this technology a feasible and credible path towards deployment out of the research laboratory.
Our society has become increasingly reliant on technical innovation to improve our quality of living. We expect our power grid to reliably supply electricity to our homes and production facilities and we expect our cars to take us safely to work every morning.
These are examples of Cyber-Physical Systems (CPSs), that is systems characterised by a strong coupling between the physical process and the software which controls it.
These types of systems are often safety critical as well, for example the failure of the power grid may cause disruption of other critical infrastructure, whereas the failure of a car may cause it to crash.
To keep up with the rising complexity of CPSs, computer assisted design (CAD) software is commonly used to simulate the individual parts of the system. The next step in this evolution is collaborative simulation (co-sim), where all components of a system are simulated at once. Rather than verifying only the individual components, this methodology makes it possible verify that whole system works as intended.
The central challenge in adopting a simulation-based approach to developing systems is the difficulty of creating the models. Today, this is very labour intensive and requires highly specialized expertise and software. This goal of this project is to develop machine learning based techniques for modelling CPSs, that provides accurate models without requiring knowledge of the internal workings of the system.
In recent years, the utilisation of deep learning has led to remarkable advances in computer vision and robotics applications such as self-driving cars. Despite this success, deep learning remains to become a prime technology for robotics. This is partly due to the prohibitive computational cost of current state of the art deep neural networks, and partly due to the lack of tailored development tools in the field. Open Deep Learning toolkit for Robotics (OpenDR) is an EU funded project aimed at the development of an open toolkit for robotics functionalities. Its focus lies on improving and making available the core AI and Cognition technologies needed in the years to come. Being a multinational project, the development effort is shared among many collaborators in both academia and the industry. My research will focus on real-time and lightweight deep learning architectures for solving computer vision tasks such as human action recognitions and object detection and tracking in recourse-constrained settings.
Deep Learning received a lot of attention in last years making lots of tasks be able to solve. However, it requires a lot of computational power to run most Deep Learning methods, making them badly suitable for low-resources systems like robotics. The main goal of my research is to create Deep Learning methods that can achieve both good performance and low computational costs for visual tasks.
Precision agriculture involves integration of new technologies such as satellite-based images to reduce soil and crop nursing. There are some important indices in agriculture, which are helpful in field caring, such as biomass, leaf area index, protein index, etc. To predict each of these indices, satellite images, multi-spectral images and radar need to be segmented separately. Then, the information of the multi-sensors is fused consequent to a map of the land areas. Deep learning algorithms show promising results in the classification and segmentation domain, and many papers have been working on improving network architectures. However, segmenting radar and satellite data demands a pixel-wise label as well. Hence, RGB ground-based high resolution images taken sparsely are going to be used in this project.
Through this estimation, three challenges will be met. Firstly, sparse ground-based images need to be globalised as the resolution of satellite-based images is much higher than plants' dimension. Secondly, multi-spectral images are going to be segmented based on labels generated in the first step. Finally, segmented results of radar and multi-spectral images will be fused together.
The IoT ecosystem enables massive sensor data acquisition, transmission and storage, as well as computation to perform data analytics for diverse smart monitoring, process automation and control services. The ever-increasing data generation brings many critical challenges to the communication, storage and computing infrastructure. The current infrastructure is not adequate to tackle these challenges in the long run. Firstly, the vast amount of generated data has to be represented more efficiently. Secondly, light-weight communication and storage secrecy is needed.
The goal of the project is to design novel schemes which provide both compression and security by using the advanced signal processing techniques. The proposed schemes will be implemented in IoT devices and a prototype will be delivered. The performance of the proposed scheme will be tested and evaluated through standard randomness tests, and energy constraints and computational requirements will be considered.
Epilepsy is a particularly challenging neurological disorder caused by relentless brain damage and aberrant rearrangements of brain wiring. In the light of the dynamic nature of the disease, the medications may stop working and neuromodulation may not completely suppress seizures. Above all, current treatments are symptomatic, while healing the epileptic brain remains unsolved so far.
This project is born from the desire to establish a new paradigm. The goal is to drive self-repair of dysfunctional brain circuits via by the symbiotic integration of bioengineered brain tissue, neuromorphic microelectronics and artificial intelligence.
Our team is taking care of the design, implementation and testing of a low power chip for brain signal recording and processing. The chip’s task is to control the way that the graft tissue develops and interacts with the host. The chip will be neuromorphic, meaning that it will mimic the behaviour of biological neural networks and resemble their structure. The neurons will be implemented through CMOS connected by memristors-based synapses.
Project title: Low-power hardware implementation of Spiking Neural Networks (SNNs) for implantable neuromorphic chips
PhD student: Margherita Ronchini
Project period: April 2019 to March 2022
Main supervisor: Assoc. Prof. Farshad Moradi
Co-supervisor: Prof. (Docent) Preben Kidmose
The operation, surveillance and maintenance of high voltage transmission lines in power systems have been a manual and costly process for decades. The process incurs not only operational expenditures but also human capital risks. With the adventure of the Internet of Things (IoT) and drone technology, this surveillance operation can be automated through the collaboration of autonomous systems. This PhD project aims to research the network structure and its protocol architecture of autonomous drones (swarm). Collaboration strategies for drones will be implemented that permit safe and efficient surveillance operation of transmission lines in the power grid.
Analytical and numerical simulations to validate proposed solutions will be considered. Promising solutions obtained from the analytical framework are expected to be prototyped and tested in a laboratory environment. There will be access (drones4energy.dk) to drone equipment and a drone test facility for piloting and field trials.
This project aims at exploring, researching and developing a novel embeddable, long life-time, ultra-low power, high resolution sensing principle for structural health monitoring (SHM), mainly in reinforced concrete (RC). The objective of the sensor is to output the corrosion rate wirelessly to a remote device, e.g. a computer, where the data is available for the owners of the structure. Corrosion damages reduce the service life of structures and can create serious safety hazards leading to fatal consequences. Today, in corrosion measurements, only very indicative and error-prone sensors exist which are based on technology developed half a century ago with slow processing, hig power usage and large formfactor. They are therefore not an attractive commodity for owners to implement. Thus, this project will challenge this status quo.
In accordance with recent years’ increased development of Wireless Sensor Networks (WSN) and Internet of Things (Iot), utilising such platforms will have tremendous benefits, such as significant lower installation cost and direct real-time access to data. This makes it possible to centralise, process and analyse a large amount of data collected from remotely placed structures, for reliable data interpretation that complies with international standards. This can then be used to launch in-time repair and maintenance operations. By further utilising energy harvesting (EH) from omnipresent sources, as the corrosion process itself, power requirements for the sensor node will be met.
Weather forecasting is an extremely important part of modern society and plays a vital role in many real-world applications. Modern weather forecasting relies on a combination of complex numerical computer models, observations of the atmosphere and pattern-recognition by meteorologists with an exceptional knowledge of the physics underlying atmospheric processes. The numerical weather prediction (NWP) models use supercomputers to create extremely large simulations of the atmosphere and its evolution in the future based on our best understanding of physics and fluid dynamics. Due to an incomplete theoretical knowledge of these processes and the inherent chaotic nature of the atmosphere, learning latent representations of the atmospheric processes using unsupervised learning can potentially extend our understanding beyond physics and climatology to improve upon the forecast accuracy of the NWP models.
This project is written in collaboration with Danske Commodities A/S and will focus on weather forecasting using deep learning approaches on time-series satellite images, applying state-of-the-art architectures within Generative Adversarial Models and/or convolutional- and recurrent neural networks to generate future sequences of satellite images. From these images, several important meteorological variables can be extracted as features which can be combined with other types of atmospheric measurements to perform supervised learning tasks in relation to weather forecasting.
Recent advances in Machine Learning have enabled us to target and successfully solve many challenging problems, most notably problems related to computer vision applications including image/scene recognition, object detection and recognition and human action localisation and recognition.
However, the current state of the art solutions based on deep neural networks require heavy computations, high memory footprint and long training processes. These requirements are restrictive in many real-life application scenarios like when they are applied in Unmanned Aerial Vehicles – UAVs (e.g. drones).
This project will research new techniques and methodologies for reducing the computational cost of deep neural network architectures (Convolutional Neural Networks and Recurrent Neural Networks). We will focus on the proposal of novel techniques for creating compact network topologies that can achieve the same (or better) performance with the state of the art. Development of the proposed approaches in UAVs and testing in real application scenarios will also lead to exciting research directions for improving existing technology.
The wellbeing of occupants in the built environment has become a rising issue in today’s society. Primary user concerns in large-scale public buildings can be described as:
The widely deployed digital process, BIM (Building Information Modelling), contains comprehensive information about buildings' design features. BIM adopts an object-oriented approach and represents building objects (storey, door, air duct, furniture, space, etc.) at a high abstraction level, instantiated by location, type, geometry and properties.
Although BIM investigates buildings’ pre-occupancy performances thoroughly, it fails to examine the “spatial artefacts” induced by people living in them. Namely, people’s lines-of-sight, physical and psychological reaction to the surroundings, and frequent activities which can vary greatly from a child to an adult, and to a disabled person with functional impairing.
The project outcome is expected to promote a holistic, integrated and human-centered design by constantly checking descriptive requirements throughout the building life-cycle. The AEC (Architecture, Engineering, Construction) industry will then have the opportunity to benefit from an intuitive, integrated and intelligible tool for validation of qualitative and numerical data.
Project title: An AI-based system for health and safety constraint checking in large public buildings
PhD student: Beidi Li
Project period: Dec 2018 to Nov 2021
Main supervisor: Prof. Peter Gorm Larsen
Co-supervisor: Assistant Prof. Carl Peter Leslie Schultz
The strict requirements for 5G and the increasing amount of data generated by both end-user and IoT devices present a set of challenges for communication networks and storage systems.
The currently employed infrastructure is unable to handle this load increase, especially with the massive increase of data. Furthermore, with the expected increase of active devices, the cost of transferring data through the current networks to the destination will be problematic. The economical cost of both maintaining and providing networks and storage systems with the ability to handle this will grow significantly. Therefore, it is critical to develop new technologies to handle the increased data load.
The goal of the project is to design technologies and architectures for future storage and communication systems which can handle the increased data loads, using compression to decrease the amount of data actually stored without loss of information.