Refine
Year of publication
- 2022 (97) (remove)
Document Type
- Conference Proceeding (97) (remove)
Conference Type
- Konferenzartikel (82)
- Konferenz-Abstract (11)
- Konferenz-Poster (2)
- Sonstiges (2)
Language
- English (97) (remove)
Is part of the Bibliography
- yes (97)
Keywords
- injury (10)
- Machine Learning (5)
- biomechanics (5)
- running (5)
- ACL (4)
- Robustness (4)
- Radar (3)
- RoboCup (3)
- sport (3)
- 3D printing (2)
Institute
- Fakultät Elektrotechnik, Medizintechnik und Informatik (EMI) (ab 04/2019) (40)
- Fakultät Maschinenbau und Verfahrenstechnik (M+V) (26)
- Fakultät Wirtschaft (W) (20)
- Fakultät Medien (M) (ab 22.04.2021) (13)
- INES - Institut für nachhaltige Energiesysteme (12)
- ivESK - Institut für verlässliche Embedded Systems und Kommunikationselektronik (11)
- IMLA - Institute for Machine Learning and Analytics (10)
- ACI - Affective and Cognitive Institute (3)
- IUAS - Institute for Unmanned Aerial Systems (3)
- POIM - Peter Osypka Institute of Medical Engineering (2)
Open Access
- Open Access (47)
- Closed (41)
- Bronze (26)
- Diamond (11)
- Closed Access (9)
- Grün (6)
- Hybrid (2)
- Gold (1)
Seismic data has often missing traces due to technical acquisition or economical constraints. A compete dataset is crucial in several processing and inversion techniques. Deep learning algorithms, based on convolutional neural networks (CNNs), have shown alternative solutions that overcome limitation of traditional interpolation methods e.g. data regularity, linearity assumption, etc. There are two different paradigms of CNN methods for seismic interpolation. The first one, so-called deep prior interpolation (DPI), trains a CNN to map random noise to a complete seismic image using only the decimated image itself. The second one, referred as standard deep learning method, trains a CNN to map a decimated seismic image into a complete one using a dataset of complete and artificially decimated images. Within this research, we systematically compare the performance of both methods for different quantities of regular and irregular missing traces using 4 datasets. We evaluate the results of both methods using 5 well-known metrics. We found that DPI method performs better than the standard method if the percentage of missing traces is low (10%) and otherwise if the level of decimation is high (50%).
The desire to connect more and more devices and to make them more intelligent and more reliable, is driving the needs for the Internet of Things more than ever. Such IoT edge systems require sound security measures against cyber-attacks, since they are interconnected, spatially distributed, and operational for an extended period of time. One of the most important requirements for the security in many industrial IoT applications is the authentication of the devices. In this paper, we present a mutual authentication protocol based on Physical Unclonable Functions, where challenge-response pairs are used for both device and server authentication. Moreover, a session key can be derived by the protocol in order to secure the communication channel. We show that our protocol is secure against machine learning, replay, man-in-the-middle, cloning, and physical attacks. Moreover, it is shown that the protocol benefits from a smaller computational, communication, storage, and hardware overhead, compared to similar works.
This paper presents a method for supporting the application of Additive Tooling (AT)-based validation environments in integrated product development. Based on a case study, relevant process steps, activities and possible barriers in the realisation of an injection-moulded product are identified and analysed. The aim of the method is to support the target-oriented application of Additive Tooling to obtain physical prototypes at an early stage and to shorten validation cycles.
In recent years, the topic of embedded machine learning has become very popular in AI research. With the help of various compression techniques such as pruning, quantization and others compression techniques, it became possible to run neural networks on embedded devices. These techniques have opened up a whole new application area for machine learning. They range from smart products such as voice assistants to smart sensors that are needed in robotics. Despite the achievements in embedded machine learning, efficient algorithms for training neural networks in constrained domains are still lacking. Training on embedded devices will open up further fields of applications. Efficient training algorithms would enable federated learning on embedded devices, in which the data remains where it was collected, or retraining of neural networks in different domains. In this paper, we summarize techniques that make training on embedded devices possible. We first describe the need and requirements for such algorithms. Then we examine existing techniques that address training in resource-constrained environments as well as techniques that are also suitable for training on embedded devices, such as incremental learning. At the end, we also discuss which problems and open questions still need to be solved in these areas.
Voice user interfaces (VUIs) offer an intuitive, fast and convenient way for humans to interact with machines and computers. Yet, whether they’ll be truly successful and find widespread uptake in the near future depends on the user experience (UX) they offer. With this survey-based study (n = 108), we aim to identify the major annoyances German voice assistant users are facing in voice-driven human-computer interactions. The results of our questionnaire show that irritations appear in six categories: privacy issues, unwanted activation, comprehensibility, response quality, conversational design and voice characteristics. Our findings can help identify key areas of work to optimize voice user experience in order to achieve greater adaptation of the technology. In addition, they can provide valuable information for the further development and standardization of voice user experience (VUX) research.
Subspace clustering aims to find all clusters in all subspaces of a high-dimensional data space. We present a massively data-parallel approach that can be run on graphics processing units. It extends a previous density-based method that scales well with the number of dimensions. Its main computational bottleneck consists of (sequentially) generating a large number of minimal cluster candidates in each dimension and using hash collisions in order to find matches of such candidates across multiple dimensions. Our approach parallelizes this process by removing previous interdependencies between consecutive steps in the sequential generation process and by applying a very efficient parallel hashing scheme optimized for GPUs. This massive parallelization gives up to 70x speedup for
the bottleneck computation when it is replaced by our approach and run on current GPU hardware. We note that depending on data size and choice of parameters, the parallelized part of the algorithm can take different percentages of the overall runtime of the clustering process, and thus, the overall clustering speedup may vary significantly between different cases. However, even
in our ”worst-case” test, a small dataset where the computation makes up only a small fraction of the overall clustering time, our parallel approach still yields a speedup of more than 3x for the complete run of the clustering process. Our method could also be combined with parallelization of other parts of the clustering algorithm, with an even higher potential gain in processing speed.
Deep learning models are intrinsically sensitive to distribution shifts in the input data. In particular, small, barely perceivable perturbations to the input data can force models to make wrong predictions with high confidence. An common defense mechanism is regularization through adversarial training which injects worst-case perturbations back into training to strengthen the decision boundaries, and to reduce overfitting. In this context, we perform an investigation of 3 × 3 convolution filters that form in adversarially- trained models. Filters are extracted from 71 public models of the ℓ ∞ -RobustBench CIFAR-10/100 and ImageNet1k leaderboard and compared to filters extracted from models built on the same architectures but trained without robust regularization. We observe that adversarially-robust models appear to form more diverse, less sparse, and more orthogonal convolution filters than their normal counterparts. The largest differences between robust and normal models are found in the deepest layers, and the very first convolution layer, which consistently and predominantly forms filters that can partially eliminate perturbations, irrespective of the architecture.
Many commonly well-performing convolutional neural network models have shown to be susceptible to input data perturbations, indicating a low model robustness. Adversarial attacks are thereby specifically optimized to reveal model weaknesses, by generating small, barely perceivable image perturbations that flip the model prediction. Robustness against attacks can be gained for example by using adversarial examples during training, which effectively reduces the measurable model attackability. In contrast, research on analyzing the source of a model’s vulnerability is scarce. In this paper, we analyze adversarially trained, robust models in the context of a specifically suspicious network operation, the downsampling layer, and provide evidence that robust models have learned to downsample more accurately and suffer significantly less from aliasing than baseline models.
Solar energy plays a central role in the energy transition. Clouds generate locally large fluctuations in the generation output of photovoltaic systems, which is a major problem for energy systems such as microgrids, among others. For an optimal design of a power system, this work analyzed the variability using a spatially distributed sensor network at Stuttgart Airport. It has been shown that the spatial distribution partially reduces the variability of solar radiation. A tool was also developed to estimate the output power of photovoltaic systems using irradiation time series and assumptions about the photovoltaic sites. For days with high fluctuations of the estimated photovoltaic power, different energy system scenarios were investigated. It was found the approach can be used to have a more realistic representation of aggregated PV power taking spatial smoothing into account and that the resulting PV power generation profiles provide a good basis for energy system design considerations like battery sizing.
We consider the local group of agents for exchanging the time-series data value and computing the approximation of the mean value of all agents. An agent represented by a node knows all local neighbor nodes in the same group. The node has the contact information of other nodes in other groups. The nodes interact with each other in synchronous rounds to exchange the updated time-series data value using the random call communication model. The amount of data exchanged between agent-based sensors in the local group network affects the accuracy of the aggregation function results. At each time step, the agent-based sensor can update the input data value and send the updated data value to the group head node. The group head node sends the updated data value to all group members in the same group. Grouping nodes in peer-to-peer networks show an improvement in Mean Squared Error (MSE).
As a university it is more and more difficult to reach all target groups equally. Common problems like information overload, numerous institutions with same focuses or multi-channel-communication make it hard to gain the attention of the target group. This paper is four-fold: we present an overview of the state of art and the importance of the study (I), based on which we highlight the approach to user experience analysis. First, we identified the irritations in the course of an expert evaluation (II) and verified them within the test, including the target groups (III). Finally, based on the results, we were able to pro-vide recommendations for action to improve the UX and to be used for the conception of an intranet (IV).
In this paper, we study the runtime performance of symmetric cryptographic algorithms on an embedded ARM Cortex-M4 platform. Symmetric cryptographic algorithms can serve to protect the integrity and optionally, if supported by the algorithm, the confidentiality of data. A broad range of well-established algorithms exists, where the different algorithms typically have different properties and come with different computational complexity. On deeply embedded systems, the overhead imposed by cryptographic operations may be significant. We execute the algorithms AES-GCM, ChaCha20-Poly1305, HMAC-SHA256, KMAC, and SipHash on an STM32 embedded microcontroller and benchmark the execution times of the algorithms as a function of the input lengths.
Biodegradable metals have entered the implant market in recent years, but still do not show fully satisfactory degradation behaviour and mechanical properties. In contrast, it has been shown that pure molybdenum has an excellent combination of the required properties in this respect. We report on PM based screen printing of thin-walled molybdenum tubes as a processing step for medical stent manufacture. We also present data on the in vivo degradation and biocompatibility of molybdenum. The degradation of molybdenum wires implanted in the aorta of rats was evaluated by SEM and EDX. Biocompatibility was assessed by histological investigation of organs and analysis of molybdenum levels in tissue extracts and body fluids. Degradation rates of up to 13.5 μm/y were observed after 12 months. No histological changes or elevated molybdenum levels in organ tissues were observed. In summary, the results further underline that molybdenum is a highly promising biodegradable metallic material.
This study aimed to compare a simplified calculation of the knee abduction moment with the traditional inverse dynamics calculation when athletes perform fake-cut maneuvers with different complexities. In the simplified calculation, we multiply the force vector with its lever arm to the knee, projected onto the local coordinate system of the proximal thigh, hence neglecting the inertial contributions from distal segments. We found very strong ranking consistency using Spearman’s rank correlation coefficient when using the simplified method compared to the traditional calculation. Independent of the tasks, the simplified method resulted in higher moments than the inverse dynamics. This was caused by ignoring the moment caused by segment linear acceleration generating a counteracting moment by about 7%. An alternative to the complex calculations of inverse dynamics can be used to investigate the contributions of the GRF magnitude and its lever arm to the knee.
Currently, many theoretical as well as practically relevant questions towards the transferability and robustness of Convolutional Neural Networks (CNNs) remain unsolved. While ongoing research efforts are engaging these problems from various angles, in most computer vision related cases these approaches can be generalized to investigations of the effects of distribution shifts in image data. In this context, we propose to study the shifts in the learned weights of trained CNN models. Here we focus on the properties of the distributions of dominantly used 3×3 convolution filter kernels. We collected and publicly provide a dataset with over 1.4 billion filters from hundreds of trained CNNs, using a wide range of datasets, architectures, and vision tasks. In a first use case of the proposed dataset, we can show highly relevant properties of many publicly available pre-trained models for practical applications: I) We analyze distribution shifts (or the lack thereof) between trained filters along different axes of meta-parameters, like visual category of the dataset, task, architecture, or layer depth. Based on these results, we conclude that model pre-training can succeed on arbitrary datasets if they meet size and variance conditions. II) We show that many pre-trained models contain degenerated filters which make them less robust and less suitable for fine-tuning on target applications. Data & Project website: https://github.com/paulgavrikov/cnn-filter-db.
Additive manufacturing with plastics enables the production of lightweight and resilient components with a high degree of design freedom. In the low-cost sector, Material Extrusion as Fused Layer Modeling (FLM) has so far been the leading method, as it offers simple 3D printers and a variety of inexpensive 3D materials. However, printing times for 6FLM are very long and dimensional accuracy and surface finish are rather poor. Recently, new processes from the field of Vat Polymerization have appeared on the market, such as masked Stereolithography (mSLA), which offer a significant improvement in component quality and build speed at equally favorable machine costs.
This paper therefore analyzes the technical and economic capabilities of the two competing additive processes. For this purpose, the achievable dimensional and surface qualities are determined using a test specimen which represents various important geometry elements. In addition, the machine and material costs are determined and compared with each other. Finally, the resulting environmental impact is determined in the form of the CO2 footprint. In order to optimize the strength of the printed components, material properties of the tensile specimens produced additively with mSLA are determined. The use of ABS-like resins will also be investigated to determine optimal processing settings.
This paper presents an extended version of a previously published Bayesian algorithm for the automatic correction of the positions of the equipment on the map with simultaneous mobile object trajectory localization (SLAM) in underground mine environment represented by undirected graph. The proposed extended SLAM algorithm requires much less preliminary data on possible equipment positions and uses an additional resample move algorithm to significantly improve the overall performance.
Synthesizing voice with the help of machine learning techniques has made rapid progress over the last years. Given the current increase in using conferencing tools for online teaching, we question just how easy (i.e. needed data, hardware, skill set) it would be to create a convincing voice fake. We analyse how much training data a participant (e.g. a student) would actually need to fake another participants voice (e.g. a professor). We provide an analysis of the existing state of the art in creating voice deep fakes and align the identified as well as our own optimization techniques in the context of two different voice data sets. A user study with more than 100 participants shows how difficult it is to identify real and fake voice (on avg. only 37% can recognize a professor’s fake voice). From a longer-term societal perspective such voice deep fakes may lead to a disbelief by default.
Robust scheduling problem is a major decision problem that is addressed in the literature, especially for remanufacturing systems; this problem is complex because of the high uncertainty and complex constraints involved. Generally, the existing approaches are dedicated to specific processes and do not enable the quick and efficient generation and evaluation of schedules. With the emergence of the Industry 4.0 paradigm, data availability is now considered an opportunity to facilitate the decision-making process. In this study, a data-driven decisionmaking process is proposed to treat the robust scheduling problem of remanufacturing systems in uncertain environments. In particular, this process generates simulation models based on a data-driven modeling approach. A robustness evaluation approach is proposed to answer several decision questions. An application of the decision process in an industrial case of a remanufacturing system is presented herein, illustrating the impact of robustness evaluation results on real-life decisions.