68-XX COMPUTER SCIENCE (For papers involving machine computations and programs in a specific mathematical area, see Section {04 in that areag 68-00 General reference works (handbooks, dictionaries, bibliographies, etc.)
Refine
Document Type
- Doctoral Thesis (6)
- Article (3)
- Preprint (2)
Language
- English (11)
Has Fulltext
- yes (11)
Keywords
- Benchmark (2)
- Black-Box (2)
- Dataset (2)
- Dynamics (2)
- Industrial Robotics (2)
- Nonlinear Identification (2)
- Scientific Visualization (2)
- System Identification (2)
- Application Framework (1)
- CWGAN (1)
Faculty / Organisational entity
In the following an introduction to the level set method will be givenso that one becomes aware of the arising problems, which lead to the needof reinitialization. The problems concerning reinitialization itself will be analysed more detailed and a solution for area loss will be proposed. This solution consists in a combination of the commonly used PDE for reinitialization and extrapolation around the zero level set. Numericalexperiments show rather satisfactory results as far as area loss and computation of curvature are concerned.
In this article we present a method to extend high order finite volume schemes
to networks of hyperbolic conservation laws with algebraic coupling conditions. This method is based on an ADER approach in time to solve the
generalized Riemann problem at the junction. Additionally to the high order accuracy, this approach maintains an exact conservation of quantities if
stated by the coupling conditions. Several numerical examples confirm the
benefits of a high order coupling procedure for high order accuracy and stable
shock capturing.
Planar force or pressure is a fundamental physical aspect during any people-vs-people and people-vs-environment activities and interactions. It is as significant as the more established linear and angular acceleration (usually acquired by inertial measurement units). There have been several studies involving planar pressure in the discipline of activity recognition, as reviewed in the first chapter. These studies have shown that planar pressure is a promising sensing modality for activity recognition. However, they still take a niche part in the entire discipline, using ad hoc systems and data analysis methods. Mostly these studies were not followed by further elaborative works. The situation calls for a general framework that can help push planar pressure sensing into the mainstream.
This dissertation systematically investigates using planar pressure distribution sensing technology for ubiquitous and wearable activity recognition purposes. We propose a generic Textile Pressure Mapping (TPM) Framework, which encapsulates (1) design knowledge and guidelines, (2) a multi-layered tool including hardware, software and algorithms, and (3) an ensemble of empirical study examples. Through validation with various empirical studies, the unified TPM framework covers the full scope of application recognition, including the ambient, object, and wearable subspaces.
The hardware part constructs a general architecture and implementations in the large-scale and mobile directions separately. The software toolkit consists of four heterogeneous tiers: driver, data processing, machine learning, visualization/feedback. The algorithm chapter describes generic data processing techniques and a unified TPM feature set. The TPM framework offers a universal solution for other researchers and developers to evaluate TPM sensing modality in their application scenarios.
The significant findings from the empirical studies have shown that TPM is a versatile sensing modality. Specifically, in the ambient subspace, a sports mat or carpet with TPM sensors embedded underneath can distinguish different sports activities or different people's gait based on the dynamic change of body-print; a pressure sensitive tablecloth can detect various dining actions by the force propagated from the cutlery through the plates to the tabletop. In the object subspace, swirl office chairs with TPM sensors under the cover can be used to detect the seater's real-time posture; TPM can be used to detect emotion-related touch interactions for smart objects, toys or robots. In the wearable subspace, TPM sensors can be used to perform pressure-based mechanomyography to detect muscle and body movement; it can also be tailored to cover the surface of a soccer shoe to distinguish different kicking angles and intensities.
All the empirical evaluations have resulted in accuracies well-above the chance level of the corresponding number of classes, e.g., the `swirl chair' study has classification accuracy of 79.5% out of 10 posture classes and in the `soccer shoe' study the accuracy is 98.8% among 17 combinations of angle and intensity.
Visualization is vital to the scientific discovery process.
An interactive high-fidelity rendering provides accelerated insight into complex structures, models and relationships.
However, the efficient mapping of visualization tasks to high performance architectures is often difficult, being subject to a challenging mixture of hardware and software architectural complexities in combination with domain-specific hurdles.
These difficulties are often exacerbated on heterogeneous architectures.
In this thesis, a variety of ray casting-based techniques are developed and investigated with respect to a more efficient usage of heterogeneous HPC systems for distributed visualization, addressing challenges in mesh-free rendering, in-situ compression, task-based workload formulation, and remote visualization at large scale.
A novel direct raytracing scheme for on-the-fly free surface reconstruction of particle-based simulations using an extended anisoptropic kernel model is investigated on different state-of-the-art cluster setups.
The versatile system renders up to 170 million particles on 32 distributed compute nodes at close to interactive frame rates at 4K resolution with ambient occlusion.
To address the widening gap between high computational throughput and prohibitively slow I/O subsystems, in situ topological contour tree analysis is combined with a compact image-based data representation to provide an effective and easy-to-control trade-off between storage overhead and visualization fidelity.
Experiments show significant reductions in storage requirements, while preserving flexibility for exploration and analysis.
Driven by an increasingly heterogeneous system landscape, a flexible distributed direct volume rendering and hybrid compositing framework is presented.
Based on a task-based dynamic runtime environment, it enables adaptable performance-oriented deployment on various platform configurations.
Comprehensive benchmarks with respect to task granularity and scaling are conducted to verify the characteristics and potential of the novel task-based system design.
A core challenge of HPC visualization is the physical separation of visualization resources and end-users.
Using more tiles than previously thought reasonable, a distributed, low-latency multi-tile streaming system is demonstrated, being able to sustain a stable 80 Hz when streaming up to 256 synchronized 3840x2160 tiles and achieve 365 Hz at 3840x2160 for sort-first compositing over the internet, thereby enabling lightweight visualization clients and leaving all the heavy lifting to the remote supercomputer.
The simulation of physical phenomena involving the dynamic behavior of fluids and gases
has numerous applications in various fields of science and engineering. Of particular interest
is the material transport behavior, the tendency of a flow field to displace parts of the
medium. Therefore, many visualization techniques rely on particle trajectories.
Lagrangian Flow Field Representation. In typical Eulerian settings, trajectories are
computed from the simulation output using numerical integration schemes. Accuracy concerns
arise because, due to limitations of storage space and bandwidth, often only a fraction
of the computed simulation time steps are available. Prior work has shown empirically that
a Lagrangian, trajectory-based representation can improve accuracy [Agr+14]. Determining
the parameters of such a representation in advance is difficult; a relationship between the
temporal and spatial resolution and the accuracy of resulting trajectories needs to be established.
We provide an error measure for upper bounds of the error of individual trajectories.
We show how areas at risk for high errors can be identified, thereby making it possible to
prioritize areas in time and space to allocate scarce storage resources.
Comparative Visual Analysis of Flow Field Ensembles. Independent of the representation,
errors of the simulation itself are often caused by inaccurate initial conditions,
limitations of the chosen simulation model, and numerical errors. To gain a better understanding
of the possible outcomes, multiple simulation runs can be calculated, resulting in
sets of simulation output referred to as ensembles. Of particular interest when studying the
material transport behavior of ensembles is the identification of areas where the simulation
runs agree or disagree. We introduce and evaluate an interactive method that enables application
scientists to reliably identify and examine regions of agreement and disagreement,
while taking into account the local transport behavior within individual simulation runs.
Particle-Based Representation and Visualization of Uncertain Flow Data Sets. Unlike
simulation ensembles, where uncertainty of the solution appears in the form of different
simulation runs, moment-based Eulerian multi-phase fluid simulations are probabilistic in
nature. These simulations, used in process engineering to simulate the behavior of bubbles in
liquid media, are aimed toward reducing the need for real-world experiments. The locations
of individual bubbles are not modeled explicitly, but stochastically through the properties of
locally defined bubble populations. Comparisons between simulation results and physical
experiments are difficult. We describe and analyze an approach that generates representative
sets of bubbles for moment-based simulation data. Using our approach, application scientists
can directly, visually compare simulation results and physical experiments.
In the avionics domain, “ultra-reliability” refers to the practice of ensuring quantifiably negligible residual failure rates in the presence of transient and permanent hardware faults. If autonomous Cyber- Physical Systems (CPS) in other domains, e.g., autonomous vehicles, drones, and industrial automation systems, are to permeate our everyday life in the not so distant future, then they also need to become ultra-reliable. However, the rigorous reliability engineering and analysis practices used in the avionics domain are expensive and time consuming, and cannot be transferred to most other CPS domains. The increasing adoption of faster and cheaper, but less reliable, Commercial Off-The-Shelf (COTS) hardware is also an impediment in this regard.
Motivated by the goal of ultra-reliable CPS, this dissertation shows how to soundly analyze the reliability of COTS-based implementations of actively replicated Networked Control Systems (NCSs)—which are key building blocks of modern CPS—in the presence of transient hard- ware faults. When an NCS is deployed over field buses such as the Controller Area Network (CAN), transient faults are known to cause host crashes, network retransmissions, and incorrect computations. In addition, when an NCS is deployed over point-to-point networks such as Ethernet, even Byzantine errors (i.e., inconsistent broadcast transmissions) are possible. The analyses proposed in this dissertation account for NCS failures due to each of these error categories, and consider NCS failures in both time and value domains. The analyses are also provably free of reliability anomalies. Such anomalies are problematic because they can result in unsound failure rate estimates, which might lead us to believe that a system is safer than it actually is.
Specifically, this dissertation makes four main contributions. (1) To reduce the failure rate of NCSs in the presence of Byzantine errors, we present a hard real-time design of a Byzantine Fault Tolerance (BFT) protocol for Ethernet-based systems. (2) We then propose a quantitative reliability analysis of the presented design in the presence of transient faults. (3) Next, we propose a similar analysis to upper-bound the failure probability of an actively replicated CAN-based NCS. (4) Finally, to upper-bound the long-term failure rate of the NCS more accurately, we propose analyses that take into account the temporal robustness properties of an NCS expressed as weakly-hard constraints.
By design, our analyses can be applied in the context of full-system analyses. For instance, to certify a system consisting of multiple actively replicated NCSs deployed over a BFT atomic broadcast layer, the upper bounds on the failure rates of each NCS and the atomic broadcast layer can be composed using the sum-of-failure-rates model.
A significant step to engineering design is to take into account uncertainties and to
develop optimal designs that are robust with respect to perturbations. Furthermore, it
is often of interest to optimize for different conflicting objective functions describing the
quality of a design, leading to a multi-objective optimization problem. In this context,
generating methods for solving multi-objective optimization problems seek to find a
representative set of solutions fulfilling the concept of Pareto optimality. When multiple
uncertain objective functions are involved, it is essential to define suitable measures for
robustness that account for a combined effect of uncertainties in objective space. Many
tasks in engineering design include the solution of an underlying partial differential
equation that can be computationally expensive. Thus, it is of interest to use efficient
strategies for finding optimal designs. This research aims to present suitable measures
for robustness in a multi-objective context, as well as optimization strategies for multi-
objective robust design.
This work introduces new ideas for robustness measures in the context of multi-
objective robust design. Losses and expected losses based on distances in objective space
are used to describe robustness. A direct formulation and a two-phase formulation based
on expected losses are proposed for finding a set of robust optimal solutions.
Furthermore, suitable optimization strategies for solving the resulting multi-objective
robust design problem are formulated and analyzed. The multi-objective optimization
problem is solved with a constraint-based approach that is based on solving several
constrained single-objective optimization problems with a hybrid optimization strategy.
The hybrid method combines a global search method on a surrogate model with adjoint-
based optimization methods. In the context of optimization with an underlying partial
differential equation, a one-shot approach is extended to handle additional constraints.
The developed concepts for multi-objective robust design and the proposed optimiza-
tion strategies are applied to an aerodynamic shape optimization problem. The drag
coefficient and the lift coefficient are optimized under the consideration of uncertain-
ties in the operational conditions and geometrical uncertainties. The uncertainties are
propagated with the help of a non-intrusive polynomial chaos approach. For increasing
the efficiency when considering a higher-dimensional random space, it is made use of a
Karhunen-Loève expansion and a dimension-adaptive sparse grid quadrature.
Deep learning has achieved significant improvements in a variety of tasks in computer vision applications with an open image dataset which has a large amount of data. However, the acquisition of a large number of the dataset is a challenge in real-world applications, especially if they are new eras for deep learning. Furthermore, the distribution of class in the dataset is often imbalanced. The data imbalance problem is frequently bottlenecks of the neural network performance in classification. Recently, the potential of generative adversarial networks (GAN) as a data augmentation method on minority data has been studied.
This dissertation investigates using GAN and transfer learning to improve the performance of the classification under imbalanced data conditions. We first propose a classification enhancement generative adversarial networks (CEGAN) to enhance the quality of generated synthetic minority data and more importantly, to improve the prediction accuracy in data imbalanced condition. Our experiments show that approximating the real data distribution using CEGAN improves the classification performance significantly in data imbalanced conditions compared with various standard data augmentation methods.
To further improve the performance of the classification, we propose a novel supervised discriminative feature generation method (DFG) for minority class dataset. DFG is based on the modified structure of Generative Adversarial Network consisting of four independent networks: generator, discriminator, feature extractor, and classifier. To augment the selected discriminative features of minority class data by adopting attention mechanism, the generator for class-imbalanced target task is trained while feature extractor and classifier are regularized with the pre-trained ones from large source data. The experimental results show that the generator of DFG enhances the augmentation of label-preserved and diverse features, and classification results are significantly improved on the target task.
In this thesis, these proposals are deployed to bearing fault detection and diagnosis of induction motor and shipping label recognition and validation for logistics. The experimental results for bearing fault detection and diagnosis conclude that the proposed GAN-based framework has good performance on the imbalanced fault diagnosis of rotating machinery. The experimental results for shipping label recognition and validation also show that the proposed method achieves better performance than many classical and state-of-the-art algorithms.
We present an identification benchmark data set for a full robot movement with an KUKA KR300 R2500 ultra SE industrial robot. It is a robot with a nominal payload capacity of 300 kg, a weight of 1120 kg and a reach of 2500mm. It exhibits 12 states accounting for position and velocity for each of the 6 joints. The robot encounters backlash in all joints, pose-dependent inertia, pose-dependent gravitational loads, pose-dependent hydraulic forces, pose- and velocity dependent centripetal and Coriolis forces as well as a nonlinear friction, which is temperature dependent and therefore potentially time varying. We supply the prepared dataset for black-box identification of the forward or the inverse robot dynamics. Additional to the data for black-box modelling, we supply high-frequency raw data and videos of each experiment. A baseline and figures of merit are defined to make results compareable across different identification methods.
Load modeling is one of the crucial tasks for improving smart grids’ energy efficiency. Among many alternatives, machine learning-based load models have become popular in applications and have shown outstanding performance in recent years. The performance of these models highly relies on data quality and quantity available for training. However, gathering a sufficient amount of high-quality data is time-consuming and extremely expensive. In the last decade, Generative Adversarial Networks (GANs) have demonstrated their potential to solve the data shortage problem by generating synthetic data by learning from recorded/empirical data. Educated synthetic datasets can reduce prediction error of electricity consumption when combined with empirical data. Further, they can be used to enhance risk management calculations. Therefore, we propose RCGAN, TimeGAN, CWGAN, and RCWGAN which take individual electricity consumption data as input to provide synthetic data in this study. Our work focuses on one dimensional times series, and numerical experiments on an empirical dataset show that GANs are indeed able to generate synthetic data with realistic appearance.