Niggemann, Oliver

Forecasting the behaviour of industrial robots, power grids or pandemics under changing external inputs requires accurate dynamical models that can adapt to varying signals and capture long-term effects such as delays or memory. While recent neural approaches address some of these challenges individually, their reliance on computationally intensive solvers and their black-box nature limit their practical utility. In this work, we propose Laplace-Net, a decoupled, solver-free neural framework for learning forced and delay-aware dynamical systems. It uses the Laplace transform to (i) bypass computationally intensive solvers, (ii) enable the learning of delays and memory effects and (iii) decompose each system into interpretable control-theoretic components. Laplace-Net also enhances transferability, as its modular structure allows for targeted re-training of individual components to new system setups or environments. Experimental results on eight benchmark datasets–including linear, nonlinear and delayed systems–demonstrate the method’s improved accuracy and robustness compared to state-of-the-art approaches, particularly in handling complex and previously unseen inputs.

Automating the structuring of Solution Principles within conceptual design is crucial for efficiently covering Function Structures while reducing time-intensive manual processes. Solution Principles are central in bridging functional requirements and technical implementations, yet traditional methods depend heavily on human expertise. To address this, a novel approach leveraging a search algorithm is proposed to automatically identify an optimal set of Solution Principles for a given Function Structure. The approach formalizes the problem and provides rules for the selection and application of Solution Principles. Key components include a function for applying Solution Principles to functions and a heuristic that minimizes principle selection, guiding the search toward optimal solutions. An evaluation shows the potential of this method to reduce time and effort in early product design.

This paper focuses on the development and implementation of a diagnosis toolchain, to identify faults and recommend actions for the system operators of the environmental control and life support system of the COLUMBUS module on the International Space Station. We present a comprehensive framework which uses different aspects of artificial intelligence to efficiently identify the necessary interventions for the system operator to stabilize the system in case of emergencies and defects. Methods such as machine learning and statistical analysis, based on time-series, are used for anomaly detection to identify potentially critical situations early and issue the corresponding warnings. Diagnostic functionality enables the identification of the causes of anomalies, integrating expert knowledge and pattern recognition algorithms to achieve accurate diagnostic results. The localization of affected system parts is crucial as fault propagation can obscure the origin of anomalies. A vital aspect of the AI system is determining possible reconfiguration measures according to the behavior of the system, offering operators various operational continuance variants in the event of damage. Based on the diagnostic results, the system identifies suitable reconfiguration measures to restore normal operation or minimize potential damage. An additional supervision module based on qualitative system models is then used to monitor, evaluate, and assess the effects of these interventions. An MLOps platform facilitates the seamless integration of the framework into existing processes, providing an agile solution for fast and reliable development, scaling, and standardized integration interfaces. The successful integration of the AI toolchain at Airbus Defense and Space exemplifies this implementation’s effectiveness, significantly reducing development times and enhancing the process’s reliability and efficiency.

This study presents an approach based on data-driven methods for determining the parameters needed to model time-dependent material behaviour. The time-dependent behaviour of the thermoplastic polymer polybutylene terephthalate is investigated. The material was examined under two conditions, one with and one without the inclusion of reinforcing short fibres. Two modelling approaches are proposed to represent the time-dependent response. The first approach is the generalised Maxwell model formulated through the classical exponential Prony series, and the second approach is a model based on fractional calculus. In order to quantify the comparative capabilities of both models, experimental data from tensile creep tests on fibre-reinforced polybutylene terephthalate and unreinforced polybutylene terephthalate specimens are analysed. A central contribution of this work is the implementation of a machine-learning-ready parameter identification framework that enables the automated extraction of model parameters directly from time-series data. This framework enables the robust fitting of the Prony-based model, which requires multiple characteristic times and stiffness parameters, as well as the fractional model, which achieves high accuracy with significantly fewer parameters. The fractional model benefits from a novel neural solver for fractional differential equations, which not only reduces computational complexity but also permits the interpretation of the fractional order and stiffness coefficient in terms of physical creep resistance. The methodological framework is validated through a comparative assessment of predictive performance, parameter cheapness, and interpretability of each model, thereby providing a comprehensive understanding of their applicability to long-term material behaviour modelling in polymer-based composite materials.

Achieving robust uncertainty quantification for deep neural networks represents an important requirement in many real-world applications of deep learning, such as medical imaging where it is necessary to assess the reliability of a neural network’s prediction. Bayesian neural networks are a promising approach for modeling uncertainties in deep neural networks. Unfortunately, generating samples from the posterior distribution of neural networks is a major challenge. One significant advance in that direction would be the incorporation of adaptive step sizes, similar to modern neural network optimizers, into Monte Carlo Markov chain sampling algorithms without significantly increasing computational demand. Over the past years, several papers have introduced sampling algorithms with corresponding theorems stating that they achieve this property. In this paper, we demonstrate that these methods can have a substantial bias in the distribution they sample, even in the limit of vanishing step sizes and at full batch size. Furthermore, for most of the algorithms, we show that convergence to the correct distribution can be restored with a simple fix at the cost of increasing computational demand.