Volume 73, Issue 10 - Special Issue: Selected contributions from the workshops “Computational Intelligence” in 2023 and 2024; Guest Editors: Ralf Mikut, Andreas Kroll and Horst Schulte

Data-driven state space models allow accurate descriptions of nonlinear systems, as they represent a general and abstract modeling approach. With the help of the model structure according to the canonical controllable form, it is possible to determine the main nonlinear influences on the process and thus, carry out learning-based categorization. Within this contribution, we explain and test categorization with the help of the Local Model State Space Network and show how conclusions about the underlying process can be drawn. Especially for time-variant behavior, the method provides the possibility to categorize the process in terms of its nonlinear characteristics to enable fault diagnosis.

Oblique model trees are nonparametric regression models that partition the input space using axis-oblique splits to define local regions for regression. The Hierarchical Local Model Tree algorithm (HILOMOT) efficiently constructs these trees and produces a continuous output from overlapping local models, making it well suited for industrial applications. We enhance HILOMOT with a novel semi-greedy splitting method to reduce the short-sightedness of its greedy splitting, eliminate an undesirable trade-off during tree construction that can cause underfitting, and improve split interpretability. Our approach first estimates the strongest direction of curvature in a partition that the existing model cannot approximate. Then, a non-greedy method determines the optimal split orthogonal to this direction. We compare our method in experiments on synthetic data against HILOMOT and other established regression techniques. Qualitative results confirm that our method produces more interpretable splits, while quantitative results show it generally constructs more accurate and compact trees. Regardless of the splitting method, HILOMOT consistently outperforms all reference methods in accuracy.

In this paper, a model-reference control scheme for wind turbine systems in the Takagi-Sugeno framework is proposed. This work deals with the model reference control approach to provide wind turbines’ fast frequency response (FFR) in fluctuating wind conditions. The proposed scheme is applied to a generic wind turbine model for the partial and full-load region. The controller coefficients are calculated using a set of proposed Linear Matrix Inequality (LMI) conditions. The design process is described in detail, including a controller performance in the wind turbine’s partial and full load region.

Renewable energies are becoming increasingly vital for electrical grid stability as conventional plants are being displaced, reducing their role in redispatch interventions. To incorporate Wind Power (WP) in redispatch planning, day-ahead forecasts are required to assess availability. Automated, scalable forecasting models are necessary for deployment across thousands of onshore WP turbines. However, irregular redispatch shutdowns complicate WP forecasting, as autoregressive methods use past generation data. This paper analyzes state-of-the-art forecasting methods with both regular and irregular shutdowns. Specifically, it compares three autoregressive Deep Learning (DL) methods with WP curve modeling, finding the latter has lower forecasting errors, fewer data cleaning requirements, and higher computational efficiency, suggesting its advantages for practical use.

The size of the training data set significantly impacts the quality of data-driven models. However, creating the data set is also one of the most significant expenses in the model development process. This is especially true for metamodels designed for structural optimization. Second, the data must be generated using computationally intensive finite element simulations. Therefore, it is essential to understand how the amount of available data influences the balance between computing costs and expected model quality to get a good trade-off. This relationship will be analyzed for a metamodel approach based on an autoencoder applied to a forming process. By applying appropriate methods of instance selection, also known as subset selection, the required data can be strongly reduced, and computational costs can be minimized.

Zusammenfassung Der industrielle Echtzeiteinsatz des zerstörungsfreien Prüfverfahrens der mikromagnetischen Materialcharakterisierung bietet ein großes Potenzial für die Analyse oberflächennaher Werkstückeigenschaften. Neben einer möglichen Zeit- und Kostenreduktion durch die direkte Integration von Prüfverfahren in den Bearbeitungsprozess, die bisher nur offline durchführbar war, ermöglicht dies eine Überwachung und gezielte Regelung relevanter Werkstückeigenschaften im laufenden Prozess. Eine zentrale Herausforderung beim Einsatz mikromagnetischer Materialcharakterisierungsverfahren (MMV) ist dabei auch auf Grund kleiner Referenzdatensätze die Kalibriermodellerstellung. Die MMV können beispielsweise bei der Prozessdatenerfassung für die Prozessparameteroptimierung in der Fertigung eingesetzt werden, ohne dabei aufwendige ex-process Referenzmessungen durchführen zu müssen. In diesem Beitrag wird ein datengetriebener Ansatz zur Kalibriermodellerstellung für die mikromagnetische Materialcharakterisierung vorgestellt, bei dem Gaußprozessregression (GPR) mit verschiedenen Verfahren zur Merkmalsselektion kombiniert wird. Die vorgestellte Methodik wird anhand eines industriellen Hartdrehprozesses evaluiert und im Hinblick auf ihre Leistungsfähigkeit gegenüber klassischen Verfahren wie multipler linearer Regression (MLR) bewertet.

Ultra-high performance concrete (UHPC) surpasses conventional concrete in performance. However, ensuring consistent mechanical properties during production, even with identical recipes, remains challenging. Using experimental data, this study investigates how material quality, environmental conditions, measurement errors in material dosing, and mixing and curing conditions influence the mechanical properties of UHPC. This broad scope of influencing factors and production conditions increases data dimensionality and, coupled with the high cost of UHPC experiments, results in a sparse dataset. Traditional evolutionary algorithms, though effective in feature selection, struggle with high-dimensional small-sized datasets. To address this, a search-space-constraining method for the non-dominated sorting genetic algorithm II (NSGA-II) is introduced, incorporating domain-specific knowledge into population initialization to reduce dimensionality and thus enhance prediction accuracy and solution stability. Comparative evaluations using various machine learning algorithms on the UHPC dataset demonstrate that population initialization to constrain the search space of NSGA-II outperforms the standard NSGA-II. Finally, the significance of each examined factor in the UHPC manufacturing process for the properties of the final product is discussed.

Microfluidic Live-Cell Imaging (MLCI) yields data on microbial cell factories. However, continuous acquisition is challenging as high-throughput experiments often lack real-time insights, delaying responses to stochastic events. We introduce three components in the Experiment Automation Pipeline for Event-Driven Microscopy to Smart Microfluidic Single-Cell Analysis (EAP4EMSIG): a fast, accurate Multi-Layer Perceptron (MLP)-based autofocusing method predicting the focus offset, an evaluation of real-time segmentation methods and a real-time data analysis dashboard. Our MLP-based autofocusing achieves a Mean Absolute Error (MAE) of 0.105 µm with inference times of 87 ms. Among eleven evaluated Deep Learning (DL) segmentation methods, Cellpose reached a Panoptic Quality (PQ) of 93.36 %, while a distance-based method was fastest (121 ms, Panoptic Quality 93.02 %).