Volume 67, Issue 10 - Schwerpunktheft: Ausgewählte Beiträge aus dem GMA-Fachausschuss 5.14 „Computational Intelligence“ / Ralf Mikut, Andreas Kroll, Frank Hoffmann

Zusammenfassung In diesem Beitrag werden Methoden zum Entwurf von Multisinus- und Multistufensignalen zur Identifikation nichtlinearer dynamischer Takagi-Sugeno-Fuzzy-Modelle vorgestellt. Diese Entwurfsmethoden arbeiten prozessmodellfrei. Sie werden am Beispiel der Modellierung eines elektromechanischen Stellglieds eines Verbrennungsmotors demonstriert.

The task of data reduction is discussed and a novel selection approach which allows to control the optimal point distribution of the selected data subset is proposed. The proposed approach utilizes the estimation of probability density functions (pdfs). Due to its structure, the new method is capable of selecting a subset either by approximating the pdf of the original dataset or by approximating an arbitrary, desired target pdf. The new strategy evaluates the estimated pdfs solely on the selected data points, resulting in a simple and efficient algorithm with low computational and memory demand. The performance of the new approach is investigated for two different scenarios. For representative subset selection of a dataset, the new approach is compared to a recently proposed, more complex method and shows comparable results. For the demonstration of the capability of matching a target pdf, a uniform distribution is chosen as an example. Here the new method is compared to strategies for space-filling design of experiments and shows convincing results.

The accuracy of many regression models suffers from inhomogeneous data coverage. Models loose accuracy because they are unable to locally adapt the model complexity. This article develops and evaluates an automated design process for the generation of hybrid regression models from arbitrary submodels. For the first time, these submodels are weighted by a One-Class Support Vector Machine, taking local data coverage into account. Compared to reference regression models, the newly developed hybrid models achieve significant better results in nine out of ten benchmark datasets. To enable straightforward usage in data science, an implementation is integrated in the open source MATLAB toolbox SciXMiner.

Industrial applications put special demands on machine learning algorithms. Noisy data, outliers, and sensor faults present an immense challenge for learners. A considerable part of machine learning research focuses on the selection of relevant, non-redundant features. This contribution details an approach to group and fuse redundant features prior to learning and classification. Features are grouped relying on a correlation-based redundancy measure. The fusion of features is guided by determining the majority observation based on possibility distributions. Furthermore, this paper studies the effects of feature fusion on the robustness and performance of classification with a focus on industrial applications. The approach is statistically evaluated on public datasets in comparison to classification on selected features only.

Zusammenfassung Der vorliegende Beitrag untersucht die Erkennung benachbarter Fahrstreifen auf Grundlage von Kamerabildern. Hierbei wird sowohl die Anzahl befahrbarer Fahrstreifen als auch deren Verlauf innerhalb eines festgelegten Bereichs vor dem Fahrzeug bildbasiert geschätzt. Die Erkennung erfolgt durch Convolutional Neural Networks. Der Beitrag bewertet die Güte der Schätzung und Genauigkeit der Fahrstreifenerkennung. Zur Interpretation der Entscheidungen werden die durch das neuronale Netz gelernten Merkmale und Zwischenrepräsentationen visualisiert.

Corneal confocal laser scanning microscopy is a promising method for in vivo investigation of cellular structures, e. g., of nerve fibers in the sub-basal nerve plexus. During recording, even slight displacements of the focal plane lead to images of adjacent tissue layers. In this work, we propose a closed-loop control of the focal plane. To detect and evaluate the visible tissues, we utilize the Bag of Visual Words approach to implement a customizable image processing pipeline for real-time applications. Furthermore, we show that the proposed model can be trained with small classification datasets and can be applied as a segmentation method. The proposed control loop, including tissue detection, is implemented in a proof-of-concept setup and shows promising results in a first evaluation with a human subject.

This thesis is settled in the field of data-based modeling (identification) and specifically focuses on the weakening of the effects of the curse of dimensionality with local model networks (LMNs). The methods for fighting the curse of dimensionality originate from the fields of input selection and design of experiments (DoE).