Susanne Töpfer

Semi-physical modelling of nonlinear processes by means of local model approaches

Abstract For many practical applications, a combination of theoretical and experimental modelling appears feasible. Qualitative knowledge about the most significant effects are often known or easily accessible. This contribution suggests a semi-physical modelling approach based on the special architecture of local models. Due to their inherent transparency, these models are very well suited for the incorporation of mainly qualitative process knowledge. The integration of prior knowledge is realised by an adaptation of the model structure to that one of the process. As a result, the final process-specific models are characterised by high generalisation performance also in situations with only few measurement data.

Effiziente Motorapplikation mit lokal linearen neuronalen Netzen

In einer Kooperation zwischen der Dr. Ing. h. c. F. Porsche AG und dem Institut fur Automatisierungstechnik der TU Darmstadt ist eine neue Versuchsplanungsmethode zur Reduktion des Messaufwands in der Motorapplikation entwickelt worden. Die neue Vermessungsstrategie kombiniert hierfur neuronale Netze mit Methoden der statistischen Versuchsplanung.

Polynommodelle, Kennfelder und neuronale Netze

Nach einer Einfuhrung des grundlegenden Vorgehens bei der experimentellen Modellbildung werden in diesem Kapitel verschiedene Modellarchitek- turen zur Approximation nichtlinearer statischer Prozesse vorgestellt. Hierzu zahlen klassische Ansatze wie die Polynommodelle und Rasterkennfelder sowie moderne Modellarchitekturen aus dem Bereich der neuronalen Netze, der Fuzzy-Systeme sowie der Neuro-Fuzzy Modelle. Fur ausgewahlte Modellvarianten werden die interne Struktur und geeignete Optimierungsalgorithmen aufgezeigt. Es erfolgt auserdem eine Diskussion der wichtigsten Vor- und Nachteile. Aufbauend auf den statischen Modellarchitekturen konnen nichtlineare dynamische Modelle fur die Identifikation dynamischen Prozess verhalt ens abgeleitet werden. Dieser Abschnitt stellt zwei Moglichkeiten der Erweiterung statischer Modelle zu dynamischen Modellansatzen vor.

Comparison of a Hierarchically Constructed Neural Network and a Hierarchical Look-Up Table

Abstract Two black-box models, LOLIMOT and hierarchical look-up tables, are presented which are well suited to deal with identification of complex nonlinear systems. LOLIMOT is a local linear neuro-fuzzy model which preferably can be used for off-line simulation. Hierarchical look-up tables represent an extension of classical look-up tables and allow the on-line simulation of the identified nonlinear systems under time-critical conditions. Both models share a recursive partitioning approach for solving the structure identification problem.

Realisation of Hierarchical Look-Up Tables on Low-Cost Hardware

Abstract Numerous applications which require nonlinear models for control, supervision, or fault detection are characterized by real-time requirements and limited computer power. The most widely used type of nonlinear models in such time-critical applications are grid-based look-up tables. However, these models fully underlie the curse of dimensionality. Therefore, their use is restricted to at most two-dimensional input spaces. In this paper, an alternative look-up table concept based on a hierarchical input space decomposition is proposed. These look-up tables can be efficiently represented in a tree-structured architecture which allows the implementation on lowcost hardware.

Efficient engine calibration with local linear neural networks

A new test planning method to reduce measurement effort in engine calibration has been developed in cooperation between Dr. Ing. h.c. F. Porsche AG and the Institute of Automatic Control at Darmstadt University of Technology. This new measurement strategy combines neural networks with statistical test planning methods.

A HINGE NEURAL NETWORK APPROACH FOR THE IDENTIFICATION AND OPTIMIZATION OF DIESEL ENGINE EMISSIONS

Abstract Characteristics such as nonlinearity, uncertainty, and time-delays make the identification of dynamic processes challenging. One means of addressing this problem is to develop approaches based on artificial neural networks that are capable of modeling highly nonlinear systems. The aim of this paper is to present the non linear mathematical models of exhaust gas formations achieved with a certain Neural Network (NN) architecture, namely Hinging Hyperplane Trees (HHT). A brief description of HHT is followed by a presentation of identification of nitrogen-oxides and opacity emissions for a Diesel engine. Experimental results that show the effectiveness of the HHT approach are included. The article concludes with the description of an optimization environment for Diesel engine management based upon previously achieved emission models, and which is used in order to obtain an optimal performance regarding fuel consumption, emissions and drivability

Hinging Hyperplane Trees for Hybrid Modeling of Alternators

Abstract Automotive alternators cause a significant proportion of the fuel consumption to fulfill the electric power demand of the vehicle. Therefore, mathematical models of fuel consumption and for on-line torque estimation in the vehicle have to incorporate representations of the alternator behavior. It has been shown that hybrid modeling approaches combining physical and black-box models are a suitable mean. In this paper, two neural black-box models, the well-known multi-layer perceptron and Hinging Hyperplane Trees are presented for the identification of the power loss of alternators. Both models utilize basis functions of the ridge construction type, which is a necessary property for the representation of the strong nonlinear behavior of the power loss output. The performance of both models is compared.