Informationstheoretische Optimierung künstlicher neuronaler Netze : für den Einsatz in Steuergeräten

DSpace Repositorium (Manakin basiert)


Dateien:

Zitierfähiger Link (URI): http://nbn-resolving.de/urn:nbn:de:bsz:21-opus-11647
http://hdl.handle.net/10900/48572
Dokumentart: Dissertation
Erscheinungsdatum: 2003
Sprache: Deutsch
Fakultät: 7 Mathematisch-Naturwissenschaftliche Fakultät
Fachbereich: Sonstige - Informations- und Kognitionswissenschaften
Gutachter: Rosenstiel, Wolfgang
Tag der mündl. Prüfung: 2003-07-23
DDC-Klassifikation: 004 - Informatik
Schlagworte: Wechselseitige Information , Informationstheorie , Neuronales Netz , Globale Optimierung , Signalanalyse
Freie Schlagwörter: Transinformation , Digitale Signalverarbeitung
Mutual Information , Information Theory , Neural Networks , Optimization , Digital Signal Processing
Lizenz: http://tobias-lib.uni-tuebingen.de/doku/lic_mit_pod.php?la=de http://tobias-lib.uni-tuebingen.de/doku/lic_mit_pod.php?la=en
Gedruckte Kopie bestellen: Print-on-Demand
Zur Langanzeige

Inhaltszusammenfassung:

Künstliche neuronale Netze werden in der Regel durch eine wiederholte Präsentation von Trainingspattern adaptiert, welche aus Meßdaten gewonnen wurden. Die zur Abbildung des Trainingsraums erforderliche Größe des Netzes muß dabei in der Regel durch den Anwender vorgegeben werden und hängt stark von der Komplexität der Daten ab. Zur Bestimmung der Komplexität von Meßdaten kann ein von Claude Shannon schon 1948 vorgeschlagenes Informationsmaß verwendet werden. Aufbauend auf der sogenannten Transinformation werden verschiedene Verfahren zur Datenanalyse und zur Erzeugung einer an die Komplexität angepaßten Netztopologie vorgestellt. Durch den Einsatz der Transinformation ist zunächst ein Verfahren entwickelt worden, das eine informationstheoretische Bewertung von Fourier-Spektren ermöglicht. Dabei werden einzelne Frequenzbereiche entsprechend ihres Informationsanteils bezogen auf das Gesamtspektrum gewichtet. Diese Gewichtung bietet eine zusätzliche Hilfe bei der Entwicklung digitaler Filter. Die informationstheoretische Bewertung von Trainingspattern konnte für die Entwicklung einer Methode zur gezielten Auswahl tatsächlich notwendiger Netzeingänge genutzt werden. Aufbauend auf so gewonnenen Erfahrungen ist schließlich ein Verfahren zur Erzeugung vollständiger Netztopologienen entwickelt worden. Dabei wird eine Topologie erzeugt, die der Komplexität des Trainingsraums entspricht. Diese neu entwickelten Verfahren können bei der Bestimmung von Einspritzmengen an einem Common-Rail Dieselmotor erfolgreich eingesetzt werden. Dies geschieht durch eine informationstheoretische Analyse und Auswertung von Meßdaten des Raildrucksignals. Dadurch kann ein kleines neuronales Netz erzeugt werden, das die Einspritzmenge für jede Verbrennung zylinderindividuell bestimmen kann. Die erzielte Genauigkeit ist dabei vergleichbar mit den Ergebnissen sehr viel größerer Netze herkömmlicher Verfahren. Diese wesentlich kleinere Netztopologie ermöglicht eine Echtzeitüberwachung der Einspritzmenge in Motorsteuergeräten mit begrenzter Rechen- und Speicherkapazität.

Abstract:

Artificial neural networks are usually adapted by repeated exposure to training patterns, which are gained from physical measurements. The size of the network required to map the training space has to be set by the user and depends on the complexity of the data. In order to analyze the complexity of measurements, the information scale proposed by Claude Shannon in 1948 can be used. Based on the so-called mutual information principle, several methods for data analysis and the creation of network topologies, which are adapted to the complexity of the data, have been developed. By using mutual information, a method was developed which allows for the evaluation of Fourier spectra based on information theory, where elaborate frequency ranges are weighted depending on their information content in relation to the whole spectrum. This weighting offers additional support in the design and development of digital filters. The evaluation of training patterns, based on information theory, could be applied to develop a method for selectively choosing the necessary network inputs. Finally, based on these experiences, a method could be developed to generate complete network topologies which match the complexity of the training space. These newly developed methods can be applied effectively in the estimation of injection quantities for a common rail diesel engine. This can be done by analyzing and evaluating measurement data from the rail pressure signal, based on information theory. Thus a small neural network can be created that determines the injected quantity for each combustion and cylinder individually. The achieved accuracy can be compared to the results of much bigger conventional networks. This considerably smaller network topology enables the real time monitoring of injection quantities in engine control units with limited calculation and memory capacities.

Das Dokument erscheint in: