123 Pages
English
Gain access to the library to view online
Learn more

Time series analysis and classification with state-space models for industrial processes and the life sciences [Elektronische Ressource] / vorgelegt von Mark Christoph Jäger

Gain access to the library to view online
Learn more
123 Pages
English

Description

Inaugural - DissertationzurErlangung der DoktorwurdederNaturwissenschaftlich-mathematischen GesamtfakultatderRuprecht - Karls - UniversitatHeidelbergvorgelegt vonDiplom-Ingenieur Mark Christoph J ageraus StuttgartTag der mundlic hen Prufung: 24.05.2007Time Series Analysis and Classi cationwith State-Space Modelsfor Industrial Processesand the Life SciencesGutachter: Prof. Dr. Fred A. HamprechtProf. Dr. Gerhard ReineltAbstractIn this thesis the use of state-space models for analysis and classi cation oftime series data, gathered from industrial manufacturing processes and thelife sciences, is investigated. To overcome hitherto unsolved problems in bothapplication domains the temporal behavior of the data is captured using state-space models.Industrial laser welding processes are monitored with a high speed camera andthe appearance of unusual events in the image sequences correlates with errorson the produced part. Thus, novel classi cation frameworks are developedto robustly detect these unusual events with a small false positive rate. Forclassi er learning, class labels are by default only available for the completeimage sequence, since scanning the sequences for anomalies is expensive.The rst framework combines appearance based features and state-spacemodels for the unusual event detection in image sequences.

Subjects

Informations

Published by
Published 01 January 2007
Reads 17
Language English
Document size 21 MB

Exrait

Inaugural - Dissertation
zur
Erlangung der Doktorwurde
der
Naturwissenschaftlich-mathematischen Gesamtfakultat
der
Ruprecht - Karls - Universitat
Heidelberg
vorgelegt von
Diplom-Ingenieur Mark Christoph J ager
aus Stuttgart
Tag der mundlic hen Prufung: 24.05.2007Time Series Analysis and Classi cation
with State-Space Models
for Industrial Processes
and the Life Sciences
Gutachter: Prof. Dr. Fred A. Hamprecht
Prof. Dr. Gerhard ReineltAbstract
In this thesis the use of state-space models for analysis and classi cation of
time series data, gathered from industrial manufacturing processes and the
life sciences, is investigated. To overcome hitherto unsolved problems in both
application domains the temporal behavior of the data is captured using state-
space models.
Industrial laser welding processes are monitored with a high speed camera and
the appearance of unusual events in the image sequences correlates with errors
on the produced part. Thus, novel classi cation frameworks are developed
to robustly detect these unusual events with a small false positive rate. For
classi er learning, class labels are by default only available for the complete
image sequence, since scanning the sequences for anomalies is expensive.
The rst framework combines appearance based features and state-space
models for the unusual event detection in image sequences. For the rst time,
ideas adapted from face recognition are used for the automatic dimension reduc-
tion of images recorded from laser welding processes. The state-space model
is trained incrementally and can learn from erroneous sequences without the
need of manually labeling the position of the error event within sequences. In
addition, a second framework for the object-based detection of sputter events
in laser welding processes is developed. The framework successfully combines
for the rst time temporal change detection, object tracking and trajectory
classi cation for the detection of weak sputter events.
For the application in the life sciences the improvement and further develop-
ment of data analysis methods for Single Molecule Fluorescence Spectroscopy
(SMFS) is considered. SMFS experiments allow to study biochemical processes
on a single molecule basis. The single molecule is excited with a laser and
the photons which are emitted thereon by uorescence contain important in-
formation about conformational changes of the molecule. Advanced statistical
analysis techniques are necessary to infer state changes of the molecule from
changes in the photon emissions. By using state-space models, it is possible
to extract information from recorded photon streams which would be lost with
traditional analysis techniques.Zusammenfassung
In dieser Dissertation wird die Anwendung von Zustandsraum-Modellen zur
Analyse und Klassi k ation von Zeitreihen, die aus industriellen Produktions-
prozessen und den Biowissenschaften stammen, untersucht. Um bisher un-
gel osten Problemen in beide Anwendungsgebiete beizukommen, wird eine Zu-
standsraum-Darstellung gew ahlt, die den zeitlichen Charakter der Daten er-
fassen kann.
Industrielle Laserschwei prozesse werden mit Hochgeschwindigkeitskameras
ub erwacht, wobei das Auftreten von ungew ohnlichen Ereignissen in den Bild-
folgen mit Fehlern am produzierten Bauteil korreliert. In dieser Arbeit werden
neuartige Auswertesysteme entwickelt um diese ungew ohnlichen Ereignisse zu-
verl assig au nden zu k onnen. Fur das Trainieren eines Klassi k ationssystems
stehen standardm assig nur \Label" fur gesamte Bildfolgen zur Verfugung, da
es sehr zeitaufwendig ist, jedes einzelne Bild der Folge einzeln auf Unregelm ass-
igkeiten hin zu untersuchen.
Das erste Uberwachungssystem kombiniert erscheinungsbasierte Merkmale
und Zustandsraum-Modelle zum Au nden ungew ohnlicher Ereignisse. Hierbei
werden erstmals Ideen aus der Gesichtserkennung fur die automatische Dimen-
sionsreduktion der aufgenommen Bilder von Laserschwei prozessen verwendet.
Dasdell wird inkrementell aufgebaut und kann den Infor-
mationsgehalt von fehlerhaften Sequenzen automatisch nutzen, ohne dass die
genaue Fehlerposition manuell spezi ziert werden muss (schwach ub erwac_ htes
Lernen). Zus atzlich wird ein objektbasiertes Klassi k ationssystem zur Erken-
nung schwacher Schwei spritzer vorgestellt. Dabei werden zeitliche Anderungs-
detektion, Objekverfolgungsalgorithmen und die Klassi k ation von Trajekto-
rien erstmals erfolgreich miteinander kombiniert, um schwache Schwei spritzer
robust zu erkennen.
Der zweite Beitrag ist die Verbesserung und Weiterentwicklung von Daten-
auswerteverfahren fur die Einzelmolekul uoreszenzsp ektroskopie (SMFS). Bei
SMFS Versuchen werden einzelne Molekule mit einem Laser angeregt und die
durch Fluoreszenz erzeugten Photonen mit hoher zeitlicher Au osung aufge-
nommen. Die Aufgabe der Datenanalyse ist, aus Anderungen in der Photon-
intensit at, auf unterschiedliche Zust ande des Molekuls zu schlie en. Durch
die Benutzung von Zustandsraum-Modellen wird es m oglich Informationen zu
extrahieren, die mit klassischen Auswerteverfahren verloren gegangen w aren.Hiermit erkl are ich, Mark Christoph J ager, dass ich die vorgelegte Dissertation
selbst verfasst und mich dabei keiner anderen als der von mir ausdruc klich be-
zeichneten Quellen und Hilfen bedient habe.
Kornwestheim, den 10. April 2007ivPreface
First and foremost I would like to express my gratitude to my doctoral advi-
sor Professor Fred A. Hamprecht from the Interdisciplinary Center of Scienti c
Computing (IWR), University of Heidelberg, Germany for his support and en-
couragement over the last years and particularly for giving me the push forward
to nish this thesis. His outstanding commitment to my work was a key factor
for its success.
I gratefully acknowledge nancial support of this thesis by the Research and
Development Department of the Robert Bosch GmbH, Schwieberdingen, Ger-
many. In particular, I would like to thank Walter Happold from the Robert
Bosch GmbH for o ering me the opportunity to gain experience solving chal-
lenging real-world problems. He always managed to combine the company’s
interest with giving me freedom to pursue scienti c research. Thanks also go
to Christian Knoll from the Robert Bosch GmbH, Schwieberdingen, Germany.
We worked together on several projects and he was always committed to trans-
ferring my work from the research stage to real applications.
Many thanks also go to Alexander Kiel and Dirk-Peter Herten from the Insti-
tute of Physical Chemistry, University of Heidelberg, Germany who introduced
me to the interesting world of single-molecule spectroscopy and provided the
data for the work presented in chapter 7. The discussions with them have
always been an indispensable source of inspiration.
I am indebted to my colleagues and my fellow doctoral students: Andreas,
Jochen, Linus, Marco, Matthias, Stefan and Thomas. If work was fun, then
it was due to them. I will de nitely miss our daily \Ice Cream-Tours". I also
want to express my gratitude to all members of the Multidimensional Image
Processing Group at the IWR for their support and the nice time we had
during several seminars. Special thanks go to Nadin, Andreas, Chris and Fred
for comments on the nal version.
Sincere thanks go to my family who continuously supported me during the
long years of my studies. And last but not least I want to thank Nadin for her
patience and support over the last years, without her I would not have nished
this work.
vvi