Series in Signal and Information Processing, Vol. 18
edited by HansAndrea Loeliger
Volker M. Koch
A Factor Graph Approach to ModelBased Signal Separation
First edition 2007, HartungGorre Verlag Konstanz, 328 pages,
ISBN10: 3866281404, ISBN13:
9783866281400, 64 €
This doctoral thesis is about a rather new modelbased signal
processing methodology that is based on factor graphs and
messagepassing algorithms. Using this, we have developed exemplary
signal processing algorithms for various biomedical applications. The
main application to evaluate and demonstrate this new methodology was in
the field of electromyographic (EMG) signal analysis. EMG signals are
electrical signals that are generated during muscle contractions. They
can be measured using various kinds of electrodes, e.g., needle or
finewire electrodes. Their analysis provides valuable information for
the diagnosis of neuromuscular disorders, the study of neuromuscular
control mechanisms, and the verification of anatomic hypotheses.
EMG signals are essentially made up of superimposed action potential
trains from several sources. For a thorough analysis, the measured EMG
signals need to be decomposed into their constituent trains. However,
this task can become difficult if action potentials from different
sources overlap. Many EMG signal decomposition methods have been
proposed. Traditional algorithms often use heuristic segmentation and
clustering approaches. Especially in the case of difficult
superpositions with many overlapping action potentials, these algorithms
could often not decompose muscle signals correctly. Many other
approaches are computationally not feasible when superpositions of many
action potentials are to be resolved.
To be able to decompose EMG signals with difficult superpositions, new
signal processing algorithms based on factor graphs were developed.
Factor graphs allow the systematic derivation of advanced modelbased
signal processing algorithms. A factor graph is a graphical model that
represents a factorization of a function. Instead of starting with such
a function, we explain how a factor graph for EMG signal decomposition
can intuitively be obtained from a block diagram. Here, the block
diagram is a simulation model for EMG signals. The factor graph approach
allowed us to integrate action potential shape information, firing
statistics, and other properties of EMG signals into the same model.
Finally, we show how decomposition is achieved by means of the
sumproduct algorithm. It performs inference by passing messages along
the edges of a factor graph. Since our factor graphs have cycles, we get
suboptimal iterative algorithms that allow handling complex models,
which could not be used just a few years ago. We have developed several
algorithms that differ, e.g., in the representation of the messages
propagating in factor graphs. Our new algorithms allow the fast
resolution of single and multichannel superpositions consisting of many
overlapping action potentials.
We have also used the factor graph language to derive novel
messagepassing algorithms for several other biomedical applications.
One example is the extraction of heart beats from pressure sensors that
are located under bedposts (seismosomnography). The advantage of this
method is that the signals can be recorded without attaching electrodes
to a human subject while sleeping. Other examples include the important
topic of multichannel neural spike sorting and blindsource separation
for electroencephalographic signals. We present exemplary factor graphs
for all these applications.
Keywords
Graphical models, factor graphs, signal modeling, modelbased signal
processing, sumproduct algorithm, message passing, superpositions,
signal decomposition, resolving superpositions, electromyography,
seismosomnography, multichannel neural spike sorting, blind source
separation.
Examiners
My examiner was Professor Dr. HansAndrea Loeliger of ETH Zurich and
my coexaminer was Professor Dr. Kevin C. McGill of Stanford University.
Links and Downloads
Buy my dissertation on
Amazon.de
Get more information on the series in which my dissertation was
published:
HartungGorre
Download
a complete PDF version [about 13 MB]
Download a complete PDF version with reduced figure quality [about 2 MB]
Download the slides of my exam
talk [about 8 MB]
Cover
