Resolución de problemas de detección y clasificación mediante soluciones óptimas no supervisadas

Abstract

The event detection problem in a signal or temporal series can be defined as an unsupervised classification problem. The detection is performed without any prior knowledge about the nature of the events. The unsupervised classification problem of signals can be managed as a classification problem of temporal sequences. The signals are classified according to their behavior. These can have different duration and their events an uncertain temporal location. Both problems, detection and classification, are linked and, as addressed in this thesis, they can be solved through grouping techniques of objects or Clustering. In this work, it is proposed to use Clustering to obtain an unsupervised optimal solution. The solutions to both problems can be applied to multichannel signals and it is interesting in different fields as diverse as bioengineering, geophysics, nuclear fusion, etc. Thanks to event detection in a signal, in addition to the signal classification, other applications arise. For example, the analysis of states in a signal. A state is the signal portion between two consecutive events. Each state of a signal can be analysed by means of its independent components. In this thesis, a method to determine its optimal number in an unsupervised way is presented.