Técnicas de clasificación, optimización y procesado de señal aplicadas a sistemas basados en sensores de gases y líquidos

  1. Acevedo Rodríguez, Francisco Javier
Supervised by:
  1. Saturnino Maldonado Bascón Director

Defence university: Universidad de Alcalá

Fecha de defensa: 01 July 2009

  1. Luis Vergara Domínguez Chair
  2. Manuel Rosa Zurera Secretary
  3. Mariano Rincón Zamorano Committee member
  4. Shezard Al Khalifa Committee member
  5. Francisco López Ferreras Committee member
  1. Teoría de la Señal y Comunicaciones

Type: Thesis


This thesis is focused on signal processing and pattern recognition methods applied to the systems known as electronic noses and electronic tongues. These kind of systems appear as analytical techniques that try to mimic the smell and taste senses by means of a matrix of gas or liquid sensors plus a stage that previoulsy preprocess and classify the obtained information. The excellence network General Olfaction and Sensing Projects on a European Level (GOSPEL) was created to develop these systems. Its main research fields are divided into the improvement in the sensor's technologies, real time implementation and the improvement in the statistical preprocessing techniques, being the last mentioned the central point of this thesis. At a first stage, the state of the art is analyzed ,either in sensor's technologies or in the techniques applied, where is important to highlight the relevance the dynamic information has in the recent applied techniques. There is also a review of the main signal processing methods applied to these systems as well as the main classification methods are studied. From this bibliographic review, new methods are proposed to extract the dynamic information provided by the sensors and a comparative methodology between the different classification methods is established. The main target of this comparative study is to go into those methods in depth and to propose improvements adapted to the problems under study. At a second stage the dynamic information extraction is studied. An extension of the wavelet transform is proposed and a regression algorithm is adapted to model the signals obtained from the sensors and to use those parameters as discriminating information. The proposed methods are tested on a variety of data sets obtained from different sensor technologies trying to get the proposed methods applied to the most possible number of the sensor's technologies and techniques. At the classification stage we have proposed a new comparative methodology among the different classification algorithms found in the state of the art and new kernel classification methods, with a high level of sucess in other fields, are suggested. In the thesis framework, an incremental learning method is developed, being useful to the considered systems, since it makes easier obtain a new classification model in corporating a new assay in the learning process. This stage is closed with the proposal of a block feature selection method allowing to find zones and to relate this information with the physical phenomena that produce the discriminating information. The main conclusion of the comparative study among classifiers is that kernel methods give a high and adequate level of flexibility to work with electronic noses and electronic tongues. Especially when their parameters are adjusted, support vector machines appear as a classification method that achieve high levels of accuracy. As a result of this conclusion, at the last stage of this thesis several ideas are proposed to improve overall accuracy, to extend support vector machines to multiclass problems and to reduce the number of operations needed to evaluate future samples.