|M.Sc Student||Batzon Meni|
|Subject||Blind Separation of IR Spectroscopy Signals|
|Department||Department of Chemical Engineering||Supervisors||Professor Yaron Paz|
|Professor Emeritus Yehoshua Zeevi|
|Full Thesis text - in Hebrew|
The purpose of the work presented here is to develop methodology for a wise separation of infra-red (IR) spectra of mixtures as a part of an approach trying to utilize I.R. for remote sensing. The spectral signal is almost always composed of a contribution from number of compounds. Therefore, the problem can be characterized as developing an algorithm for separating a multi-source signal into the particular signals of the sources. The problem of source separation was solved already by the Blind Source Separation (BSS) approach for cases such as images and audio signals. In this work we focused on the implementation of this approach for the particular case of spectral measurements of plants in the mid infra-red region.
Five plants were chosen as a model system to study the BSS problem. The first step consisted of creating a data set containing the spectra of the plants. A variety of indexes were constructed characterize the quality of the separation process.
Several series of mixtures, constructed from different couples of sources, were separated. The aim was to produce quantitative criteria and ability to foresee if the separation algorithm succeeds. It was expected to find that the quality of the separation would decrease with the similarity of the couple of sources. Indeed, this tendency was observed, albeit not too clear. A considerable improvement in quality of separation following signal dilution was observed. It was found that dilution with the "HAAR" wavelets family produced the best separation results for separation of IR spectra signals.
The limitations of the separation process were found by adding a random noise or, alternatively, by adding a known spectral signal into the measurements of the mixtures. In the latter case, the number of the measurements of the mixtures is smaller than the number of sources. It was found that the algorithm could handle quite well a situation where an additional source is added. On the other hand, the algorithm was found to be more sensitive to the addition of random noise.
Overall, our results showed that the separation method presented in this work, could efficiently separate spectral mixtures of plants in the mid infra-red region. It is sensible that it can separate also other spectral families, such as the spectra of gases in the atmosphere. Hence, this method has the potential to compete with other spectral signal separation methods, as it enables fast and automatic detection of materials.