| Title |
A study into digital denoising of nir spectroscopy measurements / |
| Translation of Title |
Skaitmeninių triukšmo šalinimo metodų artimųjų infraraudonųjų spindulių spektroskopiniuose matavimuose tyrimas. |
| Authors |
Andrejevaitė, Laima |
| Full Text |
|
| Pages |
49 |
| Keywords [eng] |
Denoising autoencoder, NIRS, spectroscopy, partial least squares regression, filtering, measurements |
| Abstract [eng] |
Near infrared spectroscopy (NIRS) provides a non-invasive, cost effective and efficient tool for monitoring compositions of various substances. Potential applications of NIRS in medicine could simplify and improve the quality of treatment for diabetic or septic patients. The precision and reliability of any tools used in medicine are of extremely high standard since the price for an error can be immense, however, measurements of any kind inherently have errors attached and data preprocessing algorithms play a crucial role in analysing the dataset. This thesis proposes an autoencoder based approach to NIRS measurement denoising that is directly comparable to the currently used filtering functions. To achieve this goal a two-step approach was designed to test the denoising autoencoder (DAE) and compare its results to the Butterworth “lowpass” filter (BLP) and the Savitzky-Golay filter (SVG). The DEA was trained on simulated (artificial) data with concentrations in the same range as the measurements. Filtering parameters of BLP and SVG were also optimized on the simulated dataset. Then the measured dataset was filtered three different ways and the results were compared through the mean absolute error (MAE) and the partial least squares regression (PLSR). BLP and SVG exhibited better results when comparing the MAE, however, DAE outperformed both filters when PLSR results were compared. This offers a conclusion that the DAE distorts some parts of the dataset (spectra) in order to preserve the important features, that are later recongnised by the PLSR. |
| Dissertation Institution |
Vilniaus universitetas. |
| Type |
Master thesis |
| Language |
English |
| Publication date |
2021 |
