1 |
Functional principal component analysis based machine learning algorithms for spectral analysisBie, Yifeng 07 September 2021 (has links)
The ability to probe molecular electronic and vibrational structures gives rise to
optical absorption spectroscopy, which is a credible tool used in molecular quantification and classification with high sensitivity, low limit of detection (LoD), and
immunity to electromagnetic noises. Spectra are sensitive to slight analyte variations, so they are often used to identify a sample’s components. This thesis proposes
several methods for quick classification and quantification of analysts based on their
absorbance spectra. functional Principal Component Analysis (fPCA) is employed
for feature extraction and dimension reduction. For 1,000-pixel spectra data, fPCA
can capture the majority variance with as few output scores as the number of expected analytes. This reduces the amount of calculation required for the following
machine learning algorithms. Further, the output scores are fed into XGBoost and
logistic regression for classification, and fed into XGBoost and linear regression for
quantification.
Our models were tested on both synthesized datasets and experimentally acquired
dataset. Our models demonstrated similar performance compared to deep learning
but with much faster processing speeds. For the synthesized 30 dB dataset, our
model XGBoost with fPCA could reach a micro-averaged f1 score of 0.9551 ± 0.0008,
while FNN-OT [1] could obtain 0.940±0.001. fPCA helped the algorithms extract the
feature of each analyte; furthermore, the output scores nearly had a linear relationship
with their concentrations. It was much easier for the algorithm to find the mapping
function between the inputs and the outputs with fPCA, which shortened the training
and testing time. / Graduate
|
Page generated in 0.0718 seconds