Event Details
Spectroscopic Analysis Using Residual Convolutional Neural Network
Presenter: Tianyang Zhao
Supervisor:
Date: Fri, December 17, 2021
Time: 08:30:00 - 09:30:00
Place: ZOOM - Please see below.
ABSTRACT
Zoom link:
Meeting ID: 857 1511 1954
Password: 409004
Note: Please log in to Zoom via SSO and your UVic Netlink ID
Summary: By analyzing the absorption spectrum, we can classify and quantify the gases and solution in mixtures. This report proposes the 1D-ResCNN algorithm for classification and quantification based on their absorption spectra. Principle component analysis is employed for noise reduction. The noise of the spectrum is effectively reduced, and it improves the performance of our algorithms. We also applied the optimal thresholding to choose the best thresholds to determine the existence of materials. We test our model on synthesized datasets, experimentally acquired datasets, and real gas mixture absorption spectrum. For the classification task, we compared our results with those of the Feed-forward neural network and PLS, our algorithm can obtain a higher F_1 score on the high signal-to-noise ratio (SNR) synthesized datasets, and the predictions for real spectrum are correct. The performance on the experimentally acquired dataset is close to that found in the 40dB SNR dataset. By comparing the results obtained by PLS in the quantification task, our algorithm is suitable for getting the gas concentration from the high SNR datasets and the gas concentration with high absorbance feature even the SNR of the spectrum is low.