(584f) In Situ FT-IR Quantitative Analysis of Amine Concentration and CO2 Loading Amount in Mixture Solvent Using Deep Neural Network

Title : In situ FT-IR
Quantitative Analysis of Amine Concentration and CO₂ Loading Amount
in Mixture Solvent using Artificial Neural Network

Yo Sung Yoon, Jay H.
Lee*

Email: jayhlee@kaist.ac.kr

This
work presents the quantitative analysis of amine based CO₂ capture
process for reclaimer and make up. The process has been widely investigated
with the first generation amine solvent, but in the long term operation, many
challenging problems regarding stability originated from amine degradation and
corrosion have been raised. These problems may cause severe solvent loss,
foaming, and fouling. Therefore, it is required to develop online monitoring
method to cope with those unfavorable phenomena. As a breakthrough, this study
suggests the in-situ FT-IR quantitative analysis model of the amine
concentration and CO₂ loading amount in MEA, MDEA, AMP mixture,
which are the well-known commercialized absorbents in the CO₂
capture process, by employing reduced order artificial neural network.

The
input data of neural network model is FT-IR data of calibration sets and the concentrations
of MEA, MDEA, and AMP as well as CO₂ loading amount are set as
supervision of corresponding data. A total of 60 calibration set is composed of
all FT-IR data, recorded in the range of 790cm^-1 to 4000cm^-1
and divided by 6660 intervals, resulting in the input matrix of [60X6660]. In
the preprocessing of data, principal component analysis(PCA) is used for model
order reduction due to highly redundant FT-IR raw data. This study involves 7
principal components in which their cumulative variance is above 99% of the
total variance. Then, the size of input matrix becomes [60X7] with minor loss
of data information. The artificial neural network with two hidden layers is
applied in regression with these reduced order input data. As a result, the coefficients
of determination of the model over 0.999 are achieved and validation results
with test sets fall within approximately 5% relative error. The leave-one-out
cross validation is also applied.