Mid-Infrared laser spectroscopy detection and quantification of explosives in soils using multivariate analysis and artificial intelligence
Cargando...
Fecha
2020
Autores
Pacheco-Londoño, Leonardo C.
Warren, Eric
Galán-Freyle, Nataly J.
Villarreal-González, Reynaldo
Aparicio-Bolaño, Joaquín A.
Ospina-Castro, María L.
Shih, Wei-Chuan
Hernández-Rivera, Samuel P.
Título de la revista
ISSN de la revista
Título del volumen
Editor
MDPI
Facultad de Ingenierías
Facultad de Ingenierías
Resumen
A tunable quantum cascade laser (QCL) spectrometer was used to develop methods for
detecting and quantifying high explosives (HE) in soil based on multivariate analysis (MVA) and
artificial intelligence (AI). For quantification, mixes of 2,4-dinitrotoluene (DNT) of concentrations
from 0% to 20% w/w with soil samples were investigated. Three types of soils, bentonite, synthetic
soil, and natural soil, were used. A partial least squares (PLS) regression model was generated for
predicting DNT concentrations. To increase the selectivity, the model was trained and evaluated using
additional analytes as interferences, including other HEs such as pentaerythritol tetranitrate (PETN),
trinitrotoluene (TNT), cyclotrimethylenetrinitramine (RDX), and non-explosives such as benzoic acid
and ibuprofen. For the detection experiments, mixes of different explosives with soils were used to
implement two AI strategies. In the first strategy, the spectra of the samples were compared with
spectra of soils stored in a database to identify the most similar soils based on QCL spectroscopy.
Next, a preprocessing based on classical least squares (Pre-CLS) was applied to the spectra of soils
selected from the database. The parameter obtained based on the sum of the weights of Pre-CLS was
used to generate a simple binary discrimination model for distinguishing between contaminated and
uncontaminated soils, achieving an accuracy of 0.877. In the second AI strategy, the same parameter
was added to a principal component matrix obtained from spectral data of samples and used to
generate multi-classification models based on different machine learning algorithms. A random forest
model worked best with 0.996 accuracy and allowing to distinguish between soils contaminated with
DNT, TNT, or RDX and uncontaminated soils.
Descripción
Palabras clave
Quantum cascade laser, Remote detection, Partial least squares, High explosives, Artificial intelligence, Machine learning