Classification of secondary explosives with a 1D convolutional neural network technique using terahertz time-domain spectroscopy in reflection geometry

Abstract

Terahertz time-domain spectroscopy (THz-TDS) stands out as a prominent spectroscopic technique ideal for identifying explosives. The integration of THz data with machine learning models enables rapid identification and classification of explosive molecules. The paper reports the terahertz time domain spectral study of premium explosives such RDX, HMX, TNT, PETN & Tetryl in reflection geometry. We have recorded the absorption spectra and refractive index of the explosives in the frequency range of 0.2 THz–3 THz. Principal component analysis (PCA) was employed for extracting the essential features from the data. In the next step, supervised machine learning algorithms such as support vector mechanism (SVM), K-Nearest neighbor (KNN), and Random Forest (RF) along with principal component analysis were used for the classification of these explosives based on the terahertz spectral data such as absorption spectra, refractive index and Fast Fourier's transform (FFT). The prediction accuracies achieved by supervised machine learning models were above 90 %. In addition, a one-dimensional convolutional neural network (1D-CNN) was implemented for classification, which has outperformed traditional machine learning models by achieving prediction accuracies greater than 95 %. The study has demonstrated that a combination of terahertz spectroscopy and 1D-CNN proves to be an efficient and practical tool for the identification of explosives.