Open Access Open Access  Restricted Access Subscription Access

A CNN-LSTM-AM APPROACH TO DECODING COMPLEX RAMAN SPECTRA IN THE CHARACTERISATION OF HMES

Nomahlubi Thunyiswa, Leon Ndhlozi, Phakamani Sikhakhane, Lisa Ngcebesha, Lamla Thungatha, Conrad Mahlase

Abstract


Detection of homemade explosives (HMEs) in heterogeneous soil matrices remains a persistent challenge when using portable Raman spectroscopy. Traditional Raman analysis suffers from fluorescence interference, low sensitivity, and poor reproducibility outside controlled laboratory conditions and in complex sample mixtures. To overcome these limitations, a multi‑output deep learning framework, CNN-LSTM-AT, was developed and it integrates convolutional feature extraction, bidirectional sequential modelling, and attention mechanisms. The architecture achieved stable convergence, avoided overfitting, and consistently localised chemically diagnostic vibrational modes such as the symmetric nitrate stretch near 1040 cm-1. Quantitative results revealed variance compression under Mean Squared Error optimisation, however, the model maintained reliable trend tracking and detection thresholds. Attention maps confirmed chemically meaningful focus, distinguishing molecular bond stretching frequencies from background noise. Together, these findings demonstrate that explainable deep learning can bridge the gap between laboratory validation and operational field reliability, establishing a foundation for future optimisation with hybrid loss functions.1


DOI
10.12783/ballistics26/37593

Full Text:

PDF

Refbacks

  • There are currently no refbacks.