Substrate-Invariant Spectral Data Analysis for Explosive Detection Using Vibrational Spectroscopy

Kurzfassung

Reliable spectroscopic trace explosive detection on real-world materials remains a major challenge for field-deployable counter-IED and CBRN screening because substrate signatures often dominate the spectral response. As a result, machine-learning models trained directly on measured spectra frequently fail to generalize to previously unseen materials. This work presents a substrate invariant spectral - data analysis approach for robust explosive detection using MIR spectroscopy on complex cotton and polymer-based fabrics contaminated with various explosives including PETN, RDX, TNT. The proposed methodology combines advanced spectral analysis with interpretable machine learning to improve discrimination of explosive signatures under challenging real-world conditions. Using a strict unseen-substrate evaluation protocol, the approach achieved an average classification accuracy of ~99% and maintained 96.3 % accuracy on the most challenging unseen substrate, where a baseline approach without substrate suppression achieved only 89 %. The system has additionally been deployed as a real-time automated detection pipeline for live measurement analysis and alarm generation. Preliminary results from an independent Raman spectroscopy analysis pipeline will also be presented, enabling comparison of data-analysis methodologies as well as performance comparison of the two spectroscopic techniques for explosive trace detection.