REMOTE MULTI-SENSOR FUSION FOR EXPLOSIVE MATERIALS DETECTION: INTEGRATING CHEMICAL GAS SENSING WITH DATA-DRIVEN PREDICTIVE MODELS FOR IED IDENTIFICATION
Abstract
This paper presents the Multi-Modal Chemical Sensing Network (MMCSN), a ROS2-based framework developed at the CSIR Detonics, Ballistics and Explosives Laboratory for staged multi-modal explosive-material detection on an unmanned ground vehicle (UGV). The system integrates three sensing modalities, metal-oxide (MOX) gas sensing via a Bosch BME690 8× Shuttle Board, Raman spectroscopy via replayed pre-extracted spectral features, and differential thermal analysis (DTA) using offline thermal curves. All classification is performed by Random Forest models, an earlier deep-learning approach did not converge on the available corpus. Five independent evaluations are reported, a controlled BME690 stimulus demonstration, an archived gas classifier evaluation, two Raman classifier cross-validations, an offline taxonomy and reference-coverage audit, and a live ROS2 bench run. The three-class Raman model achieved 97.94% grouped-validation accuracy (HNS, Tetryl, TNT). The six-class Raman baseline achieved 84.90% overall accuracy, with Tetryl recall at 56.5% as the principal weakness. The archived gas classifier reached approximately 97% under a random split but approximately 38% under a later ordered split a result that does not support standalone confirmation use. DTA remained unavailable during the live bench run, and no paired physical multi-modal validation was completed in the present study. The system is deployed on an NVIDIA Jetson AGX embedded platform aboard a Jaguar 4 tracked UGV, running concurrently with an autonomous navigation stack.
DOI
10.12783/ballistics26/37674
10.12783/ballistics26/37674
Refbacks
- There are currently no refbacks.