Aiding the Detection of Explosive Materials with Machine Learning: Extracting Relevant Features from Multivariate Sensor Data

Abstract

Reliable and effective detection of explosive substances is one of the most important tasks of civil security. Often as a countermeasure to terrorist actors, the elimination of a potential explosive is an essential means of hazard containment. Scenarios can be varied and range from threats to the public life at airports, train stations or events, to critical infrastructure of highest security, such as nuclear power plants. Procedures for determining the threat potential of an unknown, suspicious object are manifold. The minimum requirements are the same for all of them: a fast decision-making process in a time-critical situation, as well as reliability and a high degree of coverage in the detection of different types of explosives. For this application scenario, the thesis elaborates on the furtherdevelopment of an already existing model which is used for the detection of explosive materials. The use case is directed to two further, new methodologies: the detection of high-energetic explosive materials, and the identification of specific subgroups, based on the chemical structure of the substances. Sensor data from a special measurement method for explosive and harmless substances is processed for this purpose. The thematic focus will be on the extraction of features for the creation of a prediction model. The aim of this thesis is to extend the existing decision process of detection by implementing new use cases in a meaningful way, in order to achieve a high level of detail about an unknown, potentially dangerous substance. A difference in performance can be seen in the results for both use cases. While the detection of high-energetic substances is well possible, there is still potential for improvement in the identification of subgroups according to chemical structure.