You are here

Machine Learning Based Integration of Alarm Resolution Sensors


Develop an integrated Alarm Resolution sensor suite with smart algorithms to enable high collection efficiency of explosive samples and signatures and enhance AR detectors’ performance, while simultaneously reducing user information overload. Explosive threats come in all shapes, sizes, concealments, and nuanced formulations.

Due to these complexities, DHS Components employ a variety of detectors from Explosives Trace Detectors (ETD) based on both Ion Mobility Spectrometry and Mass Spectrometry, vapor detectors, bulk resolution detection (Infrared (IR) and Raman based), through barrier detection (Spatially Offset Raman Spectrometry), and colorimetric kits. However, having a multitude of AR tools can lead to information overload in end-users.

Information overload is especially pronounced in crowded and high throughput environments such as aviation and security checkpoints. As a baseline, a Transportation Security Officer (TSO) currently would have to follow a multi-step decision tree to screen and resolve an alarm of a suspicious object. As part of this, the TSO would have to characterize shapes, sizes, and color of objects and to mentally categorize objects according to their utilities and compositions. From these initial screenings, they would determine the best tool(s) at their disposal to detect and identify explosives whether it is an ETD, colorimetric, IR, Raman, vapor detector, or any combination thereof. They would then determine how best to sample and collect explosives signatures whether it is through swabbing residues at frequently touched surfaces or aiming a laser to excite unknown substances and collecting their signatures.

One way to reduce information overload is through rigorous training. TSOs are trained on how to execute the decision tree and with practices of sampling techniques, they develop muscle memory (i.e. with the use of a Pressure Sensitive Wand). However, such training could only alleviate information overload incrementally. Thus, one alarmed object after another and day after day, TSOs may experience a high level of information overload which accumulates over time leading to user fatigue.

In response to this topic, S&T seeks a proposed solution to develop a new capability that comprehensively alleviates aforementioned information overload on the users. This capability would characterize shapes, sizes, and color of objects and from this initial characterization, categorize objects according to their utilities and compositions. From these initial screenings, the capability would inform a user on the best tool(s) to detect and identify explosives. The proposed capability solution should consist of three sub-components, all integrated within a desktop size box:

1) New sensors which can scan the object and categorize material compositions according to their physical characteristics (ex: electrical conductivity, magnetic property...);

2) A Machine Learning algorithm that takes sensory output, analyzes, and suggests the best AR detector to sample or collect signatures of unknown substances; and

3) Upon user confirmation, actuate the AR detector to detect and identify explosives.

This is in essence a decision analytics tool specifically applied toward enhanced Alarm Resolution. Performance parameters for this proposed capability include the Probability of Categorizing (PC) correctly objects according to their utilities and function (Phases I and II) and the Probability of Recommending (PR) the right AR tool(s) (Phase III). For comparison, PC and PR will be collected from a well-informed TSO who employs the AR decision tree according to five different bins of characteristics and functions and four different AR tools.

US Flag An Official Website of the United States Government