PhD Thesis Proposal: John Brevard Sigman

Tuesday, May 30, 2017, 12:00–2:00pm

Jackson Conference Room, Cummings Hall

Ultra-wideband EMI sensing: Non-metallic target detection and automatic classification of unexploded ordnance


Buried explosive hazards are one of the most pressing problems worldwide. In the United States, there are up to 10 million acres contaminated with Unexploded Ordnance (UXO) from artillery testing, while in warzones UXO, improvised explosive devices (IEDs) and landmines are a threat for both military and civilians. There are three categories of explosive hazards: metallic, intermediate-electrical conducting (IEC), and non-conducting targets. Metallic target detection and classification by electromagnetic (EM) signature has been the subject of research for the last 20 years. Key to this success is modern multistatic Electromagnetic Induction (EMI) sensors, which are able to measure the complete (full relaxation including quadrature peak) EMI response from metallic buried targets. However, no hardware solutions exist which can characterize IEC and non-conducting targets. While high-conducting metallic targets exhibit a quadrature peak response for frequencies in a traditional EMI regime under 100 kHz, the response of intermediate-conducting objects manifests at higher frequencies, between 100 kHz and 15 MHz. In addition to high-quality electromagnetic sensor data and robust electromagnetic models, a classification procedure is required to discriminate Targets of Interest (TOI) from clutter. Currently, costly human experts are used for this task. This expense and effort can be spared by using statistical signal processing and machine learning. 

This thesis proposal has two main parts. In the first part, we explore using the HFEMI band (100 kHz-15 MHz) for detection of carbon fiber UXO, voids, and improvised explosive devices (IED). For our prototype HFEMI instrument, we apply the techniques of metal detection to sensing in a band of frequencies which are the transition between the induction and radar bands. 

In the second half, we present a procedure for automatic classification of UXO. For maximum generality, our algorithm is robust and can handle sparse training examples of multi-class data. It contains an unsupervised starter, semi-supervised techniques to gather training data, and concludes with supervised learning until all TOI are classified. Additionally, an inference method for remaining true positives from a partial Receiver Operating Characteristic (ROC) curve is presented and applied to UXO datasets.

Thesis Committee

For more information, contact Daryl Laware at