Methodology for anomalous source detection in sparse gamma-ray spectra

The dangers of rogue nuclear material remain a top concern despite increased attention and strides in computational protocols. Single, mobile detector methodologies for localizing sources via autonomous surveying have become popular with the maturation of the machine learning (ML) and statistical learning (SL) fields as well as increased access to drone (quad-copter) technology. These options, however, face task-inherent impediments which either degrade the quality of collected gamma-ray spectra or necessitate high-quality information on source and background spectrum compositions. Some such hurdles include: limited dwell periods, fluctuating and/or unknown background, weak source signal due to large distance and/or small/shielded activity, and the low sensitivity of mobile detectors. As such, collected gamma-ray spectra are sparse, containing many zero-count energy channels, and contain relatively large background presence. This combination of factors, as well as the natural variance in second-to-second count rates, leads to low-quality information for making navigational decisions. In this thesis, an SL algorithm is presented for extracting source count estimations from time-series, sparse gamma-ray spectra with no prior training required. A Gaussian process with a linear innovation sequences procedure is used to efficiently update ongoing spectral estimates with real-time training and hyperparameters defined by detector characteristics. Being free of prior training and assumptions allows such an algorithm to be used in a wide variety of sparse-data settings whereas a trained solution would have very narrow applications. We have evaluated the effectiveness of this approach for anomaly detection using background spectra dataset collected with a Kromek D3S and simulated source spectra. Results of anomaly detection testing with a source count rate at half that of the background displays an area under the ROC curve of 0.9. Further, deployment with an ML guided navigation scheme shows, after an anomaly is detected, estimated gross source counts and true gross source counts have an average correlation of 0.998, whereas estimated gross background counts and true gross background counts have an average correlation of 0.876.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Nuclear, Plasma, Radiolgc Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Romanchek, Gregory R.

Contributors dc:contributor

• Abbaszadeh, Shiva
• Di Fulvio, Angela

• Anomaly Detection
• Radiation
• Statistical Learning
• Machine Learning
• Gaussian Process
• Linear Innovation Sequences

Statement dc:rights

• Copyright 2020 Gregory R Romanchek

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/108545

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F108545