Influence of Machine Learning and Gamma-Ray Spectral Parameters on Novelty Detection and Novelty Localization

Gamma-ray spectra are often complex, and extracting information of interest can be challenging. Advanced spectral analysis techniques, such as machine learning, are the subject of current research and development efforts. However, a potential weakness for many of these algorithms is their reliance on supervised learning techniques. This limits predictions to the radioactive sources used for training, thus leading to erroneous classifications of foreign sources. In order to compensate for this deficiency in supervised learning, it is requisite to introduce novelty detection. The objective of this paper is to analyze various spectral and machine learning parameters to determine their influence on novelty detection performance. Gaining such insights will guide the implementation of novelty detection techniques to support supervised classification approaches. A synthetically simulated mono-energetic source was utilized as the spectral anomaly. Performance was based on a one-class novelty detection task, predicting whether a foreign source was in a spectrum. Using a channel-by-channel implementation of a density-based novelty detection algorithm with a sliding variance-window input feature resulted in F1-scores above 0.90 for spectra with 10⁵ counts, and scores above 0.97 across the foreign source energies tested. The channel-by-channel approach with a sliding variance-window input feature provides accurate novelty detection and localization results, thus acquiring information beyond the scope of supervised classification approaches.