Fig 1.
Examples of the Squared Exponential (a), periodic kernel (b), and locally periodic kernel (c).
Fig 2.
Example of a sparse gamma ray spectrum with a weak source peak with centroid at channel 100.
Fig 3.
Input spectrum from Fig 2 with explicitly labelled source counts. The algorithm does not see this labelled information.
Fig 4.
Example spectrum with estimated source counts from the KBGP-NR labelled in blue. These estimated counts are the input into sure KBGP-SE.
Fig 5.
Series of the estimated source distribution after each of 10 collections. Note the source peak (channel 100) raises at a much faster rate than the noise peaks (all other peaks).
Fig 6.
Source estimate after 1 collection.
The estimated source distribution is given in blue, the collected spectrum is given in black, and the true source spectrum is given in green. The collected spectrum is 2 s worth of data.
Fig 7.
Source estimate after 5 collections.
The estimated source distribution is given in blue, the cumulative collected spectrum is given in black, and the true source spectrum is given in green. The collected spectrum is 10 s worth of data.
Fig 8.
Source estimate after 15 collections.
The estimated source distribution is given in blue, the cumulative collected spectrum is given in black, and the true source spectrum is given in green. Note the estimated noise peak as compared to the actual counts in corresponding channels. The collected spectrum is 30 s worth of data.
Fig 9.
The average R2 test results for the 100 trials with source present.
The source peak is increasingly Gaussian with each successive collection whereas the most convincing noise peak is poorly fitted over all collections. While the average R2 value for noise never passes 0.3, this is averaged over all 100 trials, and so some noise peaks do surpass this threshold.
Fig 10.
The average density of the noise and source peaks over the 100 trials with source present.
The density within the source region increases with each successive collection while that of noise regions fall. This shows that the algorithm is correctly placing more counts in the source region with each new collection.
Fig 11.
The average R2 test results for the 100 trials with no source present.
The average R2 value for the best performing noise peak mirrors that of Fig 9, illustrating that the noise peak shape is independent of source presence.
Fig 12.
The average density of the noise peak over the 100 trials with no source present.
While density is accumulating within the peak, this is done very slowly as indicated by the small percent increases with each new collection. This high value make sense as the noise region.
Fig 13.
ROC curve for anomaly detection.
False positive defined as claiming an anomaly is present when none are. True positive defined as claiming an anomaly is present when one is. The labels refer to the threshold set on the R2 test. If the R2 value is above this threshold, then the peak is considered anomalous.