Speaker
Description
Identification of radionuclides is an essential problem especially for border security and field measurements. Today there are two widely used types of methods for identification. They are: classical methods that process separate peaks (so called “peak-by-peak” methods) and newer methods that include different types of machine learning or spectrum convolution (“fit-at-once” methods). Each type has its own pros and cons. Traditional “peak-by-peak” methods consist of two steps: peak searching/filtering and some kind of decision-making sub-algorithm.
In our work we present the method that is a combination of machine learning and peak searching. We perform peak searching as usual, but without peak filtering step and then we use machine learning algorithm as a decision-making tool. Using peak searching as a first step give us more accurate data then the whole unprocessed spectrum and more predictable and controlled behavior of machine learning algorithm. Main advantages of our method are: absence of necessity to train machine learning algorithm on real spectra, we need only library data, and we don’t need to create complicated “if-else”-based decision-making system.
The flowchart of our algorithm is following: searching peaks on the spectrum using spectrum convolution with second derivative of Gauss function; comparison of peaks found with library and creating the peak-presence matrix (at the first variant of our algorithm we use only information about presence of peaks); predicting of the result using the pretrained machine learning k-nearest neighbors-model.
We have tested our algorithm on more than 400 different gamma spectra including special nuclear materials natural occurring, industrial and medical radionuclides. As a result, we have all correct identifications except 6 false-positive results on Potassium-40. For the future work we are going to do the next steps: add more information in the peak matrix (such as peak amplitudes, deltas between library energy and energy measured and so on); collect more spectra for testing; enhance our algorithm with possibility to simultaneously identify more than one radionuclide.
