Feature Engineering: A Case Study For Radiation Source Localization In Complicated Environments

Machine learning (ML) can be a powerful data analysis technique; however, many facets must be optimized to reap the greatest benefits. Model selection and hyper-parameter tuning are important areas for optimization, but for certain scenarios, domain-aware feature engineering may lead to the greatest increase in model utility. Feature engineering can also provide insight into the physical domain, and enhance model explainability. This work explores feature engineering for the challenge of source localization in complicated and obstructed environments. Two datasets were simulated with a 60Co source located up to five meters away from a four detector NaI array, one in a modeled city block with multiple concrete obstructions, and one with no obstructions. The ML model is tasked with predicting the angle at which the source is located utilizing only a static measurement from the NaI array. Without the benefit of mobility, the model must make predictions based solely on subtle differences in the counts received in each detector due to relative differences in solid angle and array self-occlusion. The simplest choice of input features (IFs) for this scenario are the total counts received in each detector, but additional features are explored, including counts of photopeak and Compton continuum regions and simple spectral binning schemes. Improved feature engineering led to better angular predictions in previous work with unobstructed scenarios, and separate work has shown that the inclusion of obstructions reduces angular accuracy. This work expands upon previously explored feature engineering to partially overcome this decrease in performance for complex and obstructed environments. Preliminary results with these datasets illustrate the promise of good feature engineering. Using only the total counts as IFs led to an increase in average angular error of 25% with the obstructed dataset compared to the unobstructed dataset. However, using photopeak and Compton continuum regions led to only a 2% increase, and the use of a binning scheme led to only a 4% increase. This work will investigate additional feature engineering options for a variety of complex directional detection scenarios and investigate the statistical correlation between the input features and their effect on the model’s prediction.