Is the holy grail plastic? Radiation Identification from plastic scintillators

Radiation monitoring at national borders is becoming more common place and so efforts to reduce the burden of false positives (or nuisance alarms) at checkpoint operations without reducing the ability to detect potential threats is now a key objective for many nations. Plastic scintillators are relatively cheap to deploy in large volumes, are rugged and sensitive to a range of gamma energies. However, at present, plastic detectors that are available for deployment for security applications provide only minimal or no energy discrimination as the response to gamma radiation is primarily via Compton scattering so photopeaks are usually not visible. Thousands of shipping containers containing Naturally Occurring Radioactive Materials (NORM) made from ceramics, stoneware and other natural products are transported worldwide on a daily basis. Some of these NORM loads are sufficiently radioactive to trigger alarms from plastic scintillator detectors which have limited ability to also identify the radionuclides present thus necessitating secondary inspection which increases the operational overhead. Previous studies have been carried out to ascertain if radionuclide discrimination using plastic scintillators is possible with a variety of approaches including deconvolution and computer learning. In this paper, a two stage algorithm is described. An example implementation of the algorithm is presented, applied to operational data, and has been installed in real time operation on a polyvinyltoluene (PVT) detector. The approach requires the collection of a large library of spectra using examples of the detectors to be deployed. In this study, data from both actual freight loads passing through a port and predefined freight containing various radionuclides were collected. To ascertain the 'ground truth' content of the freight loads, a sodium iodide detector was deployed along side the PVT detector. A support vector machine (SVM) was used to classify the transformed data into subcategories. This is done by forming a curved surface that separates the principle components when they are plotted in n-dimensions. The library represents freight loads that may contain industrial, medical, nuclear, and NORM radionuclides. The radionuclides in the predefined freight were placed in various orientations and in various amounts of shielding to mimic many different scenarios. Spectra from mixed sources and sources in the presence of NORMs were also taken in typical quantities seen in freight. Steel and aluminium shielding was used in 1mm incremental steps to achieve scattering and shielding effects that may be seen in typical freight loads. Preliminary results on an initial subset of data containing industrial and NORM sources show the number of misclassifications to be less than 1% of the total test data. The training set of data was chosen at random each time the analysis was undertaken, and comprised of approximately 50% of the total data set; the remaining 50% made up the test data. Good initial results were obtained even for low energy radionuclides such as 241Am. Where discrimination is not possible, and principle components overlap, this region or "cloud" of the n-dimensional plot can be put aside. Those spectra that fall in the "cloud" can be regarded as suspect and in these cases, some secondary screening will still be necessary. It is predicted that the algorithm will enable recognition of NORM loads by plastic scintillator detectors to be increased by as much as 80% and for high energy industrial radionuclides to be identified with accuracy approaching 100%. A demonstration system has been produced which provides classification in real time.