Logo of the Physikalisch-Technische Bundesanstalt

Deconvolution algorithm for element reconstruction in neutron resonance radiography

09.07.2014

Within the scope of the joint research project "ACCIS", which is supported by the BMBF, the local concentration of certain elements (C, N, O and H) can be determined by means of an energy-selective radiographic procedure with neutron radiation in order to detect and identify contraband in cargo units (air cargo containers, palette freight, etc.). The element distribution is reconstructed from a complex, multidimensional data set which contains, besides the spatial information of a detected neutron and gamma quantum, also its energy. The element reconstruction is based on a statistical MCMC (Marcov Chain Monte Carlo) procedure. This procedure was implemented on the basis of the software "WinBUGS", and then tested and optimized by means of different experimental and simulated data sets. WINBUGS is frequently used by our department for various unfolding tasks. One of the major advantages of this method, compared to other deconvolution procedures, is that it also allows a quantitative statement to be made with regard to the uncertainty of the detected amount of a substance, so that – based on the ratio of the absolute quantity to its uncertainty – false-positive hits are reliably eliminated (false-positive substances are substances which are not contained in the sample but which – according to the reconstruction of the simulated or of the measured data – are allegedly contained in the sample).

Besides the - not yet completed - analysis of experimental data from neutron transmission measurements of test samples at PTBxs accelerator (PIAF), especially data sets obtained from the numerical simulation of neutron resonance transmission images of air cargo containers by means of the GEANT program were reconstructed in order to assess and to optimize the method. Another significant objective of this work was to determine the detection sensitivity and the false-positive rate at a given neutron fluence for different loads of the container. For a given radiation generator and a defined irradiation geometry, this is directly correlated with the minimum scanning time required for the inspection of a container or a freight unit.

To this end, the simulation results, which first, at a very high neutron fluence, exhibit only low statistical uncertainties, were progressively corrupted by modulating successively an increasing statistical uncertainty onto the transmission spectra, and were then deconvoluted again using the reconstruction algorithm. The results were then compared with the input values. Figure 1 shows the transmission spectrum of a container within a detector cell at a neutron fluence of 9•106 cm-2 in the detector plane (at a distance of 8 m from the source), and a spectrum which results if the sample is irradiated with only 1 % of this fluence.

Figure 2 and Table 1 show the results of the calculation in the case of an air cargo container (LD3 type) filled with cotton and an embedded nitrogen sample (Fig. 2) as well as samples made of other elements (Table 1) at the lowest neutron fluence we calculated, which amounted to a total fluence of 9104 cm-2 in the detector plane. The table shows the deviation of the various reconstructed elemental ratios to the known input values.

An important conclusion that was drawn from these calculations is that the reconstruction procedure is more robust vis-à-vis statistical uncertainties in the transmission spectrum than was expected. The neutron fluence calculated to obtain a detection limit of approx. 300 g of explosives is, hence, approx. 10 times lower than the value obtained by an analytical estimation at the beginning of the project. This estimation would have led not only to borderline irradiation times of 10 minutes (and more) per container, but also to requirements for the radiation generator which would have been close to the limit of the present state of the art.

A limiting factor that must, however, be pointed out is that this optimistic result only applies to freight that does not exhibit too many hydrogen-containing compounds. The last row of Table 1 ("H+cotton") presents results where a hydrogen sample of 10 cm in length with a density of 0.3 g/cm3, embedded in cotton, was investigated. The reconstruction results with up to 16 % deviation from the real value cannot be accepted. In this case, due to the high hydrogen content, the statistics of the transmitted neutron signal are too low to provide useful results. As a consequence, we recommend, for a practical inspection scenario, that the irradiation time is varied in accordance with the transmitted neutron fluence in such a way that transmission spectra are obtained with statistical uncertainties that are still acceptable for the respective detection objective. In practice, this requires a variable scanning speed controlled by the transmitted neutron fluence.

Figure 1: Calculated neutron transmission through an LD3 container filled with cotton and a nitrogen sample, as a function of the energy at varied incident neutron fluence values. In the case of the low fluence, the fluctuations of the transmission are visibly determined by the counting statistics; at high fluence, in contrast, the energy-dependent neutron cross sections of the radiographed substances prevail. The latter bear the information relevant to the element reconstruction.

Figure 2: Frequency distribution of the element ratios at an incident neutron fluence of 9•104 cm-2 at a distance of 8 m from the source (detector plane) - specified (red) and reconstructed after irradiation (blue).

Material C/H N/C O/H N/H O/C
Cotton 1.97 % 0 % 1.20 % 0 % 0.79 %
C+cotton 0.25 % 0 % 10.38 % 0 % 10.29
N+cotton 2.85 % 1.46 % 1.85 % 1.43 % 0.67 %
O+cotton 4.38 % 0 % 2.15 % 0 % 6.26 %
H+cotton 14.73 % 0 % 4.07 % 0 % 16.11 %

Table 1: Relative deviation of the reconstructed element ratios at the lowest calculated neutron fluence of 9•104 cm-2 in the detector plane for the different samples inside the cotton-filled LD3 transport container.