Blind tests involving 134 luggage items and 8 different plastic explosives have been completed. Spectra of fast neutrons transmitted through the luggage were measured with a 16-detector linear array by time-of-flight techniques as the bag was stepped through the neutron fan beam. Neutron attenuations for each pixel of the bag were least-squares fitted with total cross sections to determine projected number densities of H, C, N, and O and X (everything else). Those numbers plus measured suitcase thickness were used to deduce an explosive probability for each pixel. Individual pixel probabilities were combined through spatial-correlation programs to produce an overall explosive probability for the bag. Additional information concerning goodness of fit, average projected atomic number for each pixel as determined from simultaneous gamma-ray attenuation measurements, total projected atomic number densities, and identification of edges, were used in ambiguous cases to determine whether rescanning was desirable. A discussion of the false-alarm and missed-explosive rates from the tests is presented.

Keywords: explosives detection, neutrons, time-of-flight spectroscopy, HCNO analysis

1. INTRODUCTION

A series of single-blind tests for explosives in luggage was conducted over a two-week interval in September 1996. The purpose of the tests was to evaluate the effectiveness of the Fast-Neutron Transmission Spectroscopy (FNTS) technique and the explosives detection algorithms we have developed. The tests were carried out in cooperation with Federal Aviation Administration (FAA) personnel. They involved 134 different luggage items and 8 different nitrogen-based explosives.

Previous publications^{1, 2, 3} and another paper presented at this conference⁴ describe our FNTS technique. Briefly, attenuation of a neutron energy continuum with energies between 0.5 and 8.2 MeV is measured by pulsed-beam time-of-flight methods at each 3x3 cm² pixel of a suitcase. Measurements are currently made with a neutron fan beam and a 16-detector linear horizontal array. Attenuations are least-squares fitted with total cross sections for H, C, N, and O to determine projected number densities of those elements. Contributions to the attenuation from other elements are lumped together as a fictitious element, X, taken to have an energy-independent cross section of 3 barns.

The detection algorithm is based on the observation⁵ that plastic explosives contain relatively high densities of N and O and low concentrations of H and C. In fact, if local densities of H, C, N, and O could be determined, plastic explosives could be reliably identified. When only projected number densities are obtained because of time constraints, the problem is more difficult. Our solution to the problem involves construction of distributions of projected number densities of H, C, N, O, and X through computer simulations of both explosive and nonexplosive situations. Those distributions are converted to a quantity which we call a B-matrix (for Boom-matrix), whose elements are related to the probability that a pixel of a suitcase is explosive. An attenuation measurement at a given pixel addresses an element of that matrix. Results for all of the pixels in a suitcase are combined to produce an overall explosive probability. Additional information such as an average atomic number for each pixel and a quantitative measure of goodness of fit is available to aid the analysis.

This paper describes the blind-test procedures and the detection algorithms in some detail. We will present examples of measurements and summarize results of the tests. We conclude with suggestions for improvement of the technique.

2. TEST PROCEDURES

Test procedures were prescribed in a formal test protocol. Explosives were delivered to the accelerator laboratory each morning by the City of Eugene Police Department. During the day, explosives were stored in a day-box in the laboratory shop area under the supervision of FAA explosives handlers. FAA personnel prepared the luggage samples and recorded the type, amount, configuration, and location of each explosive insertion. Locations of other items such as toiletries, books, foods, and unusual items were recorded for all luggage items. Since the sample-preparation area was in the shop, FAA personnel had access to a wide variety of metals, chemicals, and unusual materials, which they used freely in loading suitcases. All information concerning luggage content was retained confidentially by FAA personnel until the tests were completed, final decisions concerning explosive content had been made, and a report had been submitted.

Each bag was delivered to a staging area by FAA personnel where its weight, dimensions, and external description were recorded. Most of the bags were of the large international variety, typically with a long dimension of 60-75 cm. Only a few of the bags had a long dimension of 50 cm or less. The bags were both soft- and hard-sided. A number of them were duffel bags and a few were back packs.

Each bag was then carried into the target room and placed on the luggage lift. Care was taken to locate the handle and lower left corner of the bag in a manner consistent with the convention used by FAA personnel to locate contents of the bag. Usually the large dimension of the bag was vertical, and the neutron beam passed through its smallest dimension. Despite care in locating the bags, a few inconsistencies in bag orientations became apparent after the tests.

All personnel then left the target room, the deuteron beam was put on target, and the bag was scanned automatically under computer control. Initially, 80 m C of integrated beam current was used for each bag location. Time-average beam current was about 1 m A, and the beam pulse repetition period was 1.0 m s. For a typical large suitcase, 20-25 slices were required, for a total scan time of 30-40 minutes.

Transmitted spectra were corrected for background, converted to attenuations, and analyzed while data were being accumulated for the succeeding slice. Background spectra were measured by raising the detector array completely out of the penumbra of the neutron fan-beam. They were recorded with no sample in place, and with particle boards 11-cm thick in place on the luggage lift to simulate effects of suitcases. Background spectra were recorded only at the beginning of the two-week test interval, and were used throughout the tests. Sample-out (incident) spectra were measured at the beginning of each day for calibration purposes, and occasionally during the day if beam-pulse tuning required adjustment.

According to the test protocol, a decision as to whether or not a bag was explosive was to be made within eighty minutes. It was possible to rescan portions of a bag, and that was done in a few cases if improved resolution or statistics seemed desirable. In most instances, however, the explosive decision was made within a few minutes of completing a scan. Since the tests were single-blind, FAA personnel could have immediate knowledge of the decision.

After completion of the tests and submission of a preliminary test report to the FAA, we were allowed up to four additional weeks to further analyze the data. We then submitted an additional report specifying the location and extent of each explosive. Relative probabilities of each explosive type were provided as well. For that purpose, samples of the eight pure explosives were scanned near the conclusion of the tests. Results of measurements of elemental composition via FNTS are given in Table 1. The amounts of element X deduced were essentially zero for each explosive.

3. EXPLOSIVES DETECTION ALGORITHMS

The method of constructing the table of probabilities (B-matrix) used to assess the explosive likelihood of each pixel, has been described previously². Briefly, a data base of 78 different items was assembled. A weighted list of 184 items was constructed through multiple entries of those items. A computer program was developed to select various amounts of materials from that list to simulate packing of a pixel. Neutron transmissions through each pixel are calculated from measured transmissions for poorly characterized materials such as foods, or from calculated transmissions for well-characterized materials. Fluctuations due to counting statistics are inserted. We have been using fluctuations characteristic of 30 m C of beam charge to simulate our live-charge measurement conditions. Attenuations, the logarithm of the ratio of incident to transmitted spectrum, are then calculated. Results are unfolded to obtain projected number densities of H, C, N, O, and X. Thickness-dependent cross sections for C, N, and O measured in our suitcase-scanning geometry are used in the calculations.

Results for each element are normalized by the total projected number density to reduce dependence on suitcase thickness. Simulation results are tallied in a binned 5-dimensional space to produce nonexplosive distributions of elements. A 24x36x40x20 set of bins of normalized projected number density, and 6 bins of N+O density are currently used. Those bins are sized to span the entire range of materials encountered. Typically, 10⁹ simulations are done in about three days. Distributions are subsequently smoothed, as described previously².

Explosive distributions of elements are built up by substituting various volume fractions of explosive for benign pixel content. The B-matrix is then constructed by taking the ratio of explosive to explosive plus nonexplosive simulations in each bin. For the preliminary decisions, the six explosives DATB, PETN, RDX, TATB, Tetryl, and TNT were used to construct the B-matrix, The amounts of each were weighted to compensate for their relative destructive power. Destructive power was taken to be proportional to detonation pressure. For subsequent evaluations, the 8 explosives listed in Table 1 were used. Those explosives were equally weighted, and the density of each was assumed to be 1.6 g/cm³.

As the latter B-matrix was being constructed, the number of simulations of each explosive falling in each number-density bin was also tallied. The ratio of the number of occurrences of a particular explosive type to the total number of explosive simulations falling in a bin provides a relative probability of an explosive type. The resulting matrix was used to identify explosives in the blind tests.

A determination of projected H, C, N, O, and X content of a pixel partially addresses an element of the B-matrix. The remaining N+O density address is obtained from measured suitcase thickness and N and O projected number densities. To account for variable packing density, measured thickness is modified in the following way. After a bag is scanned, the total projected number densities are averaged over those pixels above the threshold for assigning a B-value. That threshold is typically set at a projected mass density of 1.6 g/cm² for H+C+N+O, corresponding to a 1-cm thickness of pure explosive. A computed thickness is obtained by multiplying this average by 57 cm/(atom/barn), a value obtained from simulations. The modified thickness used in the analysis is the average of the measured and computed thickness.

The B-matrix address provides a probability that a pixel is explosive. As data are accumulated, a gray-scale map of pixel B-values is produced on-line. Gray-scale maps of total projected number density and average atomic number <Z> for the pixels are also presented. Methods of determining <Z> from measured gamma -ray attenuations are described in another paper presented at this conference⁴.

Our analysis of pixel content is inaccurate if a pixel is only partially filled with material, as can happen at an edge, or if it is filled in a laterally inhomogeneous way. Therefore, after the unfolding is done, each pixel is assumed to be partially empty in 1% increments and the unfolding is repeated until values of reduced chi-square, X²_r, reach a minimum. Pixels which are more than 5% empty are flagged in the B-value maps. This provides a way of indicating a possibly-faulty analysis. Another indication is provided by the values of X²_r themselves. If those values exceed a threshold value of 1.4, the corresponding pixels are flagged in the gray-scale maps of total projected number density and <Z> values.

Figure 1 is a set of gray-scale maps of a bag that had an explosive placed in an iron pipe sloping up to the right. Values of X²_r and <Z> are high in the region of the pipe. The quality of the fit was improved when the cross section for iron was

Figure 1. An example of the gray scale maps used to aid our on-line scanning decisions. The left inset is a map of B-values with numerical values shown at each pixel. Partially filled pixels are indicated by a circular flag. Lighter shading indicates higher values. The center inset 2 is a map of total projected number densities. Numerical values are given in atoms/barn and values of X²_r above 1.4 are indicated by a circular flag. The right inset shows <Z> values for the bag with high values of X²_r again flagged.

included in the unfolding basis set, resulting in the explosive being more easily identified. Locally high values of <Z> and X²_r provide a valuable clue that additional elements should be included in the unfolding. At times, values of X²_r would be generally high throughout the gray-scale maps. That indicated the pulsed deuteron beam had altered its characteristics, and that it should be retuned and a new sample-out spectrum obtained.

Two spatial correlation algorithms are employed to determine how a bag should be classified. The first is a "contiguous pixel" algorithm. Twelve B-value threshold levels are adopted. The number of contiguous pixels above each threshold is counted. Those numbers are converted to explosive probabilities derived from computer-simulated packings of complete suitcases, both with and without explosives. The packing algorithm is described in a third paper⁶ presented at this conference. Only a few contiguous pixels are required to be above a high B-value threshold to yield a high probability. For a low B-value threshold, many pixels are required.

The twelve explosive probabilities, P_i, are combined to produce an overall explosive probability for the suitcase according to the expression:

In this expression, x = n/m, where n is the effective number of independent tests represented by the m = 12 probabilities. We find empirically that n ~1.5. Although P is not strictly a "real world" explosive probability, it is an effective ordering parameter.

At the completion of a suitcase scan, the contiguous pixel algorithm is automatically invoked. The explosive region of highest probability is highlighted in the B-value map, and the overall explosive probability is provided. A threshold probability based upon our previous studies of explosives in luggage was used to aid in deciding whether or not the bag contained an explosive.

The second spatial-correlation algorithm is a "shape test." The algorithm considers 15 sizes of objects, defined by the number of pixels they occupy. A number of differently shaped objects are specified for each size. Shapes involving 1-10 pixels include L-shapes, lines at various orientations, and squares of 4 and 9 pixels. In addition, five larger squares occupying 16, 25, 36, 49, and 100 pixels are considered. The geometric mean of B-values is calculated for each of the shapes and assigned to the central pixel for that shape. The geometric means are termed "C-numbers", for correlation numbers. Each of the various shapes is swept through the entire suitcase. The highest C-number among all shapes for each size is retained, resulting in 15 C-numbers as descriptors of the bag.

Frequency distributions of the fifteen C-numbers for the explosive and non-explosive suitcases that we have examined in the past have been compiled, one for each of the fifteen sizes. The distributions are smoothed and converted to a set of tables of explosive probability versus C-number. The probabilities, P_i, for a scanned suitcase are obtained from these 15 tables and are then combined to yield an overall probability, P, for the suitcase according to the above expression with m = 15 and n = 2.

The three gray-scale maps, the highlighted explosive region, and the two overall explosive probabilities were available immediately upon the conclusion of scan. A decision to declare the suitcase explosive, nonexplosive, or to acquire more data was then made. In several cases, additional scans of several slices with as much as 320 m C of incident beam charge were obtained in regions of high total projected number density, where improved statistics might be desirable for an accurate analysis. In a few other cases, bags were rescanned over limited regions in half-slice increments to improve spatial resolution. In other cases the suitcase was rotated slightly and questionable regions rescanned.

3. TEST RESULTS

We now know that 75 of the 134 bags contained explosives, and of that number, four contained two types. Explosives varied in amount and configuration. Nitroglycerin dynamite and the water gel were invariably in stick form about 40-cm long and 5-cm in diameter with a mass of about 1 kg. In other cases, masses of explosives ranged from 0.12 kg to 1.8 kg. Explosives were in sheets as well as blocks. In most cases, explosives were at least 1 cm thick to the beam, but in several instances thicknesses were less than that. Some of the explosives were placed on edge and were less than one pixel wide.

The initial on-line analysis resulted in six false alarms and six missed explosives. The missed explosives were all 450 grams or less, and less than 1.2-cm thick. In one case the sheet was on edge to the beam. We should note that in other instances we were able to detect explosives of that size. On the basis of the information available to us, it is difficult to determine why false alarms occurred. In several cases, the explosive areas were identified as large sheets, and our algorithms may just need refinement in those cases.

During the three weeks following the tests the data were studied and re-analyzed with the new B-matrix. The following conclusions which are based on those studies should be regarded as tentative. There is a large amount of information to be assimilated and we have not yet thoroughly digested it.

Two lists of bags ordered by their explosive probability were constructed from the contiguous-pixel and shape tests. For the contiguous-pixel list, the data were evaluated and a decision made as to which rescan data and reprocessed data should be used to construct the list. A gap appeared in the ordered list, with no suitcases between 0.17 and 0.28 in explosive probability. The location of the gap is consistent with results from previous data when those data are analyzed with the new B-matrix. An explosive threshold was set at that level, resulting in what we now know are eight missed explosives and eight false alarms. The four bags whose assignment changed from the preliminary one were all close to the threshold.

For the shape-test, bags were ordered by explosive probability after reprocessing with the new B-matrix. All old data were also reprocessed and an explosive threshold of about 0.35 was selected. That value minimized the number of errors for previous data. Then the rescan data were examined, and effects of lowering the projected mass density threshold for assigning B-values to 0.3 g/cm² were considered. As a result of operator judgment, 11 bags were then changed from nonexplosive to explosive, and one was changed from explosive to nonexplosive. The final result was 13 false alarms and 4 missed explosives. Had the original order been retained there would have been 5 false alarms and 7 missed explosives.

During the scanning process we began to suspect that multiple explosives were present in some bags. Upon further study, we identified 12 bags as containing multiple explosives. Three of those cases were indeed correct. The one instance of missed multiple explosives was a bag containing two explosives on edge, close enough together to be unresolved. Many of the errors involving multiple explosives again involved large areas identified as containing sheet explosives.

Attempts to identify explosives were moderately successful. The most likely explosive was correct 38 times; the second most likely was correct 16 times. Many of the errors involved thin explosives or explosives on edge to the beam. It should also be noted that TNT and TNT boosters are similar in composition and often appeared in pairs as the two most probable explosives.

5. CONCLUSIONS

Operator judgment should be reduced or eliminated from the decision-making process. Adaptive reprocessing of data to include elements other than H, C, N, O, and X should be automated. B-matrices should be altered to include the effects of adaptive reprocessing, and automated inclusion of total projected number density and effective Z-values should be incorporated into the detection algorithm. Methods of improving spatial resolution should be considered.

Evidence is emerging that the distribution of elements in bags is different from the distribution of elements in our assumed data base. Accurate distributions of elements should improve explosive detection efficiency. Efforts to accumulate that data base should be continued and expanded.

It is difficult to extrapolate the 10% error rate of these tests to real world experience. The statistical sample in these tests is small, and incorporation in the luggage of unusual materials, such as a lead brick, may skew those statistics.

We have demonstrated that Fast-Neutron Transmission Spectroscopy is useful for explosives detection. Efforts to speed up the process to make it practical should be continued.

5. ACKNOWLEDGMENTS

This material is based upon work supported by the Federal Aviation Administration under Grant No. 94-G-020. The efforts of the technical monitor for the grant, Dr. Curtis Bell, are greatly appreciated. Dr. Bell also served as manager for these tests. J Smith, M. Barrientos, and E. Ocker of the FAA Technical Center served as explosives handlers. Alex Brown, William Fanselow, and Barbara Telecki have contributed substantially to the project.

6. REFERENCES

1. J. C. Overley, M. S. Chmelik, R. J. Rasmussen, R. M. S. Schofield, and H. W. Lefevre, "Detection of Explosives through Fast-Neutron Time-of-Flight Measurements," Report No. DOT/FAA/CT-94/103, 54 pp, (National Technical Information Service, Springfield VA 22161) 1994.

2. J. C. Overley, M. S. Chmelik, R. J. Rasmussen, R. M. S. Schofield, G. E. Sieger, and H. W. Lefevre, "Results of tests for explosives in luggage from fast-neutron time-of-flight transmission measurements," Proceedings of the 5th International Conference on Applications of Nuclear Techniques, Crete, Greece, June 1996. To be published by SPIE, The International Society for Optical Engineering.

3. H. W. Lefevre, R. J. Rasmussen, M. S. Chmelik, R. M. S. Schofield, G. E. Sieger, and J. C. Overley, "Using a fast-neutron spectrometer system to candle luggage for hidden explosives," ibid.

4. H. W. Lefevre, M. S. Chmelik, R. J. Rasmussen, R. M. S. Schofield, and J. C. Overley "Using Fast-Neutron Transmission Spectroscopy (FNTS) to candle luggage for hidden explosives. Can it be made practical for first-line airport use?" Proceedings of the present conference. A more detailed paper describing determination of <Z> has been submitted for puplication in Nucl. Instr. & Meth. in Phys. Research.

5. A. Fainberg, "Explosives Detection for Aviation Security," Science, Vol. 255, pp1531-1537, 1992.

6. G. E. Sieger, "Use of Simulated Suitcases in Refining Explosives Detection Algorithms," Proceedings of the present conference.