Physical Sciences Inc. VG13-158 Utilization of advanced clutter suppression algorithms for improved spectroscopic portal capability against radionuclide threats Bogdan R. Cosofret, Kirill Shokhirev and Phil Mulhall, Andover MA cosofret@psicorp.com David Payne, Bernard Harris and Richard Moro Raytheon Integrated Defense Systems, Tewksbury MA 2013 IEEE Conference on Technologies for Homeland Security (HST 13) 12-14 November 2013 ACKNOWLEDGEMENT: This work has been supported by the US Department of Homeland Security, Domestic Nuclear Detection Office, under competitively awarded contract/iaa HSHQDC - 10 - C - 00171 and HSHQDC-11-C-00117. This support does not constitute an express or implied endorsement on the part of the Government. 20 New England Business Center Andover, MA 01810
Agenda VG13-158 -1 Motivation and General Objectives Overview of PCS Algorithm and Optimization for ASP Experimental Setup Results Conclusions
Motivation Detection of threat R/N sources in moving cargo is difficult due to the need to acquire spectra at short integration time Low SNR regime Poisson noise and clutter mask weak threat signals VG13-158 -2 Advanced Spectroscopic Portal Current systems: PVT-based RPM and NaI-based ASP Capability gaps: RPMs are sensitive, cost effective, but lack energy resolution necessary for threat ID high false warning rates that require secondary screenings PVTs have discrimination capability, but are expensive reduced sensitivity imposes limits on how fast traffic moves through portal
General Objectives Improve overall performance of current ASP systems using advanced algorithms for noise and clutter suppression VG13-158 -3 Demonstrate the achievement of ASP Key Performance Parameters under improved and cost effective operational capability: Utilization of only 4 out of 12 NaI detectors currently integrated with RTN s ASP units Vehicle speeds through the portal in excess of 20 mph (> 6x improvement over current throughput) ASP Key Performance Parameters targeted: False alarm rate of 1 in 1000 occupancies P d,id > 90% for weak activity sources (< 10 µci)
Poisson/Clutter Split Model (PCS): Conceptual Approach and ASP Optimization
GLRT Framework GLRT Methodology for Threat Detection Background Estimation Data Set: Background-only Spectra VG13-158 -5 Likelihood of H 0 Statistical Model (no threat) Likelihood ratio CFAR threshold Detection and ID Algorithm Likelihood of H 1 (threat present) Test Spectra: Background + (Threat Signal) PCS algorithm is based on the GLRT framework, where the background is estimated within the Poisson/Clutter model
PCS Model: Separation of Poisson and Clutter Noise The variability among background radiological spectra can be attributed to two mechanisms: Background clutter, i.e. the changes of the energy-dependent count rate due to variations in isotopic composition depending on particular environments, weather conditions, etc. The random process of radioactive decay, described by Poisson statistics VG13-158 -6 The key innovations behind the Poisson Clutter Split (PCS) algorithm are: The use of a novel probabilistic representation of radiological backgrounds Accurate modeling of gamma counts based on Poisson statistics The use of the Generalized Likelihood Ratio Test (GLRT) to simultaneously perform detection and identification of sources. PCS algorithm s non-linear probabilistic model provides a better characterization of the radiological environment than traditional linear methods
PCS Model: Separation of Poisson and Clutter Noise Observed counts x, obey Poisson statistics corresponding to the local background rate, b, and the integration time, Δt. ~ P( b t) PCS calculates the mean rate as a function of energy and the dominant modes of spectral variations, z k, observed across sampled environments: The underlying rate, b, can be accurately parameterized with a limited number of coefficients which determine the spectral variability of the rate: Clutter is reflected in the varying parameters, w z k capture the spectral features of the environment x b b b({ z 1,.., z }, w) w ( w 1,.., w K ) 1 x D vector, D is the number of channels In the presence of a radioactive source, the background rate, b, is elevated by an energy-dependent contribution from the source: b s b( w) s ( w, ) k p b pb ( w) f ( w) x Test f(w) is the probability distribution of the clutter parameters ~ P( t) VG13-158 -7
PCS-based Background Estimation and Threat Detection and Identification Estimate the background (train) in single or multiple environments VG13-158 -8 N background spectra, { x i }, i = 1,..,N, Estimate the modes of spectral variability and find parameter combinations, w i for each spectrum Fit a probabilistic model to the distribution of w s z k, k 1,.., K distribution of w s Detection and ID: analyze new spectrum x Given new spectrum x, maximize likelihood under two hypotheses: H 0 : x is generated from a rate consistent with the background model H 1 : x is generated from the rate consistent with the estimated background spectrally perturbed by a threat isotope Alarm and ID if likelihood ratio exceeds threshold
PCS Optimization for ASP PCS background model developed using previous ASP field data Background model developed for two operational modes: 1 sec and 0.5 sec integration time Spectra from 60 empty occupancies were used to create background model VG13-158 -9 PCS CFAR threshold determination (Objective 1 in 1000 occupancies): 1000 available no-source occupancies contain ~200,000 1/10 sec live time spectra Processed spectra through PCS and analyzed results Set isotope specific CFAR threshold to be the highest PCS signal value recorded under the 1000 observed occupancies 16 isotopes included in the PCS spectral library
PCS Integration with ASP Real-time PCS software installed on Windows laptop VG13-158 -10 Laptop connected to ASP database server which resides on an Ethernet backbone 32-bit API allows calls across the Ethernet backbone to pull spectra and packet sequence number (PSN) from database Groups of 5 PSN were accumulated to generate 0.5 sec integration time spectra for ingestion into PCS
Experimental Setup
Low Cost Identifier Portal (LCIP): Only 4 out 12 ASP detectors used: Aa3, Ba3, Ca1, Da1 VG13-158 -12 Ca3 Da3 Ca2 Ca1 2.36m 96 Da2 Da1 Aa3 Ba3 Aa2 Aa1 2.30m 69 Ba2 Ba1 Speed (MPH) Observation time (sec) 5 2.6 10 1.3 20 0.65 30 0.43 NaI Detectors (4 x2 x16 ) Neutron Detector Spectra acquired at 0.5 second integration time
Experimental Parameters VG13-158 -13 Vehicle: PSI Truck Check sources emplaced inside truck: (1x) Cs-137 (8 µci) (1x) Ba-133 (7 µci) (1x) Co-57 (4 µci) Interferant: three 50 lb salt bags (~ 40 µci of K-40 signal) Shielding: ¼ steel cap 30% reduction in peak count for Cs-137 50% reduction in peak count for Ba-133 67% reduction in peak count for Co-57 Salt bags Steel Cap
Source Locations LCIP Evaluation: Check Source Locations Inside Vehicle VG13-158 -14 12' Shielding (0.25 steel cap) (when used) 29 49 Salt Bags 40 3x50 lbs Salt bags Positions for Cs-137 and Ba-133 used during multi-source runs. Co-57 was located on second stand
Truck Through Portal at 30 mph VG13-158 -15
Background Only (30 mph) 150 lbs of Salt Inside the Truck VG13-158 -16 Several runs through the portal were conducted without sources present Weak PCS responses observed for all 16 isotopes in the library No PCS responses exceeded the CFAR (1 in 1000 occ) isotope specific thresholds No false alarms were reported Continuous acquisition of spectra over ~ 2 hrs also yielded no false alarms
Truck with Unshielded Cs-137 (8 µci) + Salt Vehicle Speed: 30 mph VG13-158 -17 PCS results with 4/12 NaI detectors: P d,id = 95% against unshielded Cs-137 at 30 mph (detected/id in 18 out of 19 runs), CFAR = 1 in 1000 occ. No false alarms or mis-identifications were reported Note: Standard ASP software using with all 12/12 NaI detectors yielded P d,id = 10% (2 out of 19 runs)
Truck with Shielded Cs-137 (attenuated 8 µci) + Salt Vehicle Speeds: 20 and 30 mph VG13-158 -18 20 mph 30 mph PCS results with 4/12 detectors: P d,id = 93% at CFAR of 1 in 1000 occ. against shielded Cs-137 at 20-30 mph (detected in 14 out of 15 runs) No false alarms or mis-identifications were reported Note: Standard ASP software with all 12/12 ASP detectors yielded P d,id = 0%
Truck with Unshielded Ba-133 (7 µci) + Salt Vehicle Speed: 30 mph VG13-158 -19 PCS results with 4/12 NaI detectors: P d,id = 93% against unshielded Ba-133 at 30 mph (detected/id in 14 out of 15 runs), CFAR 1 in 1000 occ. No false alarms or mis-identifications reported Ba-133 presence leads to correlated I-131 PCS responses, but not strong enough to exceed the I-131 isotope specific threshold. Cross-talk also addressed using Dominant PCS Note: Standard ASP software using all 12/12 NaI detectors yielded P d,id = 0%
Truck with Shielded Ba-133 (attenuated 7 µci) + Salt Vehicle Speed: 20 mph VG13-158 -20 PCS results with 4/12 NaI detectors: P d,id = 86% against shielded Ba-133 at 20 mph (detected/id in 12 out of 14 runs), CFAR 1 in 1000 occ. No false alarms or mis-identifications were reported Note: Standard ASP software using all 12/12 NaI detectors yielded P d,id = 0%
Truck with Unshielded Multiple Sources (Co-57, Ba-133, Cs-137) Vehicle Speed: 20 mph VG13-158 -21 PCS Results with 4/12 NaI detectors: P d,id (Cs-137) = 100%, P d,id (Ba-133) = 100%, P d,id (Co-57, 4 µci) = 85% at 20 mph when all sources inside the truck Note: Standard ASP software using all 12/12 NaI detectors yielded P d,id (Cs-137) = 30%, P d,id (Ba-133) = 0%, P d,id (Co-57) = 0%
Conclusions
Conclusions and Next Steps Demonstrated P d,id = 100%, CFAR = 1 in 1000 occupancies at 20 mph VG13-158 -23 Successfully demonstrated P d,id > 90%, CFAR = 1 in 1000 occupancies at 30 mph Successfully demonstrated isotope identification/discrimination capability with no reported false alarms or mis-identifications Demonstrated the ability to detect shielded check sources Next Steps: Integrate C version of PCS (demonstrated < 100 msec/spectrum processing time with 28-isotope library)