Multifaceted Radiation Detection and Classification System

Abstract

To provide a system of identifying a source of radiation.SOLUTION: The system includes a radiation source detector and a radiation source identifier. The radiation source detector receives measurements of radiation; for one or more sources, generates detection metrics indicating whether the sources are present in the measurements; and evaluates the detection metrics to detect whether a source is present in the measurements. When the presence of a source in the measurements is detected, the radiation source identifier generates, for one or more sources, identification metrics indicating whether the sources are present in the measurements, generates a null-hypothesis metric indicating whether no source is present in the measurements; and evaluates the one or more identification metrics and the one or more null-hypothesis metrics to identify the source, if any, which is present in the measurements.SELECTED DRAWING: Figure 9

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This patent document claims the benefit of priority to U.S. Provisional Patent Application No. 62 / 673,750, filed May 18, 2018, and entitled "Radiation Detection Analysis Kit (RDAK)," and No. 62 / 805,825, filed February 14, 2019, and entitled "Radiation Detection Analysis Kit (RDAK)," both of which are incorporated herein by reference.

[0002] STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT The United States Government has rights in this invention pursuant to Contract No. DE-AC52-07NA27344 between the United States Department of Energy and Lawrence Livermore National Security, LLC for the operation of the Lawrence Livermore National Laboratory. [Background technology]

[0003] Detecting radioactive sources is particularly important for safeguarding nuclear materials to ensure the safety of the general population, military personnel, and first responders. Radioactive sources are classified into source classes based on the use of the radiation. For example, radioactive materials are dispersed among weapons (e.g., dirty bombs) and radiopharmaceuticals (e.g., in medical patients). 99mThe radiation source classes for radioactive materials are considered to be different for radioactive materials (e.g., 1.2Tc). The radiation source classes for many sources (e.g., medical patients) correspond to uses that are approved because any safety risks are deemed acceptable, while the radiation source classes for other sources (e.g., dirty bombs) correspond to uses that are not approved because the safety risks are deemed unacceptable. Since some uses are approved and others are not approved, simply detecting the presence of a radioactive source is of little use. Thus, one purpose of radiation source detection is to distinguish between radiation source classes that may present an unacceptable safety risk and those that do not present such a safety risk.

[0004] To aid in radiation detection, a radiation detector, also called a detector (i.e., a physical device that detects photons), is used to collect the number of photons having energies within a radiation spectrum. A radiation detector can only distinguish the energy of a photon to a certain level of accuracy. As a result, the radiation detector can categorize the photon counts into ranges of energy. The radiation detector collects the counts over a period of time (e.g., 1 second) and can indicate the counts within each energy range as a measurement of radiation (i.e., a radiation spectrum).

[0005] To detect radiation sources over a large area (e.g., a city), a drone equipped with a radiation detector can be flown over the area in a measurement collection pattern. The detector can also be run on a vehicle, carried by hand, or placed in strategic fixed locations. These measurements can be provided to an analysis system to identify whether there is a radiation source of interest, which may be a source that presents an acceptable or unacceptable safety risk. Since different areas may have significantly different background radiations (e.g., due to different topography or building materials), measurements representing the same radiation source may be significantly different in different areas. As a result, such analysis systems must take into account the background radiation present in the area where the measurements were collected. Summary of the Invention [Problem to be solved by the invention]

[0006] The speed and accuracy of detecting and classifying radioactive sources (e.g., as an acceptable or unacceptable safety risk) is important. Also, the nuclide that is the source (e.g., U or 137 It is also important to identify the radioactive material (Cs) from the radioactive material. However, some state-of-the-art analysis systems require huge amounts of computational resources to process the measurements, identify background radiation, identify the source, and classify the source class of the source. In addition, many of the currently deployed analysis systems employ detection algorithms that are not very accurate. Therefore, the use of such analysis systems is not practical due to the required computational resources or low accuracy. There is a strong need for algorithms that are both computationally efficient and accurate. [Brief description of the drawings]

[0007] [Figure 1] FIG. 1 is a block diagram illustrating the overall architecture of an RDA system in some embodiments. [Diagram 2] FIG. 2 is a block diagram illustrating components of a radiation source detector of an RDA system in some embodiments. [Diagram 3] FIG. 2 is a block diagram illustrating components of a radiation source identifier of an RDA system in some embodiments. [Figure 4] 1 is a graph showing detection efficiency. [Diagram 5] FIG. 13 illustrates calibration of measurements based on a new binning structure. [Figure 6] FIG. 2 is a flow diagram illustrating high-level processing of the OSP matched filter component of the RDA system in some embodiments. [Figure 7] FIG. 2 is a flow diagram illustrating the process of the calibration component of the RDA system in some embodiments. [Figure 8]FIG. 11 is a flow diagram that illustrates the processing of the binning component of the RDA system in some embodiments. [Figure 9] FIG. 2 is a flow diagram illustrating a background estimation process based on the rank average component of the RDA system in some embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] A method and system for detecting and classifying radioactive sources is provided. A radiation detection and analysis ("RDA") system processes a set of radiation measurements (i.e., gamma ray spectrum data and neutron counts) and analyzes these measurements in real time to detect the presence of a radioactive source and classify the radioactive source (e.g., U or 137 Cs) to determine the source class of the radioactive source. The RDA system provides a radioactive source detector component ("Radiation Source Detector") that can operate on a wide range of detector types, including different materials and detector sizes, using multiple detection algorithms. Each detection algorithm generates one or more detection metrics that are updated with each new measurement in a series of measurements. The detection metrics optimize the signal-to-noise ratio under various conditions. The Radioactive Source Detector employs various background estimation algorithms to estimate background radiation ("background") so that the contribution of the radioactive source to the measurements can be determined. The Radioactive Source Detector continually retrains the background estimation algorithms based on background measurements (i.e., measurements where no radioactive source is detected). The RDA system determines the source class of the radioactive source (e.g., 137 Cs) and a radiation source identifier component ("radiation source identifier") that identifies the radiation source class. Once a radiation source class of interest (which may be an approved or unapproved use) is identified, the RDA system may output an alarm indicating the radiation source class.

[0009] The RDA system can be used in fixed or mobile detection applications. Not all algorithms offered by the RDA system are suitable for all detection applications. The RDA system provides an architecture that allows the selection of different combinations of detection and background estimation algorithms that optimize performance. Various combinations of algorithms can be used for sample measurements (real or simulated). For each application, the combination of algorithms that best suits that application can be selected. The most suitable combination can then be used within the RDA system to operate in real time with high accuracy. Additionally, the RDA system can also be used to process previously collected measurements much faster than they would be processed in real time.

[0010] 1 is a block diagram illustrating the overall architecture of an RDA system in accordance with some embodiments. The RDA system 100 includes a radiation source detector 110 and a radiation source identifier 120.

[0011] The radiation source detector inputs measurements and outputs a source detection flag for each measurement (or group of measurements) indicating whether a source is present in the measurements. The radiation source detector includes a calibrate measurements component 111, a calculate metrics component 112, an evaluate metric component 113, an estimate background component 114, and a detection source definitions store 115. The measurement calibration component calibrates the measurements to account for drift and non-linearity of the detector's sensors. The metric calculation component calculates metrics for the measurements to indicate whether any source is present in the measurements as defined by the source definition. The source definition of a source includes a source signature (e.g., spectral shape) of the source that takes into account specified shielding, such as shielding type and shielding thickness. Thus, a radiation source can be associated with a source definition for each specified shielding. A source definition can include multiple sources. Some algorithms operate without a source definition to detect any anomalies different from the background. A metric evaluation component sets a source detection flag individually or collectively based on an evaluation of the metrics to determine whether a radioactive source may be present in the measurements. A background estimation component updates the estimated background based on measurements where the presence of a radioactive source was not detected. A metric calculation component takes the estimated background into account when calculating the metrics. A detection source definition store includes source definitions that define source signatures of radioactive sources to be detected.

[0012] The radiation source identifier identifies the source type and its source class when the radiation source detector detects the presence of a source. The radiation source identifier 120 includes a metric calculation component 121, a source identification component 122, and an identification source definition store 123. The metric calculation component calculates metrics similar to the metric calculation component of the radiation source detector, but may use algorithms that are particularly effective for source identification. The source identification component identifies the source based on evaluation of the metrics. The source identification component may output a probability of each radiation source class. The identification source definition store includes source definitions for the sources to be identified. The identification source definition store includes a more comprehensive set of source definitions than those of the detection source definition store to increase the accuracy of source identification.

[0013] The RDA system inputs spectral measurements represented as a spectral histogram representing a detection energy range (or spectrum) that includes a fixed number of energy bins, each of which represents an energy range within the detection energy range. For example, if the energy range is 100 keV to 1 MeV, the histogram may have 50 energy bins, each of which represents an energy range of 20 keV. A first energy bin has an energy range of 100 keV to 120 keV, a second energy bin has an energy range of 120 keV to 140 keV, ..., up to 980 keV to 1 MeV. The histogram has energy edges at 100 keV, 120 keV, 140 keV, ..., and 1 MeV. Thus, there is one more energy edge than there are energy bins. Each energy bin contains the number of detected photons that were present within the energy range of that energy bin over one measurement interval (e.g., one second).

[0014] The following table provides an explanation of the terms used to describe the RDA system. [Table 1]

[0015] FIG. 2 is a block diagram illustrating the components of the radiation source detector of the RDA system in some embodiments in more detail than 110 in FIG. 1. The radiation source detector 200 inputs measurements, detects whether a source may be present in the measurement, and outputs a source detection flag. However, the radiation source detector does not perform the computationally expensive process of actually identifying the source. The radiation source detector includes a measurement calibration component 201, a generate metric component 202, an aggregate metrics component 203, a metric evaluation component 204, a train metric component 205, a veto component 206, a background estimation component 207, and a generate calibration component 208. The radiation source detector also includes a detected source definition store 211 and a history store 212. Each component may also include a store for locally storing data generated by the component or received from other components.

[0016] The measurement calibration component generates calibrated measurements to account for drift and nonlinearity in the detector measurements using a rebinning process. The measurement calibration component inputs measurements and calibration data and outputs calibrated measurements. The calibration data indicates how to rebin the measurements based on the detector drift.

[0017] The metric generation component can include metric generation components for different types of detection algorithms. The metric generation component inputs calibration measurements, optionally navigation information (e.g., GPS information to identify the presence of buildings), and coefficient sets for the source signatures, and generates a metric for each source signature that indicates whether the measurements (i.e., calibration measurements) represent the presence of the source of that source signature, or the presence of any source that is not background. Each coefficient set is associated with a combination of a source signature and a detection algorithm. For example, a first coefficient set does not include occlusion and an orthonormal subspace projection ("OSP") matched filter (described below). 137 The second coefficient set may be a source signature of 10 cm lead shielding and an OSP matched filter. 137 The first coefficient set may be a source signature of Cs, and the third coefficient set may be a source signature of U that does not include shielding and a likelihood ratio test ("LRT") matched filter (described below). A metric generation component for an associated detection algorithm uses these coefficients to generate a metric for each set of coefficients that is a semi-definitive indication that the source represented by the source signature is present. The source identifier provides a more definitive indication of each source. Continuing with this example, the detection algorithm for the matched filter may be a source signature of U that does not include shielding. 137 A set of coefficients for the Cs source signature is applied to generate the first metric, including 10 cm lead shielding and the OSP matched filter. 137 A set of coefficients for the source signature of Cs is applied to generate a second metric. Each metric can be expressed as a numerator and denominator pair called metric partials. For example, a metric that does not include occlusion for a particular detection algorithm can be expressed as a metric that does not include occlusion for a particular detection algorithm. 137The metric for the Cs source signature can be (1,5). The detection algorithm is primarily described as a matched filter, but the detection algorithm can also be an anomaly detector. The anomaly detector produces a metric that indicates whether the deviation of the measurement from the estimated background may be sufficient to indicate the presence of a source, regardless of what the source is.

[0018] The metric aggregation component generates an aggregated metric for each combination of source signature, detection algorithm, and window of measurements (e.g., a rolling window). Each window includes the number of most recent measurements, such as 1, 2, and 4. For example, a window that does not include occlusions may be generated. 137 The metric partials for a particular detection algorithm for a source signature of Cs and four consecutive measurements may be (1,4), (3,5), (1,2), and (3,5). To aggregate the metrics for a window for the source signature, the metric aggregation component sums the numerators of the metric partials, sums the denominators, and sets the aggregate metric to the sum of the numerators divided by the square root of the sum of the denominators. Continuing with this example, an individual metric partial (1,5) would result in a time window of 1, and a metric of 1 divided by the square root of 4 would be 0.5. If the first two metric partials are aggregated over a time window of 2, the sum of the numerators is 4 and the sum of the denominators is 9, which multiplied by the square root of 9 is 1.33. If all four metric partials are aggregated over a time window of 4, the sum of the numerators is 8 and the sum of the denominators is 16, which divided by the square root of 16 gives an aggregate metric of 2. Thus, no occlusion is involved. 137 For the specific detection algorithm for Cs, and for windows containing 1, 2 and 4 measurements, the aggregation metrics are 0.5, 1.33 and 2. Aggregation can also be done based on measurements taken at the same location, even if at different times.

[0019] The metric evaluation component inputs the aggregate metrics and outputs a single decision metric that indicates whether a source has been detected. First, the metric evaluation component normalizes each aggregate metric by subtracting the measured mean and dividing by the measured standard deviation. Then, it determines a decision metric from an evaluation of all the normalized aggregate metrics, including different time windows, different algorithms, and different sources. Typically, this evaluation uses the maximum value of each normalized aggregate metric, and may also use weighting factors determined based on performance analysis of different methods or using machine learning as described below.

[0020] The metric evaluation component then compares the decision metric to a number of thresholds set to control the operation of the radiation source detector. One such threshold is a source present threshold. If the decision metric exceeds the source present threshold, the metric evaluation component sets a source detection flag (i.e., true or false). The source detection flag triggers the radiation source identifier. The source present threshold can be set to ensure a certain percentage of missed source detections (false positives or false alarms) when the source is not present, such as a false alarm every 8 hours. The certain percentage can be based on a period (e.g., 8 hours), a measurement rate, a number of rolling windows, and a number of metrics. As another alternative, the metric evaluation component can calculate the decision metric using a classifier generated via a machine learning algorithm (e.g., a support vector machine, a Bayesian classifier, and a neural network). The classifier can be trained using training data including a feature vector and a label. Features of the feature vector can include a detection algorithm, a source, a window size, and an aggregation metric. The feature vector can be generated based on data collected and generated while performing actual detection of different sources. The feature vector can include 10 cm of iron shielding. 137Cs, and the angular velocity between the source and detector (identified, for example, based on GPS readings) are different. 137 The labels may also be generated using a mathematical model that generates simulated measurements of different types of sources, such as Cs. The labels may be an indication (e.g., true or false) of whether the feature vector represents the presence or possible presence of a source. The labels may be generated manually or based on data generated when the radiation source identifier identifies radiation sources and source classes.

[0021] The metric evaluation component also compares the decision metric to an extended aggregation threshold. If this threshold is exceeded, the source identifier is triggered to aggregate until the decision metric falls below an end aggregation threshold. The decision metric can also be compared to a veto threshold, which is typically set very low. If the decision metric exceeds the veto threshold, a lockout flag is sent to the veto component 206 that controls whether the measurement is used for calibration or background estimation.

[0022] The background estimation component may include a background estimation component for each background algorithm. The background estimation component inputs measurements to generate background statistics for the background estimation. Each background algorithm may generate different statistics, such as the average number in the measurements that represents the background (called a rate), the average measurement of the background, or the basis of the background. The detection algorithms may employ different background statistics. For example, one detection algorithm may employ a rate statistic, while another may employ a rate statistic and a basis statistic. The background estimation component may generate a statistic that is a weighted sum of the new statistic and the previous statistic. For example, if the new rate is 160, the previous rate is 200, and the weight is 0.25, then the weighted rate to use as the current rate would be 190 (i.e., 200*0.75+160*0.25). Thus, the weight controls how quickly the estimate adjusts based on the new statistic. The background estimation component can generate background statistics periodically (eg, every 5 minutes) or based on an analysis that the current background is significantly different from a previous background.

[0023] The reject component provides uncalibrated measurements to the calibrated measurements generation component and calibrated measurements to the background estimation component. Meanwhile, the reject component filters out measurements that should not be used to estimate the background. For example, each measurement used to generate the aggregate metric that is above the reject threshold can be filtered out. Measurements before and after these measurements can contribute to a source. The reject component can filter out measurements within an extended window (fixed or variable) before and after the measurement used to generate the aggregate metric to remove this contribution from the background. Thus, the extended window can be used to filter out measurements that may be influenced by a source but are not included in the aggregate metric. These measurements would otherwise be considered in the background calculation, making the estimated background less likely to be influenced by a source. The reject threshold can be set to favor a certain percentage of missed vetoes, such as one missed veto every 10 minutes. This certain percentage can be based on a period (e.g., 10 minutes) and a measurement rate.

[0024] The metric training component inputs the source signature and background statistics to generate a coefficient set for a combination of the source signature and the detection algorithm. However, a coefficient set may not be generated for each combination. For example, a detection algorithm may 137 It is particularly effective at identifying Cs, but may not be effective at identifying U. In such cases, the detection algorithm: 137 Generate a coefficient set for Cs and none for U.

[0025] The history store contains each measurement and calibration measurement. The detected source definition store contains a source definition for each source (eg, one per shielding) whose presence is to be detected.

[0026] 3 is a block diagram illustrating components of a radiation source identifier of an RDA system in some embodiments. When the radiation source detector detects the possible presence of a source, the radiation source identifier is executed to identify the source (e.g., 137 Cs) and the source class that is the main contributor to the measured non-background photons. The source identifier inputs the measurements and background statistics used by the source detector to detect the presence of the source. The measurements may be the sum of the measurements contained in each window. For example, if a window contains 512 measurements, the source identifier sums these measurements to give a single total measurement. The source identifier outputs an indication of the sources present. For each source, the source identifier may output a probability that that source is present. For example, the source identifier may: 137 The source identifier may output a probability of 0.67 for Cs and a probability of 0.11 for U. The radiation source identifier also applies radiation source class rules to identify the radiation source class of the source. For example, the radiation source classes of the source may be medical, industrial, and fission, etc. The radiation source classes may also be represented by probabilities. The radiation source identifier may include a measurement calibration component 313, a metric generation component 301, a metric training component 302, a select source metrics component 303, a generate null-hypothesis metric component 304, a generate source scores component 305, and an identify radiation source class component 306. The radiation source identifier also includes an identification source definitions store 311 and a radiation source class rules store 312.

[0027] The metric calibration component of the source classifier inputs the measurements used by the radiation source detector to detect the source. The metric calibration component aligns the measurement bin range with the source signature bin range. The metric calibration component can shift either the measurement bin range or the source signature bin range. Shifting the measurement bin range induces correlations between bins that need to be tracked to ensure statistical accuracy. Shifting the source definition increases statistical accuracy but requires more computational resources as each source signature must be shifted.

[0028] The discrimination source definition store is similar to the detection source definition store, but has a more comprehensive set of source definitions than the detection source definition store. This set is more comprehensive to take into account the detection efficiency of the source definitions. Detection efficiency refers to the likelihood that a source signature is effective in detecting a source despite shielding. For example, a source signature for a source with 5 cm of lead shielding may have a detection efficiency of 0.25, while a source signature for a source with 7 cm of lead shielding may have a detection efficiency of 0.15. The goal is to have source signatures for enough shielding configurations to minimize shielding configurations with detection efficiency below the maximum possible detection efficiency. Maximizing overall performance requires a tradeoff between detection efficiency across all shielding configurations of interest and false alarm opportunities, which increase with the number of signatures in use.

[0029] In some embodiments, it is useful to capture signatures observed in the field. A form of semi-supervised learning can be used to extract significant detections from measurements that have spectra similar to, but not as consistent as expected with, the source signature to create a new source signature, which is added to the source signature store. The detection efficiency of the extracted spectrum is compared to the detection efficiency of the existing source signatures in the source signature store to determine whether the measured spectrum is adequately represented by the source signatures in the source signature store. This method can increase sensitivity to sources under test that are intermediate between the shielding configurations used in the source signature store and unexpected source combinations, or that have some differences from the source signatures in the source signature store. This new source signature can be added locally to the source signature store of the detector in use, or can be distributed to source signature stores used by other detectors of the same type as needed.

[0030] For the purpose of discrimination, it is assumed that the source has already been detected and the false alarm rate is no longer taken into account. Thus, the limit on the number of source signatures for discrimination is not so critical, but is limited only by the computational cost. Figure 4 is a graph showing the source discrimination efficiency. The graph 400 includes an x-axis representing thickness and a y-axis representing detection efficiency. The solid line 401 represents the discrimination efficiency of the source signature without shielding. The discrimination efficiency has a maximum value of 0.5 at a thickness of 0 cm and drops to 0.0 at a thickness of 5 cm, i.e. the source shielded by 5 cm is not discriminated by the source signature without shielding. The solid lines 402 and 403 show the discrimination efficiency of the source signatures of 5 cm and 7 cm, respectively. The dotted line 404 connects the detection efficiencies of these thicknesses to the source signatures. The regions 405 and 406 represent the gap between the maximum discrimination efficiency and the actual discrimination efficiency. Such a gap may cause a source to be misidentified. The area of ​​the gap can be reduced by adding a source definition with a larger shielding thickness, for example gap 405 is reduced by adding a source definition with a thickness of 3 cm.

[0031] The metric generation component includes a metric generation component for each discrimination algorithm used by the radiation source discriminator. In some embodiments, only one discrimination algorithm is used, such as one similar to the OSP matched filter, but with a different regularization term to account for the difference between detection efficiency and discrimination efficiency. The metric training component can be similar to the metric trainer component of the radiation source detector.

[0032] The source metric selection component inputs metrics, identifies the best metrics for each source, and outputs these metrics as source metrics. 137 Cs, and has 1cm and 2cm shielding 137 The metrics for Cs can be 1.0, 2.0, and 0.5, respectively. In such cases, the source metric selection component is 137 Select 2.0 as the source metric for Cs.

[0033] The null hypothesis metric generation component inputs the measurements, background statistics, and source metrics and generates a null hypothesis metric that represents the statistical worst case scenario for the background. The null hypothesis metric generation component calculates the variance of each channel of the measurements and the expected background. The null hypothesis metric generation component also calculates the variance of each metric and the covariance of the source metrics. The null hypothesis metric generation component calculates a null hypothesis score based on the variance and covariance of the source metrics and outputs the null hypothesis variance (e.g., the average of the variances of the metrics) and the null hypothesis score.

[0034] The source score generation component inputs the target metrics (i.e., source metrics and null hypothesis metrics) and the initial probability (prior probability) of each target (i.e., source and background). The score generation component calculates the probability of each target given a measurement that is the maximum of the probabilities that the target's metric is greater than the target metrics of each of the other targets. The maximum probability can be based on a multivariate normal distribution centered on the measurement given the measurement and the variance of each target metric. The source score generation component generates the target probability (posterior probability) for each target using, for example, a Bayesian estimator to generate the target probabilities.

[0035] The radiation source class identification component inputs the target probabilities and applies the radiation source class rules to generate a radiation source class probability for each radiation source class. The radiation source class rule store contains the radiation source class rules. Alternatively, the radiation source class identification component can select for each radiation source class the maximum radiation source class probability that the target is of that radiation source class. The maximum radiation source class probability can be based on a multivariate normal distribution and variance for that radiation source class, similar to the target probabilities.

[0036] Detection Algorithm The RDA system supports the use of various combinations of detection algorithms, metric trainers, and aggregation methods. Supporting such combinations allows the RDA system to be used in, and tailored to, a wide range of applications.

[0037] The RDA system can be used with projection-based and quadratic-based detection algorithms. A projection-based detection algorithm can be decomposed into a set of linear operations, each of which operates on the measurements individually. For example, the detection algorithm can generate a metric by applying a linear transformation to a sum of measurements over a period of time to calculate a numerator and denominator as represented by the following equation:

number

number

number

[0038] In a quadratic based detection algorithm, the numerator or denominator is generated as a product of measurements through a transformation matrix as represented by the following equation:

number

number

[0039] Orthonormal Subspace Projection ("OSP") Matched Filter The OSP matched filter metric trainer constructs an optimal multiplicative spectral filter that best rejects the background. The metric trainer uses an estimated background basis function B that is an orthonormal subspace projection. This projection is maximal onto the source signature S with a weighting matrix W that is based on the estimated background. The metric produced by the OSP matched filter has a Gaussian distribution and can be normalized to have zero mean and unit variance. The OSP matched filter metric can be expressed by the following equation:

number

[0040] The metric trainer uses the source signature S i , estimated background Input the image, JPEG0007660169000008.jpg6150, and the background basis vectors. The estimated background can be calculated using an exponential smoothing function. The weighting matrix can be calculated by dividing the previous background measurements B. The metric trainer can calculate the optimal projection vector T as follows: 1) An estimated background where each energy bin variance is over 1 + a scalar over the square root of the expected background Calculate the inverse of the diagonal of the weighting matrix W from JPEG0007660169000009.jpg6150. 2) Calculate the weighted background. B W =(B t WB) -1 B t W 3) Weight the source signature by a weighting matrix. S t W 4) Calculate the weighted projection of the source signature. S t WB 5) Calculate the background weighted projection. S t WB(B W ) 6) Set the optimal projection vector T to the difference between the source signature and the background weighted projection. 7) Set the variance scalar V to the square of the optimal projection vector T times the sum of the estimated background. JPEG0007660169000010.jpg141508) Divide the terms of the optimal projection vector T by the square root of the expected variance so that the expected variance becomes 1. JPEG0007660169000011.jpg6150

[0041] The performance characteristics of the OSP matched filter can be modeled using the following equations:

number

[0042] Hybrid Matched Filter A hybrid matched filter combines the properties of a matched filter with those of gross count type algorithms. This combination can allow for more effective detection when the systematic and statistical noises are similar to each other on the timescale of the maximum integration. Such similarity between systematic and statistical noise is often found in small detectors (e.g. pagers).

[0043] The hybrid matched filter does not remove all background related features, but rather removes those parts that do not correspond to the source signature. Because some of the background is retained, the hybrid matched filter estimates and removes the bias. The hybrid matched filter metric trainer generates a transformation T as represented by the following equation:

number

number

number

number

[0044] The performance characteristics of a matched filter can be modeled using the following equations:

number

[0045] Two-Dimensional ("BD") Matched Filters The BD matched filter helps to address a problem that occurs when detection algorithms use background rates. If the background rate suddenly increases, the detection algorithm may underestimate the background. This results in a higher mean for the metric as this underestimated background is subtracted from the measurement. The variance of the metric also increases as there is more Poisson noise present as the background increases. Thus, measurements with increased background are more likely to cause the metric to exceed the threshold and produce a false positive.

[0046] The BD matched filter, like the hybrid matched filter, allows the strength of the measurement over the background (i.e., the number of counts) to contribute to changes in the metric and measurement shape. The BD matched filter introduces a nonlinear element to limit the effect of this contribution. The BD matched filter balances these nonlinear elements such that an increase in background reduces the metric, so that the chance of a false positive is constant regardless of condition.

[0047] The BD matched filter algorithm does not use oblique OSP. Rather, it calculates the total counts in the source signature and each background basis vector that represents the background. The decomposition uses both the measured and expected background measurements. The BD matched filter is two-dimensional in the sense that it produces both an estimated source and an estimated background for the measurement.

[0048] Since the estimated background of a measurement is based on both the measurement and the estimated background, the estimated background will be below the actual background. Therefore, the BD matched filter algorithm calculates the penalty by applying a penalty function based on the difference between the estimated background and the expected background. The penalty tends to balance the increase in the mean and variance of the metric. However, when the estimated background is below the expected background, the metric is less likely to produce false positives, and therefore no penalty is necessary. When using a linear penalty function, the noise in the estimated background is amplified whenever the estimated background is low. The nonlinear penalty function allows for a constant false positive rate at the estimated background. Furthermore, unlike the OSP matched filter, the estimated variance is calculated using the expected background. Thus, the metric is JPEG0007660169000020.jpg10150 rather JPEG0007660169000021.jpg10150. This results in a higher detection count per unit for the BD matched filter, and therefore a higher sensitivity.

[0049] The BD matched filter can be based on the partitioned matrix problem as represented by the following equation:

number

[0050] Solving this partition matrix problem directly can be computationally expensive since the vector projections of each of the source and background components must be calculated at every time step. This partition matrix problem can be solved by decomposing it into a problem whose solution is expressed by the following equation:

number

[0051] Once the estimated source counts and the estimated background counts have been generated, the metric of the DB matched filter can be calculated as represented by the following equation:

number

[0052] As with other matched filter algorithms, the numerator and denominator terms can be summed and integrated over multiple time segments. The penalty can be calculated so that the total integrated distribution is constant for different background strength and background estimate term conditions. Because the distribution function is not linear or fixed as a function of a threshold, the slope can be chosen so that the distribution integral never exceeds an expected value. The slight over penalty this produces may reduce overall sensitivity, but this sensitivity loss can be at least as good as the sensitivity gained from the use of the estimated background in the denominator term.

[0053] Likelihood Ratio Test ("LRT") Matched Filter The LRT matched filter may employ the same metric evaluator as the hybrid matched filter. The LRT matched filter uses Poisson probability statistics on measurements to detect source signatures when the background shape is known. The LRT matched filter only removes one source signature at a time. The LRT matched filter has excellent sensitivity in static detection scenarios and may be suitable for small detectors where statistical noise is a major factor. The LRT matched filter may not be suitable for large moving detectors.

[0054] The LRT matched filter constructs two hypotheses as expressed by the following equations: H0=B s B r Δt

number

number

number

[0055] The resulting LRT is based on a two-sided test, whereas other matched filters are based on one hypothesis tests with normal statistics. Therefore, the resulting LRT can be transformed by subtracting the mean with the expected background and dividing by the expected noise of the estimated background. This result represents a one-hypothesis test of the background. The expected background rate is required to generate the metric. The metric can then be expressed by the following equation,

number

[0056] The LRT matched filter follows a behavioral model given by:

number

[0057] Targeted NSCRAD Algorithm NSCRAD Algorithm The Nuisance-Rejection Spectral Comparison Ratio Anomaly Detection ("NSCRAD") algorithm was developed by Pacific Northwest National Laboratory. The NSCRAD algorithm first transforms the measurements into a set of regions of interest and applies a transformation to remove the expected background vectors. Since the regions of interest may overlap, the NSCRAD algorithm takes into account a correlation matrix estimated from background samples.

[0058] Region of interest feature vector JPEG0007660169000032.jpg6150 and the background vector of the region of interest B = [b i ] is given, the NSCRAD transformation α is JPEG0007660169000033.jpg6150, which can be represented by the following equation:

number

[0059] In addition to removing the background components, the NSCRAD algorithm can also employ a subspace projection matrix γ. The NSCRAD algorithm uses the matrix The calculation is performed from TIFF0007660169000035.tif6150. The NSCRAD algorithm calculates the subspace projective transformation as expressed by the following equation: γ(N)=(IN(N t N) -1 N t )

[0060] Since the feature vectors are projected and not independent, the NSCRAD algorithm can use oblique subspace projection. The transformed covariance matrix can be expressed by the following equation: β = (αΣα t ) -1 α where Σ represents covariance. The NSCRAD algorithm adjusts the subspace projection to remove the background as expressed by the following equation: γ(N;α,β)=(I-αN(N t β t αN) -1 N t β t )

[0061] Many different algorithms can be developed from these transformations. The NSCRAD algorithm can be expressed by the following equation: DM=X t (β t γα)X This equation is Q=β t This metric is a classical quadratic form where γα is the number of energy features minus one for background removal, and then minus the number of background sources being removed, χ 2 The implementation of the NSCRAD algorithm can be as described below. [Table 2]

[0062] Because the NSCRAD algorithm has a quadratic form, aggregation is performed on a spectral comparison ratio ("SCR") vector and a projected SCR vector, rather than simply on a numerator and denominator as in other algorithms. This results in increased memory requirements when aggregating measurements spatially. Thus, the NSCRAD algorithm may be appropriate when using temporal aggregation, but may not be appropriate when using spatial aggregation.

[0063] The NSCRAD algorithm is essentially a 2 Therefore, as will be explained later, this algorithm may not work with metric aggregation. The NSCRAD algorithm removes only a single dynamic degree of freedom and a specified number of fixed background dimensions. This approach reduces the complexity of the background estimator, but may reduce systematic variability attenuation. Thus, the maximum aggregation time is limited for large detectors. The NSCRAD algorithm may produce poor statistical distribution when the population is small. This tendency may be enhanced by the χ 2 This requires increasing the threshold for the level expected by the statistics.

[0064] NSCRAD matched filter Unlike the NSCRAD algorithm, the NSCRAD matched filter has statistics with unit variance and normal distribution. This results in a lower tendency to have outliers and a wider operating range with increased systematic variance rejection. The downside of this approach is that the region of interest for optimizing performance is different for each target source. Therefore, every detection metric evaluator needs to be paired with a corresponding metric trainer. The implementation of the NSCRAD matched filter is the same as that of the NSCRAD algorithm, except that the matched filter coefficients are calculated for each region of interest as represented by the following equation: G=S t β t γ(N;α,β)α The decision metric is represented by the following equation:

number

[0065] The NSCRAD matched filter, unlike the NSCRAD algorithm, produces metrics for one source signature. To use a filter to cover a range of sources, a source signature description is required that covers the sources to be detected.

[0066] Metric Aggregation The metric aggregation component can aggregate metrics spatially or temporally. Many spectral detection algorithms perform aggregation within a fixed time window. Spatial aggregation methods generate aggregate metrics from measurements collected at the same location but at different times (e.g., a detector makes multiple sweeps over an area). Temporal aggregation methods generate aggregate metrics for each time interval based on the immediately preceding time interval. The metric aggregation component can employ different aggregation methods to generate aggregate metrics. If the RDA system employs multiple aggregation methods, the metric aggregation component generates aggregate metrics for each aggregation method. The metric evaluation component can use the maximum value of the aggregated metrics to detect whether a source is present.

[0067] Temporal aggregation uses a rolling window of measurements in generating aggregate metrics. The aggregation component generates aggregate metrics for multiple rolling windows, each containing a different number of measurements.

[0068] The spatial aggregation method is a simple 1 / R 2 A weighted back projection method assuming a kernel can be applied to generate an aggregate metric.

[0069] Metric Standardization The metric generation component can employ a metric standardization method to correct the statistical distribution of the detection algorithm, which produces a standard normal distribution on each metric. The metric standardization method maintains a running estimate of the recent mean and variance of each metric. From this running estimate, the metric standardization method calculates the deviation of the mean from zero (bias) and the excess variance. The metric standardization method then corrects the detection metric to remove this bias and reduce the variance.

[0070] The metric standardization method may employ a bias forgetting factor and a variance forgetting factor. The metric standardization method uses these factors in a first order infinite impulse response filter to maintain bias and variance estimates. The smaller this number is, the slower the algorithm learns about changing conditions. For example, the forgetting factor may be about 0.01, which means that the metric standardization method learns about 100 measurements of changes. These forgetting factors are small when the background is quite static and large when the background is changing significantly. The bias and variance forgetting factors may have the same or different values. The metric standardization method may exclude metrics that clearly contain sources when calculating these bias and variance estimates. For example, metrics that are above a threshold may be excluded.

[0071] Measurement Calibration The RDA system calibrates the measurements to account for drift of the detector's sensor during collection of photons. As a result of the drift, photons of one energy level may be identified as a different energy level and the spectrum may be scaled wider or narrower. To account for this drift, the measurement calibration component assigns counts from a measurement bin to a different bin, which is called changing the measurement binning structure or rebinning. The radiation source detector may periodically collect calibration measurements of known sources and / or measurements from the background from the detector to identify characteristics of the drift. The measurement calibration component identifies the required calibration by comparing the calibration measurements to expected measurements of known sources to determine the drift. For example, if the source signature of a known source and the calibration measurements have corresponding peaks in different bins, the measurement counts need to be reassigned based on the bin difference. For example, in a simple example, the measurements may have exactly the same shape as the source signature, except that they are shifted by one bin to a lower energy level due to drift. The calibration component may also correct for nonlinearities in the detector response, including nonlinearities inherent in the photon absorber and nonlinearities from the amplifier. Typically, these nonlinearities are captured in characterization measurements made prior to deployment and provided in the form of polynomial or spline coefficients, or energy / channel pairs.

[0072] In some embodiments, robust measurements of sources with well-separated features can be used for monitoring and possibly further correction of nonlinearities. To do this, each source signature is augmented with a lookup table that contains the effective energy of each feature in the source signature (e.g., gamma emission lines or Compton edges) that is useful for calibration. The effective energy is the energy found for a perfect measured source signature that accounts for shifts from the true energy due to the effects of shielding and scattering on the feature location extraction procedure (e.g., centroiding or line fitting). In addition to the effective energy, the table also lists the expected range of effective energy (e.g., standard deviation) along with the range of the ratio of peak counts to counts across the spectrum. The effective energy, range of effective energies, and peak-to-total count ratio can all be functions of the signal-to-noise ratio of the measurement and the total number of counts in the measurement. These dependencies are included in the lookup table of the source signature. The lookup table also includes the settings of the feature extraction procedure (e.g., region of interest). Some source signatures, such as those with large amounts of shielding or those of nuclides that do not have clearly defined characteristics, are not suitable for calibration monitoring and correction, do not have useful energies, and do not contain associated values.

[0073] Once a source is detected with a sufficiently high decision metric and source identification probability, the effective energy of each feature is calculated and compared to the energies in the lookup table of source signatures. If the extracted energy and the effective energy listed in the lookup table differ significantly from the expected range for the measurement conditions encountered, a flag can be set indicating a calibration problem (calibration monitor) and / or this information can be used to update the energy scale (calibration correction). The measured peak-to-total count ratio is also compared to the appropriate range in the lookup table to ensure that the correct feature and source signature are found regardless of the energy scale. Multiple measurements of the calibration error are required to trigger a correction. In this case, a constrained spline can be used to generate the corrected energy scale, which maintains the overall shape of the nonlinearity from the more finely controlled pre-deployment calibration measurements for a specific detector or general detector type, but also adjusts to more closely match the measurements at the available energies.

[0074] The measurement calibration component reallocates the counts of each measurement bin to the next higher energy level to account for changes in energy scale due to drift and nonlinearity. As another example, drift in the energy range can vary across the energy range, e.g., low energy bin counts can be compressed into fewer bins and high energy bin counts can be expanded into more bins. As another example, when measurement bin counts need to be allocated to multiple bins, correlation can exist in the statistics of the data. Statistical analysis methods may introduce bias due to this correlation that may affect the detection metric. To prevent such statistical bias, the measurement calibration component employs a statistical rebinning method. For example, the measurement calibration component can randomly select a binomial variable for the number of measurement bin counts that need to be reallocated to other bins. By introducing this randomness, the Poisson statistics of the rebinning are preserved. The nonlinearity correction can account for environmental factors such as temperature, or computational factors such as count rate. Thus, the coefficients can be functions of these factors.

[0075] FIG. 5 illustrates calibration of measurements based on detector drift. Measurement 510 includes extended bin 511 with extended edges 512 and 513. Calibrated measurement 520 includes bins 521-523 with edges 524-527 specified by the binning structure. The calibration measurements represent a compression of the magnitude of the measurements. Bin 511 overlaps bins 521-523. The measurement calibration component allocates a fraction of the counts of bin 511 to bin 521-523 to distribute the counts of bin 511 to bins 521-523. This fraction is based on the ratio of energy range 528 to the energy range of bin 511. Similarly, the fraction of the counts corresponding to energy range 529 is allocated to bin 523. The remaining counts are allocated to bin 522. Rather than using the exact fraction of the counts, the measurement calibration component can select the number of counts based on a binomial distribution of the fraction of the counts.

[0076] Dynamic Background Estimation RDA systems provide a dynamic estimation of the background of the measurement. If the detector is moving, the measured background can change significantly (e.g., by a factor of 5). Gross count algorithms, such as the gross counts k-sigma algorithm, are dominated by these variations and are largely insensitive to weak sources. Spectral detection algorithms detect sources by looking for changes in the spectral signature rather than changes in overall intensity. Thus, the spectral detection algorithm is trained using a set of representative backgrounds for the area to be measured. Typically, such representative backgrounds are developed based on collecting hours of background data that is used as a fixed background training set.

[0077] In some cases, generating a representative background may be difficult. For example, if a detector is to be deployed in a new area, collecting background data may not be feasible. Also, the detector response may change over time. These changes may include gain drifts, changes in nonlinear response, and changes in energy resolution and efficiency. Although the measurement calibration component corrects for gain drift, it is limited in correcting other changes in the detector response. As a result, a representative background generated from background data collected on one day may not represent the background detected on another day. The RDA system can use a dynamic background estimate that is trained "on the fly" during normal operation to make the detection sensitivity and false positive rate more reliable, robust, and predictable.

[0078] Because the RDA system receives measurements from detectors that measure sources as well as background, the background estimation component filters out the contribution of sources. These sources can include intentional sources, unknown sources found as part of a source search, and accidental encounters with sources commonly found in (e.g., medical) environments. To improve the estimated background, the background estimation component recognizes when measurements represent accidental sources and filters out these measurements when generating the estimated background. The RDA system employs a rejection component that analyzes metrics to determine which measurements should be filtered out. It is possible not to filter out source measurements. For example, a source may not be detected and therefore not filtered out if it is present when the detector begins collecting measurements, or if radiation from the source increases gradually. This can result in a small amount of the source signature being introduced into the background, making the background estimation less sensitive to this source. In such cases, measurements of the source based on slightly larger intensities will not be recognized as a source and will not be further filtered out.

[0079] To prevent such sources from going undetected and therefore unfiltered, the RDA system employs a sensitivity tester. The sensitivity tester periodically checks the sensitivity to many sources by injecting sources of various intensities into an estimated background to generate test measurements. The sensitivity tester then runs a detection algorithm on the test measurements using the estimated background generated by the background estimation component to determine the minimum intensity of the source required for detection. The sensitivity tester compares this minimum intensity to a previous minimum intensity to determine whether sensitivity has faded. If so, it restarts the background estimation component to generate a new estimated background.

[0080] Rank Average The rank averaging algorithm generates a background rate for all measurements received. The RDA system can use the rank averaging algorithm in conjunction with the gross count detection algorithm during operating conditions where the background of measurements temporarily increases or decreases. For example, the background rate may decrease when an object (e.g., a truck) passes the detector, and the background rate may increase when the detector passes a structure containing a source (e.g., a building).

[0081] The rank average algorithm sorts the total counts of each measurement collected over a period of time to account for temporary dips or rises in background. The rank average algorithm then generates an average of the top (or bottom) counts. This average is a biased estimate of the background. The rank average algorithm calculates a bias factor using standard statistical assumptions. The rank average algorithm then corrects the biased estimate based on the bias factor. For example, the rank average algorithm can calculate the expected offset of a Gaussian distribution in the average total counts and subtract this offset from the estimated background.

[0082] The rank-averaging algorithm can be particularly useful when the detector is stationary, when the background is obscured by nearby objects, and when an estimated background is required in the absence of these objects.

[0083] Progressive Projection The RDA system employs a progressive projection algorithm that estimates a background vector set that spans a linear combination of all possible background shapes. This spanning set is called a basis, but the background vectors are not orthogonal. The progressive projection algorithm queues background measurements (i.e., backgrounds where no source is detected) until a change occurs in the spectral shape of the background measurement or until a certain minimum time has elapsed. The progressive projection algorithm then compares the background measurement with the changed spectral shape to the current basis vectors to determine which basis vector the background measurement best represents. If the background measurement is found not to be strongly related or separate from any of the existing basis vectors, then this background measurement does not represent anything new about the background. Conversely, if the background measurement is found to be strongly related or separate from the basis vectors, then the progressive projection algorithm averages this background measurement into the existing basis vectors. The progressive projection algorithm then attempts to maximize the eigenvalues ​​associated with the correlation matrix formed by the basis vectors such that the basis vectors have maximum diversity. The progressive projection algorithm employs a forgetting function so that if the background measurements remain the same over a long period of time, the progressive projection algorithm does not update the basis vectors to be non-collinear.

[0084] The progressive projection algorithm can initialize basis vectors from background measurements collected after the algorithm is started or cleared. If the detector is exposed to a source during initialization, the source can be trained to the basis vectors and therefore become unresponsive until the sensitivity tester is activated. The progressive projection can have three start methods. The hot start method uses a predefined set of basis vectors specified in the configuration data. The warm start method generates an average of the background measurements over a specified period and mixes this average with a predefined background measurement from the configurator (or previous run). The cold start method injects a known background measurement after taking the average over a specified period.

[0085] Each start method has different advantages and disadvantages. The cold start method is robust in the sense that it will work even if the source is co-located with the detector, unless the source is removed. However, this robustness comes at the cost of a long period of time that it may take for the method to be fully trained and reach full sensitivity. If the known background measurements of the cold start method do not accurately reflect the variability of the background measurements, the RDA system may not detect the source correctly and must be restarted. The warm start method uses the previous background measurements and is therefore trained quickly. However, if the background measurements are significantly different from the previous background measurements (e.g., a source is present in the background measurements), the progressive projection algorithm will no longer be sensitive to the source and will need to be restarted. The hot start method can generate basis vectors immediately. However, if the background measurements are not represented by the predefined basis vectors, the RDA system will generate a false positive, i.e., the background measurements are indicated as having a source present. As a result, the progressive projection algorithm will not update the basis vectors, the false positives will continue, and it will need to be restarted using either the cold start method or the warm start method.

[0086] classification In some embodiments, the metric generation component employs a discrimination algorithm similar to the OSP matched filter detection algorithm but with a different regularization term. The metric generation component uses the expected background calculated before the measurement to calculate the expected variance in each channel, Σ B When considering a mixture of source signatures of sources, the metric generation component calculates Σ B X=Σ B The metric generation component creates a mixed source signature S = Ax by calculating the mixture of the source signatures of each source that optimally solves Ax. If mixing is not considered, the metric generation component i Use.

[0087] The metric generation component calculates, for each source signature (mixed or unmixed), a transformation as represented by the following equation:

number

[0088] The null hypothesis generation component uses the expected background to estimate the variance of the measurements, Σ Y The null hypothesis generation component generates the variance of each metric and the covariance of the metrics using the following equations: σ i 2 =Q i Σ X Q i t c ij =Q i Σ X Q j t Here, σi 2 is S i represents the variance of ij is S i and S j represents the covariance of the null hypotheses. The null hypothesis generation component selects the maximum of the null hypothesis variances. The null hypothesis generation component then calculates a metric for the null hypothesis based on the maximum of a random walk around zero with a fixed number of draws as represented by the following equation:

number

[0089] Computer Systems A computer system (e.g., a network node or collection of network nodes) capable of implementing the RDA system and other described systems may include a central processing unit, input devices, output devices (e.g., display devices and speakers), storage devices (e.g., memory and disk drives), network interfaces, graphics processing units, cellular or other wireless link interfaces, global positioning system devices and inertial navigation, etc. Input devices may include keyboards, pointing devices, touch screens, gesture recognition devices (e.g., for air gestures), head and eye tracking devices, and microphones for voice recognition, etc. Computer systems may include high performance computer systems, cloud-based servers, desktop computers, laptops, tablets, electronic readers, personal digital assistants, smartphones, gaming consoles, servers, etc. For example, simulation and training may be performed using high performance computer systems, and classification may be performed by mobile devices that are part of the network nodes. Computer systems may access computer readable media, including computer readable storage media and data transmission media. Computer readable storage media are tangible storage means that do not include transitory propagating signals. Examples of computer-readable storage media include primary memory, cache memory, secondary memory (e.g., DVD), and other storage devices. The computer-readable storage media can store or encode computer-executable instructions or logic that implement the RDA system and other described systems. Data transmission media are used to transmit data via ephemeral propagating signals or carrier waves (e.g., electromagnetic) over wired or wireless connections. The computer system can include a secure cryptographic processor as part of the central processing unit that generates and securely stores keys and uses these keys to encrypt and decrypt data.

[0090] The RDA system and other described systems may be described in the general context of computer-executable instructions, such as program modules and components, executed by one or more computers, processors, or other devices. Generally, program modules or components include routines, programs, objects, data structures, etc. that perform the tasks or implement data types of the RDA system and other described systems. Typically, the functionality of the program modules may be combined or distributed as desired in various examples. Aspects of the RDA system and other described systems may be implemented in hardware using, for example, application specific integrated circuits ("ASICs") or field programmable gate arrays ("FPGAs").

[0091] Flow Diagram 6 is a flow diagram illustrating the high-level processing of the OSP matched filter component of the RDA system in some embodiments. t The component 600 passes the source definition S and an indication of the source. In block 601, the component computes a transformation projection T that complements the background B. In block 602, the component computes the background-subtracted X t In block 603, the component calculates the partial projection of the partial noise n t In block 604, the component updates the moving window projection to add the partial projection over time t and subtract the partial projection over time t-Δ. In block 605, the component updates the moving window noise to add the partial noise over time t and subtract the partial noise over time t-Δ. In block 606, the component computes the detection metric as the projection divided by the noise. The component then terminates.

[0092] FIG. 7 is a flow diagram illustrating the process of the calibration component of the RDA system in some embodiments. Invokes component 700 to correct existing bins based on radiation detector sensor drift in signal energy detection. Invokes component to forward indication of edge and measurement of current bin and scale s. In blocks 701-205, the component initializes the number of overflow bins (#OFbins) and the number of underflow bins (#UFbins). In block 701, the component sets the number of overflow bins to 0. In decision block 702, if scale is less than 1.0, the component proceeds to block 703, else the component proceeds to block 704. In block 703, the component sets the number of underflow bins to the number of bins minus the integer part of the number of bins (#bin) multiplied by the scale. In block 704, the component sets the number of underflow bins to 0. In decision block 705, if scale is equal to 1.0, the component ends since no rebinning is required, else the component proceeds to block 706. In block 706, the component initializes variables b0 and b1 to track the energy of the bin edges and the bin index i. In blocks 707-713, the component loops performing the rebinning. In block 707, the component increments index i to the next bin. In decision block 708, if all bins have already been indexed, the component ends, else the component proceeds to block 709. In block 709, the component sets b0 to the current edge and b1 to the next edge. In block 710, the component sets indexes i0 and i1 to the integer parts of index b0 and index b11, respectively. In decision block 711, if index i0 is equal to index i1, the component proceeds to block 713, else the component proceeds to block 712.In block 712, the component calls the Split Bin component to split the bins and then loops to block 707 to select the next bin, else the component continues in block 713. In block 713, the component increments the new bin count out[i0] by the measurement count X[i] and then loops to block 707 to select the next bin.

[0093] 8 is a flow diagram that illustrates the processing of the split bins component of the RDA system in some embodiments. The component invokes the split bins component 800. In block 801, the component sets a variable n to the counts of bin[i] and a variable split to a fraction of these counts to assign to out[i0]. In block 802, the component applies a binomial distribution to select the number of counts x1 to add to out[i0] based on the variable n and split. In block 803, the component adds the count x1 to out[i0] and decrements the remaining count of variable n by count x1. In block 804, the component sets index j to index i0 plus 1. In decision block 805, if the variable j is greater than the minimum value of variable x1 and the number of bins #bin, the component ends; otherwise, the component increments variable j and proceeds to block 806. In block 806, the component sets variable x1 to a value based on the binomial distribution of variable n and 1.0 divided by scale s. In block 807, the component decrements the variable n by the variable x1. In block 808, the component increments the count of out[i1] by the variable x1. In decision block 809, if the index i1 is less than the number of bins, the component continues at block 801, else the component continues at block 811. In block 810, the component increments out[i1] by the variable n and loops to block 805. In block 811, the component sets the count of overflows to the variable n and loops to block 805.

[0094] FIG. 9 is a flow diagram illustrating the processing of the estimate background based rank average component of the RDA system in some embodiments. The component calls the adjust background estimate component 900. In block 901, the component initializes a variable i to index across bins. In decision block 901, the component increments the variable i and if it is greater than the number of bins, the component ends, otherwise the component proceeds to block 903. In block 903, the component creates a sort of counts within a moving window of the indexed bin. In block 904, the component calculates the average of the top counts. In block 905, the component calculates the expected offset of the top count based on the average. In block 906, the component adjusts the count of the indexed bin by the offset and loops to block 902 to select the next bin.

[0095] The following paragraphs describe various embodiments of the RDA system and other system aspects. A system implementation may employ any combination of the embodiments. The processes described below may be performed by one or more computer systems having a processor executing computer-executable instructions stored on a computer-readable storage medium that implements the system.

[0096] In some embodiments, one or more computer systems for identifying a radiation source are provided. The one or more computer systems include one or more computer readable storage media that store computer executable instructions and one or more processors that execute the computer executable instructions stored on the one or more computer readable storage media. The instructions implement a radiation source detector that receives measurements of radiation, generates a detection metric for one or more sources that is indicative of whether the source is present in the measurements, and evaluates the detection metric to detect whether the source is present in the measurements. The instructions also implement a radiation source identifier that, if the presence of a source is detected in the measurements, generates an identification metric for the one or more sources that is indicative of whether the source is present in the measurements and generates a null hypothesis metric that is indicative of whether the source is absent in the measurements, and evaluates the one or more identification metrics and the one or more null hypothesis metrics to identify the source, if present, in the measurements. In some embodiments, the radiation source detector generates a detection metric for a source based on a detection source signature of the source and generates an identification metric for the source based on a plurality of identification source signatures of the source. In some embodiments, the identification source signatures of the source represent sources with different shielding. In some embodiments, the radiation source identifier identifies a radiation source class for the identified source. In some embodiments, the measurements are calibrated to account for drift in the detector that collected the measurements. In some embodiments, the radiation source identifier generates an aggregate metric for each source. In some embodiments, measurements within a window are aggregated and radiation is detected and identified based on the aggregate measurements.

[0097] In some embodiments, a method for detecting the presence of a radiation source from measurements of radiation performed by a computer system. The method uses, for one or more sources, a detection algorithm that considers an estimate of background radiation generated based on previous measurements of a current measurement, for one or more detection algorithms, to generate a metric indicative of a similarity between a current measurement and a source signature of the source, and generates an aggregate metric for the source from the one or more metrics of the source. The method analyzes the aggregate metric for the one or more sources to determine whether the source is present in the current measurement. The method indicates that the presence of the source has been detected upon determining that the source is likely present in the current measurement. In some embodiments, the method further generates an estimate of background radiation based on previous measurements, excluding previous measurements in which the presence of the source was detected. In some embodiments, the estimate of background radiation is a weighted average based on the measurements and the prior estimate of background radiation. In some embodiments, the source signature is associated with source identification and source shielding. In some embodiments, the method generates a metric for each source signature associated with the source for each detection algorithm. In some embodiments, the measurements are represented by a histogram of energy ranges including the number of photons in each energy range. In some embodiments, a metric is generated for each of a plurality of window measurements, each window including the current measurement and a number of adjacent previous measurements. In some embodiments, the detection algorithm is associated with coefficients, and the method dynamically adjusts the coefficients based on an estimate of background radiation. In some embodiments, the analysis of the aggregate metric includes applying a classifier trained using training data including a feature vector having features including the detection algorithm, the window size, and the aggregate metric, and a label for each feature vector indicating whether a source is present. In some embodiments, the feature vector is generated from collected radiation measurements. In some embodiments, the feature vector is generated from simulated radiation measurements.In some embodiments, the method further calibrates the current measurement to account for drift associated with the detector used to collect the measurement.

[0098] In some embodiments, a method is provided for identifying a radiation source from radiation measurements, performed by a computer system. The method receives a current measurement in which the presence of a source is detected based on a detected source signature. The method generates a metric for each of a plurality of discriminatory source signatures using a discrimination algorithm. The metric is indicative of a similarity between the current measurement and the source signature for that source. The discrimination algorithm takes into account an estimate of background radiation generated in detecting the presence of the source. The discriminatory source signature is more comprehensive than the detected source signature. The method generates a source metric for each source based on one or more metrics generated using the source signature for that source. The method generates a null hypothesis metric indicative of a similarity between the current measurement and the estimate of background radiation. The source and background radiation estimates are targets. The source metric and the null hypothesis metric are target metrics. The method generates a target probability for each target indicative of the presence of that target in the current measurement. The target probability is based on the target metric for that target. In some embodiments, the method generates, for each of a plurality of radiation source classes, a radiation source class probability that the current measurement represents a source of that radiation source class. The radiation source class probability is generated based on the target probabilities. In some embodiments, the generation of the target probabilities considers the prior probability of each target. In some embodiments, the identification algorithm is based on an orthonormal subspace projection matched filter algorithm.

[0099] A matched filter algorithm for orthonormal subspace projection. In some embodiments, a method is provided for generating a metric associated with a radiation source in a measurement of radiation, the method being executed by one or more computer systems. The method accesses a source signature of the source, an estimated background, and a background basis vector. The method generates a projection vector based on the source signature, the estimated background, and the background basis vector. The method accesses the measurement. The method generates a metric based on the source signature, the estimated background, and the background basis vector. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, generating the projection vector includes generating a weighting matrix from the estimated background, the variance of each energy bin being based on an expected background. In some embodiments, generating the projection vector further includes generating a source signature weighted projection and a background weighted projection, and setting the projection vector to the difference between the source signature weighted projection and the background weighted projection. In some embodiments, the source signature represents an obscuration of the source. In some embodiments, the method further generates aggregate measurements of different numbers of measurements and, for each aggregate measurement, generates a metric based on the aggregate measurement, the projection vector and the expected variance of the aggregate measurement. In some embodiments, the method indicates that the presence of a source is detected when the metric meets a source presence threshold. In some embodiments, the method indicates that the presence of a source is detected semi-definitively when the metric meets a source presence threshold. In some embodiments, generating the detection metric includes dividing the product of the projection vector and the measurement by the square root of the expected variance.

[0100] In some embodiments, one or more computer systems are provided for generating a metric associated with a radiation source in a radiation measurement. The one or more computer systems include one or more computer readable storage media storing computer executable instructions and one or more processors executing the computer executable instructions stored in the one or more computer readable storage media. The instructions, when executed, access a source signature of the source, an estimated background, and a background basis vector. The instructions generate a projection vector based on the source signature, the estimated background, and the background basis vector. The instructions access the measurement. The instructions generate a metric based on the measurement, the projection vector, and an expected variance of the measurement. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the instructions for generating the projection vector generate a weight matrix from the estimated background, and the variance of each energy bin is based on the expected background. In some embodiments, the instructions to generate the projection vector further generate a source signature weighted projection and a background weighted projection, and set the projection vector to the difference between the source signature weighted projection and the background weighted projection. In some embodiments, the source signature represents an occlusion of the source. In some embodiments, the instructions further generate aggregate measurements of different numbers of measurements, and for each aggregate measurement generate a metric based on the aggregate measurement, the projection vector, and an expected variance of the aggregate measurement. In some embodiments, the instructions further indicate that a presence of a source has been detected when the metric meets a source presence threshold. In some embodiments, the instructions further indicate that a presence of a source has been semi-definitively detected when the metric meets a source presence threshold. In some embodiments, the instructions to generate the detection metric divide the product of the projection vector and the measurement by the square root of the expected variance.

[0101] In some embodiments, a method is provided for generating a metric associated with a radiation source in a sequence of collected radiation measurements, the method being executed by one or more computer systems. The method accesses a source signature of the source, an estimated background, and background basis vectors. The method generates a projection vector as represented by the following equation: T=S t W(IB(B t WB) -1 B t W) where T represents the projection vector, S represents the source signature, B represents the estimated background, and W represents the background basis vector. The method accesses the measurements. The method generates a metric as represented by the following equation:

number

[0102] Hybrid Matched Filter In some embodiments, a method is provided for generating a metric associated with a radiation source in a measurement of radiation, the method being executed by one or more computer systems. The method accesses a source signature and an estimated background for the source. The method generates a background vector that removes the source signature. The method generates a projection vector based on a weighting matrix derived from the estimated background, the source signature and the estimated background. The method accesses the measurement. The method generates a metric based on the measurement, the projection vector and an expected variance an average background. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on the expected background. In some embodiments, the source signature represents an obstruction of the source. In some embodiments, the method generates aggregate measurements for a different number of measurements, and for each aggregate measurement generates a metric based on the aggregate measurement and the projection vector. In some embodiments, the method indicates that the presence of a source is detected when the metric meets a source presence threshold.In some embodiments, the method indicates that the presence of a source is detected semi-definitively when the metric meets a source presence threshold.

[0103] In some embodiments, one or more computer systems are provided for generating a metric associated with a radiation source in a radiation measurement. The one or more computer systems include one or more computer readable storage media storing computer executable instructions and one or more processors executing the computer executable instructions stored in the one or more computer readable storage media. When executed, the instructions access a source signature and an estimated background of the source. The instructions generate a background vector that removes the source signature. The instructions generate a projection vector based on a weighting matrix derived from the estimated background, the source signature and the estimated background. The instructions access the measurement. The instructions generate a metric based on the measurement, the projection vector and the expected variance average background. In some embodiments, the source signature hs is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on the expected background. In some embodiments, the source signature represents an obscuration of the source. In some embodiments, the instructions further generate aggregate measurements of the different numbers of measurements and, for each aggregate measurement, generate a metric based on the aggregate measurement and the projection vector. In some embodiments, the instructions further indicate that a presence of the source has been detected when the metric satisfies a source presence threshold. In some embodiments, the instructions further indicate that a presence of the source has been semi-definitively detected when the metric satisfies a source presence threshold.

[0104] In some embodiments, a method is provided for generating a metric indicative of the presence of a radiation source in sequentially collected radiation measurements, the method being executed by one or more computer systems. The method accesses a source signature of the source, an estimated background, and background basis vectors. The method generates a projection vector as represented by the following equation:

number

number

number

number

[0105] Two-Dimensional ("BD") Matched Filters In some embodiments, a method is provided for generating a metric associated with a radiation source in a radiation measurement, the method being executed by one or more computer systems. The method accesses the measurement, a source signature of the source, and an estimated background. The method generates a background vector that removes the source signature. The method generates a projection vector based on a weighting matrix derived from the estimated background, the source signature, and the estimated background. The method accesses the measurement. The method generates a metric based on the measurement, the projection vector, and an expected variance average background. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on the expected background. In some embodiments, the source signature represents an obstruction of the source. In some embodiments, the method further generates aggregate measurements of different numbers of measurements, and for each aggregate measurement generates a metric based on the aggregate measurement and the projection vector. In some embodiments, the method indicates that a source presence is detected when the metric meets a source presence threshold. In some embodiments, the method indicates that the presence of a source has been semi-definitively detected when the metric meets a source presence threshold.

[0106] In some embodiments, one or more computer systems are provided for generating a metric associated with a radiation source in a radiation measurement. The one or more computer systems include one or more computer readable storage media storing computer executable instructions for controlling the one or more computer systems, and one or more processors executing the computer executable instructions stored in the one or more computer readable storage media. The instructions access the measurement, a source signature of the source, and an estimated background. The instructions generate a background vector that removes the source signature. The instructions generate a projection vector based on a weighting matrix derived from the estimated background, the source signature, and the estimated background. The instructions access the measurement. The instructions generate a metric based on the measurement, the projection vector, and the expected variance average background. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on the expected background. In some embodiments, the source signature represents an obstruction of the source. In some embodiments, the instructions further generate aggregate measurements of different numbers of measurements and, for each aggregate measurement, generate a metric based on the aggregate measurement and the projection vector. In some embodiments, the instructions further indicate that a presence of the source has been detected when the metric satisfies a source presence threshold. In some embodiments, the instructions indicate that a presence of the source has been semi-definitively detected when the metric satisfies a source presence threshold.

[0107] In some embodiments, a method is provided for generating a metric indicative of the presence of a radiation source in sequentially collected radiation measurements, the method being executed by one or more computer systems. The method accesses a source signature of the source, an estimated background, and background basis vectors. The method solves the following equation:

number

number

number

[0108] LRT Matched Filter In some embodiments, a method is provided for generating a metric associated with a radiation source in a measurement of radiation, the method being executed by one or more computer systems. The method accesses a source signature and an estimated background of the source. The method generates a first hypothesis based on the estimated background and the estimated background rate. The method generates a second hypothesis based on the estimated background, the estimated background rate, and the source signature. The method generates a first likelihood of the first hypothesis given a Poisson distribution. The method generates a second likelihood of the second hypothesis given a Poisson distribution. The method generates a projection vector based on the first likelihood and the second likelihood. The method accesses the measurement. The method generates a metric based on the measurement, the projection vector, and the estimated background rate. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins. The energy bins have values ​​representing the number of photons emitted by the source over a time interval. In some embodiments, the source signature represents shielding of the source. In some embodiments, the method further generates aggregate measurements of different numbers of measurements, and for each aggregate measurement generates a metric based on the aggregate measurement and the projection vector. In some embodiments, the method indicates that a presence of a source has been detected when the metric satisfies a source presence threshold. The method indicates that a presence of a source has been detected semi-definitively when the metric satisfies a source presence threshold.

[0109] One or more computer systems are provided for generating a metric associated with a radiation source in a measurement of radiation. The one or more computer systems include one or more computer readable storage media storing computer executable instructions for controlling the one or more computer systems and one or more processors executing the computer executable instructions stored on the one or more computer readable storage media. The instructions access a source signature and an estimated background of the source. The instructions generate a first hypothesis based on the estimated background and the estimated background rate. The instructions generate a second hypothesis based on the estimated background, the estimated background rate and the source signature. The instructions generate a first likelihood of the first hypothesis given a Poisson distribution. The instructions generate a second likelihood of the second hypothesis given a Poisson distribution. The instructions generate a projection vector based on the first likelihood and the second likelihood. The instructions access the measurement. The instructions generate a metric based on the measurement, the projection vector and the estimated background rate. In some embodiments, the source signature is a histogram representing an energy range divided into energy bins, with each energy bin having a value representing the number of photons emitted by the source over a time interval. In some embodiments, the source signature represents occlusion of the source. In some embodiments, the instructions generate aggregate measurements of different numbers of measurements and, for each aggregate measurement, generate a metric based on the aggregate measurement and the projection vector. In some embodiments, the instructions indicate that a source presence has been detected when the metric meets a source presence threshold. In some embodiments, the instructions indicate that a source presence has been detected semi-definitively when the metric meets a source presence threshold.

[0110] In some embodiments, a method is provided for generating a metric indicative of the presence of a radiation source in sequentially collected radiation measurements, the method being executed by one or more computer systems, comprising: constructing a hypothesis represented by the following equation: H0=B s B r Δt

number

number

number

number

[0111] Rank Average In some embodiments, a method is provided for adjusting measurement counts of measurements to account for temporal variations in the measurements, the method being performed by one or more computer systems. The method accesses measurements. For each measurement, the method generates a measurement count for that measurement. The method generates an upper average measurement count that is an average of the measurement counts higher than a threshold count. The measurement calculates an expected offset of a probability distribution based on the upper average measurement count and the measurement counts above the threshold. For each measurement, the method subtracts an estimated offset from the measurement count for that measurement. In some embodiments, the measurements are measurements of radiation and the temporal variations are a result of changes in background radiation. In some embodiments, the method further detects when the measurements include a radiation source. In some embodiments, the detection is performed using a gross count detection algorithm. In some embodiments, the upper average measurement count is based on a probability distribution.

[0112] In some embodiments, one or more computer systems adjust measurement counts of measurements to account for temporal changes in the measurements. The one or more computer systems include one or more computer readable storage media storing computer executable instructions for controlling the one or more computer systems and one or more processors executing the computer executable instructions stored on the one or more computer readable storage media. The instructions generate, for each measurement of a plurality of measurements, a measurement count for that measurement. The instructions generate a top average measurement count that is an average of the measurement counts above a threshold count. The instructions calculate an expected offset of a probability distribution based on the top average measurement counts and the measurement counts above the threshold. For each measurement, the instructions subtract an estimated offset from the measurement count for that measurement. In some embodiments, the measurements are measurements of radiation and the temporal changes are a result of changes in background radiation. In some embodiments, the instructions further detect when the measurements include a radiation source. In some embodiments, the detection is performed using a gross count detection algorithm. In some embodiments, the top average measurement count is based on a probability distribution.

[0113] progressive projection In some embodiments, a method is provided for estimating a background histogram of background count measurement histograms of measurement counts of collected measurements to detect the presence of a source histogram in the measurements, the method being executed by one or more computer systems. The method accesses a spanning vector of the histogram spanning a linear combination of previous background histograms derived from previous measurement histograms. The method calculates a current background histogram of a current sequence of measurement histograms. The method determines whether the current background histogram has a strong association or strong separation with the spanning vector. If the method determines that the current background histogram has a strong association or strong separation with the spanning vector, the method modifies the spanning vector based on the current background histogram. After modifying the spanning vector, the method adjusts the spanning vector to increase diversity. In some embodiments, the adjustment attempts to maximize eigenvalues ​​association with a correlation matrix formed by the spanning vector. In some embodiments, the method applies a detection algorithm to detect the presence of a source in the target measurements that takes into account the adjusted spanning vector. In some embodiments, the source is a radiation source. In some embodiments, the previous measurement histogram is a predefined histogram. In some embodiments, the previous measurement histogram includes a histogram collected during the same collection process but prior to collection of the sequence of the current measurement histogram. In some embodiments, a previous measurement histogram is not collected during the same collection process.

[0114] In some embodiments, one or more computer systems are provided that estimate a background histogram of background count measurement histograms of measurement counts of collected measurements to detect the presence of a source histogram of a source in the measurements. The one or more computer systems include one or more computer readable storage media storing computer executable instructions for controlling the one or more computer systems and one or more processors executing the computer executable instructions stored in the one or more computer readable storage media. The instructions access a spanning vector of a histogram that spans a linear combination of previous background histograms derived from previous measurement histograms. The instructions calculate a current background histogram of a current measurement histogram sequence. The instructions determine whether the current background histogram has a strong association or strong separability with the spanning vector. If the instructions determine that the current background histogram has a strong association or strong separability with the spanning vector, then the instructions modify the spanning vector based on the current background histogram. After modifying the spanning vector, the instructions adjust the spanning vector to increase diversity. In some embodiments, the instructions attempt to maximize eigenvalue associations with a correlation matrix formed by the spanning vector. In some embodiments, the instructions further apply a detection algorithm that detects the presence of a source in the target measurements that take into account the adjusted spanning vector. In some embodiments, the source is a radiation source. In some embodiments, the previous measurement histogram is a predefined histogram. In some embodiments, the previous measurement histogram comprises a histogram collected during the same collection process and prior to collection of the sequence of the current measurement histogram. In some embodiments, previous measurement histograms are not collected during the same collection process.

[0115] Although the subject matter has been described in language specific to structural features and / or acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims. Accordingly, the invention is not limited except as by the appended claims. [Explanation of symbols]

[0116] 900 Estimated background RA 901 i=0 902 ++ i>N 903 x[]=sort(x[i,t]) 904 Calculate the average of the top x counts 905 Off=Expected offset 906 Adjust x[i,t] by offset

Claims

1. 1. A method for generating a metric relating to a radiation source in a measurement of radiation, comprising: accessing measurements, a source signature of the source, and an estimated background; generating a background vector that removes the source signature; generating a projection vector based on a weighting matrix derived from the estimated background, the source signature and the estimated background; generating the metric based on the measurements, the projection vector, and an expected variance calculated from an average background; The method according to claim 1, further comprising:

2. The source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. The method of claim 1.

3. the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on an expected variance calculated from the average background; The method of claim 2.

4. the source signature represents the shielding of the source; The method of claim 1.

5. generating aggregate measurements of different numbers of measurements; and for each aggregate measurement, generating a metric based on the aggregate measurement and the projection vector. The method of claim 1.

6. generating an output indicating that the presence of the source has been detected when the metric satisfies a source presence threshold. The method of claim 1.

7. generating an output indicating that the presence of the source has been semi-conclusively detected when the metric satisfies a source presence threshold. The method of claim 1.

8. The projection vector is expressed by the equation [0010] is generated according to where T represents the projection vector, S represents the source signature, and W represents the basis vectors. 【number】 is the equation [0025] is represented by k and μ represent variables that are fitted to generate a metric with unit variance, and μ is expressed in the equation: [0030] where 【number】 represents the average background, and the metric is calculated using the equation [0045] is generated according to where DM represents the metric, X represents the measurement, t represents the time of measurement, and Δt represents the change in time period. 【number】 represents the expected background at time t, The method of claim 1.

9. The method comprises the steps of: [0050] [0033] where S represents the source signature, 【number】 represents a row vector with components equal to the number of background components, s represents the estimated count of the source, 【number】 represents a vector of background intensities for each component, B is a matrix of background basis vectors, X is the measurement, and e b is the estimated total count in the background, λ is a tuning parameter based on the reliability of the estimated background total count, and the metric is calculated using the equation [006] is generated according to where DM represents the metric and κ represents the penalty for any difference between the estimated total counts in the background and the expected total background counts in the background. The method of claim 1.

10. 1. One or more computer systems for generating a metric relating to a radiation source in a measurement of radiation, comprising: one or more computer readable storage media storing computer executable instructions; one or more processors capable of executing computer-executable instructions stored on said one or more computer-readable storage media; Equipped with The computer executable instructions are for executing the one or more computer systems: accessing a source signature and an estimated background for said source; generating a background vector that removes the source signature; generating a projection vector based on a weighting matrix derived from the estimated background, the source signature and the estimated background; Access the measurements, generating the metric based on an expected variance calculated from the measurements, the projection vector, and an average background; To execute One or more computer systems.

11. The source signature is a histogram representing an energy range divided into energy bins, each energy bin having a value representing the number of photons emitted by the source over a time interval. One or more computer systems according to claim 10.

12. the weighting matrix is ​​generated from the estimated background, and the variance of each energy bin is based on an expected variance calculated from the average background; 12. One or more computer systems according to claim 11.

13. the source signature represents the shielding of the source; One or more computer systems according to claim 10.

14. Execution of the instructions further includes generating aggregate measurements of the different numbers of measurements, and for each aggregate measurement, generating a metric based on the aggregate measurement and the projection vector. One or more computer systems according to claim 10.

15. When the metric satisfies a source presence threshold, execution of the instructions further results in generating an output indicating that the presence of the source has been detected. One or more computer systems according to claim 10.

16. When the metric satisfies a source presence threshold, execution of the instructions further results in generating an output indicating that the presence of the source has been semi-conclusively detected. One or more computer systems according to claim 10.

17. The projection vector is expressed by the equation [0070] is generated according to where T represents the projection vector, S represents the source signature, and W represents the basis vectors. 【number】 is the equation [0080] is represented by k and μ represent variables that are fitted to generate a metric with unit variance, and μ is expressed in the equation: [0090] where 【number】 represents the average background, and the metric is calculated using the equation [0010] is generated according to where DM represents the metric, X represents the measurement, t represents the time of measurement, and Δt represents the change in time period. 【number】 represents the expected background at time t, One or more computer systems according to claim 10.

18. Execution of the computer-executable instructions further comprises: ##EQU00011## Let them solve the where S represents the source signature, 【number】 represents a row vector with components equal to the number of background components, s represents the estimated count of the source, 【number】 represents a vector of background intensities for each component, B is a matrix of background basis vectors, X is the measurement, and e b is the estimated total count in the background, λ is a tuning parameter based on the reliability of the estimated background total count, and the metric is calculated using the equation ##EQU00012## is generated according to where DM represents the metric and κ represents the penalty for any difference between the estimated total counts in the background and the expected total background counts in the background. One or more computer systems according to claim 10.

Images