U.S. patent application number 11/541935 was filed with the patent office on 2009-04-23 for agent detection in the presence of background clutter.
This patent application is currently assigned to SPARTA, INC.. Invention is credited to Philip D. Henshaw, Pierre C. Trepagnier.
Application Number | 20090101843 11/541935 |
Document ID | / |
Family ID | 40562527 |
Filed Date | 2009-04-23 |
United States Patent
Application |
20090101843 |
Kind Code |
A1 |
Henshaw; Philip D. ; et
al. |
April 23, 2009 |
AGENT DETECTION IN THE PRESENCE OF BACKGROUND CLUTTER
Abstract
The present invention generally provides systems and methods for
detection of agents of interest in a bulk quantity of matter, which
also contains clutter and other constituents that typically
interfere with the detection of one or more agents of interest. A
detection system of the invention generally contains a collection
subsystem for obtaining a bulk sample, an interrogation subsystem
for generating one or more analytical signals representative of the
composition of the bulk sample, and an analytical subsystem
according to the teachings of the invention that implements the
methods and algorithms of the invention for analyzing the sample
analytical signals to determine whether one or more agents of
interest are present, e.g., at quantities above a certain
threshold, in the bulk sample.
Inventors: |
Henshaw; Philip D.;
(Carlisle, MA) ; Trepagnier; Pierre C.; (Medford,
MA) |
Correspondence
Address: |
NUTTER MCCLENNEN & FISH LLP
WORLD TRADE CENTER WEST, 155 SEAPORT BOULEVARD
BOSTON
MA
02210-2604
US
|
Assignee: |
SPARTA, INC.
|
Family ID: |
40562527 |
Appl. No.: |
11/541935 |
Filed: |
October 2, 2006 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60723985 |
Oct 3, 2005 |
|
|
|
Current U.S.
Class: |
250/484.4 ;
356/300 |
Current CPC
Class: |
G01N 2021/6421 20130101;
G01N 2021/6417 20130101; G01N 21/6408 20130101; G01N 2201/129
20130101; G01J 3/28 20130101; G01J 3/44 20130101 |
Class at
Publication: |
250/484.4 ;
356/300 |
International
Class: |
H05B 33/00 20060101
H05B033/00; G01J 3/00 20060101 G01J003/00 |
Goverment Interests
[0002] This invention was made with U.S. Government support under
contract number DAAD13-03-C-0077 awarded by the Department of
Defense. The government has certain rights in the invention.
Claims
1. A method of detecting an agent in a bulk sample including at
least one background constituent, comprising: utilizing a
measurement modality to interrogate the sample with electromagnetic
radiation so as to generate spectral data corresponding thereto,
deriving principal components of the sample spectral data, applying
a rotate-and-suppress transformation to said principal components
of the sample data, wherein said transformation suppresses a
contribution of at least one background constituent, if present,
and comparing said transformed principal components of the sample
data with background-suppressed principal components of
corresponding spectral data of an agent to determine whether said
agent is present in the bulk sample.
2. The method of claim 1, further comprising utilizing said
measurement modality to obtain spectral data corresponding to the
agent, deriving principal components of the agent spectral data,
and applying said rotate-and-suppress transformation to said
principal components of the agent data to generate said
background-suppressed principal components of the agent data.
3. The method of claim 1, wherein the step of deriving principal
components of sample spectral data comprises utilizing Principal
Component Analysis (PCA).
4. The method of claim 1, wherein said comparing step comprises
determining whether a spectral angle between a vector representing
the background-suppressed principal components of the sample data
and a vector representing the background-suppressed principal
components of the agent data lies within a pre-defined range.
5. The method of claim 4, further comprising indicating detection
of the agent in the bulk sample when said spectral angle lies
within said range.
6. The method of claim 4, further comprising indicating detection
of the agent in the bulk sample when said spectral angle lies with
said range and a dot product of said vectors exceeds a predefined
threshold.
7. The method of claim 1, further comprising: utilizing said
measurement modality to obtain spectral data for at least one
background constituent, deriving principal components of said
background spectral data, and generating said rotate-and-suppress
transformation based on said background principal components of the
background constituent data.
8. The method of claim 1, wherein said measurement modality
comprises any of fluorescence excitation-emission spectroscopy,
fluorescence lifetime spectroscopy, laser-induced breakdown
spectroscopy, Raman spectroscopy, Terahertz transmission or
reflection spectroscopy, and X-ray fluorerscence.
9. The method of claim 1, wherein said sample comprises a
bio-aeosol.
10. The method of claim 1, wherein said sample comprises any of a
biowarfare agent, a chemical warfare agent, a radiological agent,
an explosive agent, or a pollutant agent.
11. The method of claim 1, wherein said radiation comprises
ultraviolet radiation and said measurement modality comprises
fluorescence spectroscopy.
12. A method of processing spectral data taken from a sample,
comprising utilizing a plurality of measurement modalities to
interrogate the sample with electromagnetic radiation so as to
generate a plurality of sets of spectral data each corresponding to
one of said modalities, combining the plurality of sets of spectral
data into a composite feature vector, and transforming said
composite feature vector into a single principal component
vector.
13. The method of claim 12, wherein said combining step further
comprises applying a normalization factor to at least one of said
spectral data sets.
14. A method of detecting an agent in a bulk sample including at
least one background constituent, comprising: utilizing a plurality
of measurement modalities to interrogate the sample with
electromagnetic radiation so as to generate a plurality of sets of
spectral data each corresponding to one of said modalities,
combining the plurality of sets of spectral data into a composite
feature vector, transforming said feature vector into a principal
component vector, applying a rotate-and-suppress transformation to
said principal component vector, wherein said transformation
suppresses a contribution of said background constituent, if
present, to said vector, and comparing said transformed principal
component vector with a respective background-suppressed agent
vector to determine whether said agent is present in the bulk
sample.
15. The method of claim 14, wherein said combining step further
comprises applying a normalization factor to at least one of said
two spectral data sets.
16. The method of claim 14, wherein one of said data sets comprises
fluorescence emission-excitation data and the other data set
comprises fluorescence lifetime data.
17. The method of claim 16, wherein each fluorescence lifetime
datum is normalized utilizing the amplitude of the
emission-excitation space at which it was measured.
18. A system for detecting an agent in a bulk sample having at
least one background constituent, comprising: a background library
for storing data corresponding to at least one background
constituent, an agent library for storing principal components of
measurement data corresponding to said agent, a preprocessing
module in communication with said libraries and adapted to receive
measurement data corresponding to a test sample, said module
calculating a rotate and suppress transformation and applying said
rotate-and-suppress transformation to measurement data
corresponding to said test sample so as to generate transformed
principal component vectors corresponding to the test sample,
wherein said transformation suppresses said background constituent,
if present in the test sample, said preprocessing module further
applying said transformation to the agent so as to generate
background-suppressed principal component vectors corresponding to
the agent, and an agent detection module in communication with said
preprocessing module to receive said transformed principal
component vectors corresponding to the agent and the sample, said
agent detection module comparing said vectors to determine whether
said agent is present in the sample.
Description
RELATED APPLICATION
[0001] The present application claims priority to provisional
application No. 60/723,985 entitled "Agent detection in the
presence of background clutter," filed on Oct. 3, 2005, which is
herein incorporated by reference.
BACKGROUND OF THE INVENTION
[0003] The present invention relates generally to methods and
systems for detecting agents in a bulk sample that can include one
or more background constituents, and more particularly, for
detecting bio-aerosol warfare agents.
[0004] The detection of bio-aerosol warfare agents in the presence
of either indoor or outdoor backgrounds is a difficult problem.
Natural backgrounds are variable and can simultaneously include
multiple constituents. The variation of each constituent may be
larger than the concentration level of an agent whose detection is
desired. The detection problems can be further exacerbated by the
presence of spikes in measurement data of a naturally-occurring
background, which may be an order of magnitude larger than the
contribution of the normal quiescent background. Such spikes may
last for minutes and may exhibit large variations in particle
count.
[0005] The detection of other important agents share some of the
difficulties associated with the detection of bioaerosols. For
example, chemical warfare agents may need to be detected in the
presence of industrial cleaners or insecticides. Nuclear materials
may be hidden by background radiation from rocks and cements, as
well as by residual radiation from medical treatment or radiation
from shipments of medical equipment. Signatures of explosives
traces can be mimicked by foods preserved with nitrates as well as
by legitimate shipments of fertilizers. Detection of pollutants and
contaminants share the same problems as detection of biological,
chemical, and radiological warfare agents. A solution to all of
these problems requires the ability to detect low levels of agents
in an ambient environment. The detection sensitivity can be
increased by concentrating the sample to be analyzed, but at the
risk of having both large amounts of background and small amounts
of agent in the same sample. Further, simulants for a variety of
agents, such as those mentioned above, are often used for detector
development and testing. Thus, the detection of simulants is also
an important problem requiring solution.
[0006] Some workers in the field have attempted to solve the
problem of detecting low levels of agents against an ambient
background by finding signatures that are unique to the agents
whose detection is desired. This normally requires that signatures
of agents and background constituents be unique and
non-overlapping. This approach may work with signatures that have
many very narrow features, such as those typically exhibited by
LIBS (Laser Induced Breakdown Spectroscopy), Raman spectra, and
FTIR (Fourier Transform Infrared) spectra. However, it is not
suitable for signatures that have broad features, such as
UV-induced fluorescence spectra and lifetime measurements, x-ray
fluorescence spectra, and terahertz (THz) spectra. Hence, this
conventional approach has the disadvantage that it limits the
detection techniques that can be used to solve a given agent
detection problem.
[0007] Another conventional method for detecting agents utilizes
single particle flow-through systems, such as BAWS
(Biological-Agent Warning Sensor) to make a small number of
simultaneous measurements, a single particle at a time. Each
particle could be classified based on this small number of
measurements, and a histogram of particle locations in measurement
space could be built up over time.
[0008] However, single particle flow-through systems have several
disadvantages. First, the signal from a single particle is small.
Expensive hardware, such as large collection optics or more
powerful lasers to excite a larger return signal, can be used to
compensate for this low sensitivity. However, even with more
expensive hardware, the detection rate for very small particles is
generally negligible, leading to an inability to detect aerosols
composed of small particles (such as viruses), even if the particle
number density is large. Even for large particles, a low detection
rate can render sufficient data collection for a statistically
meaningful detection (build-up of a particle count that is
sufficient for agent detection) cumbersome and time-consuming.
Second, only a small number of measurements can be made
simultaneously. A small number of measurements implies a small
number of histogram bins. This can result, in turn, in placing
different particle types in the same histogram bin, leading to a
high false alarm rate. Finally, the flow of particles near the
large aperture collection optics of such systems can lead to
fouling of the optics, thus lowering the optical efficiency of the
system and driving up maintenance costs.
[0009] Accordingly, there is a need for enhanced methods and
systems for detecting agents in a variety of backgrounds.
SUMMARY OF THE INVENTION
[0010] The agent-detection methods of the present invention can be
utilized in conjunction with bulk collection and immobilization of
a sample under investigation to achieve greater sensitivity, lower
cost than conventional techniques, and to render a large variety of
measurements of the sample feasible. As discussed in more detail
below, the methods and systems of the invention allow suppressing
contributions of unwanted background constituents in a bulk sample
to information obtained about the sample (e.g., via spectral
measurements) by utilizing previously-obtained information about
the signatures of the agents and the background constituents of
interest. The present invention advantageously increases the types
of signatures that can be employed for agent detection.
[0011] In one aspect, the invention provides a method of detecting
an agent in a bulk sample, which includes at least one background
constituent, that comprises utilizing a measurement modality to
interrogate the sample with electromagnetic radiation so as to
generate spectral data corresponding to the sample. The principal
components of the spectral data can then be derived and represented
as a vector in a principal component space. A rotate-and-suppress
transformation can be applied to the principal component vector,
wherein the transformation suppresses, and preferably eliminates,
the contributions of the background constituent, if present,
thereby generating a background-suppressed principal component
vector corresponding to the sample. This transformed sample vector
can be compared with a background-suppressed principal component
vector corresponding to the agent to determine whether the agent is
present in the bulk sample. The background-suppressed principal
component vector of the agent can be generated by application of
the rotate-and-suppress transformation to the principal components
of the spectral data of the agent obtained by employing the same
measurement modality as that utilized to interrogate the
sample.
[0012] The above step of comparing the background-suppressed
principal component vectors of the sample and the agent can
comprise determining a spectral angle that separates the two
vectors in the principal component space. If the spectral angle is
within a pre-defined range, a detection of the agent in the sample
can be indicated. In some embodiments, this criterion constitutes
only one prong of a two-prong test for indicating the detection of
the agent within the sample. The other prong requires that the dot
product of the two vectors obtained, for example, in selected
two-dimensional sub-spaces of the principal component space, be
above a pre-defined threshold. In some embodiments, once it is
determined that the angle between the two vectors lies in a
predefined range, a warning is issued. If not only the angle is
within the range, but also the dot product is above a selected
threshold, an Alert/Alarm processor is triggered. Such a two-step
test procedure can lower the risk of issuing false alarms.
[0013] In a related aspect, the same measurement modality as that
utilized for obtaining the agent and the sample spectral data is
employed to generate spectral data corresponding to the background
constituent. The principal components of the background spectral
data are then derived and represented as a vector in the principal
component space. The following two transformations of the principal
component vector are then combined to form the aforementioned
rotate-and-suppress transformation. First, the principal component
vector is rotated so as to be aligned along an axis of the
principal component space (that is, it is rotated so as to include
only one component). Subsequently, the dimension of the principal
component space corresponding to that axis is eliminated. In other
words, the remaining component of the rotated vector is eliminated.
The combination of these two transformations generates the
rotate-and-suppress transformation.
[0014] As will be discussed, this method is equivalent to simply
projecting into a subspace along the background vector, so that the
vector is along an axis for computational convenience.
[0015] In many embodiments, Principal Component Analysis (PCA) is
utilized for deriving the principal components of spectral data
associated with the agents, background constituents, as well as the
samples under investigation. Principal Component Analysis (PCA) is
a well-known technique for determining the most important
components of multi-dimensional measurements of a collection of
agents and backgrounds. Using PCA, the dimensionality of the
measurement space can be reduced while maintaining the
distinguishing features of the original measurements.
[0016] Typically, the principal components for varying
concentrations of a given agent or background will form a straight
line in principal component space pointing toward the origin. The
origin is the location of the signature of the measurement
apparatus in the principal component space in the absence of any
agent or background. The straight line for each agent makes an
angle with other agents or background constituents. This "spectral
angle" can be defined for any number of principal component
dimensions. As the number of dimensions increases, the spectral
angle between randomly chosen vectors tends toward 90.degree..
[0017] In many embodiments of an agent-detection method of the
invention, the method begins by rotating the principal component
(PC) space to place the major background constituent normally seen
at the sample collection location into a single component in PC
space. Projection into a new PC space orthogonal to the major
background constituent is achieved by simply dropping that
component axis. The result is suppression of variations of that
background constituent but with loss of one dimension. The process
can be repeated for each background constituent that is larger than
a desired level of detection of the agent(s) of interest, as long
as there are remaining dimensions that can be discarded. The agent
library is rotated and projected in the same manner as the sample.
As noted above, the agent detection is determined by requiring that
the sample (the background-suppressed sample vector in the
principal component space) be within a small spectral angle of an
agent (the background-suppressed agent vector in the principal
component space), and that the dot product of the sample and agent
vectors simultaneously exceed a threshold determined by the level
of detection desired. False alarms can be reduced, at the expense
of response time, by requiring successive detections to declare an
alarm. We refer to this successive detections requirement as
"Alert/Alarm" processing.
[0018] A variety of measurement modalities can be employed in
agent-detection methods and systems of the invention. Some of these
measurement modalities include, without limitation, fluorescence
excitation-emission spectroscopy (e.g., UV-induced fluorescence
excitation-emission spectroscopy), fluorescence life-time
spectroscopy (e.g., UV fluorescence decay time histories), X-ray
fluorescence spectroscopy, laser-induced breakdown spectroscopy,
Raman spectroscopy, Terahertz transmission or reflection
spectroscopy, hyperspectral imaging, and performing optical
reflectance or scattering measurements. Agent and background
measurements can be made with the same or similar instruments as
those utilized to collect samples for detection of agent(s)
therein, or can be made with different instruments. Alternatively,
the methods of the invention can be applied to data collected and
disseminated by third parties. Further, as discussed below,
measurements corresponding to background constituents can be
obtained separately or can be determined by techniques such as
end-member analysis.
[0019] Further, the methods and systems of the invention can
applied to and utilized with a variety of sample collection
modalities. For example, these methods can be utilized to
interrogate an ambient volume of air or a fluid, or interrogate a
sample collected and concentrated into a fluid, or interrogate a
sample collected onto a substrate.
[0020] In another aspect, a method of the invention for detecting
an agent in a bulk sample having at least one background
constituent includes utilizing a plurality of (e.g., two)
measurement modalities to interrogate the sample with
electromagnetic radiation so as to generate a plurality of sets of
spectral data, each of which corresponds to one of those
measurement modalities. The sets of the spectral data are combined
into a composite data vector (herein also referred to as feature
vector), and that vector is transformed into a principal component
vector comprising the principal components of the data. The
principal component vector is then transformed, e.g., via
application of a rotate-and-suppress transformation thereto, so as
to generate a transformed principal component vector. The
transformation suppresses, and preferably eliminates, the
contribution of the background constituent, if present. The
transformed principal component vector can then be compared with a
background-suppressed principal component agent vector (a principal
component vector corresponding to the agent to which a
rotate-and-suppress transformation has been applied) to determine
whether that agent is present in the sample in a quantity that
would warrant indicating detection.
[0021] In a related aspect, in the above detection method, one or
more of the data sets are normalized prior to combining them into a
single data vector. In some embodiments, one of the measurement
modalities comprises fluorescence excitation-emission spectroscopy
and another comprises fluorescence life-time spectroscopy.
[0022] In some embodiments of the invention, by stringing together
the excitation-emission spectra and the lifetime measurements in a
suitable weighted vector, as discussed in more detail below, the
principal components can be calculated by applying a matrix
transformation to the vector.
[0023] In other aspects, the invention provides systems for
detecting agent(s) in a bulk sample. One such system can include a
spectrofluorometer with a lifetime measurement option that includes
modules for implementing the above analysis methods of the
invention. For example, such a system can include a background
library for storing data corresponding to at least one background
constituent, and an agent library for storing principal components
of measurement data corresponding to the agent. The system can
further include a preprocessing module in communication with the
libraries and adapted to receive measurement data corresponding to
a test sample. The module calculates a rotate-and-suppress
transformation and applies that transformation to measurement data
corresponding to the test sample so as to generate transformed
principal component vectors corresponding to the test sample. The
transformation suppresses the contribution of the background
constituent, if present. Further, the preprocessing module applies
the rotate-and-suppress transformation to principal component
vectors corresponding to the agent. An agent detection module,
which is in communication with the preprocessing module, receives
the transformed principal component vectors corresponding to the
sample and the agent, and compares those vectors to determine
whether the agent is present in the sample.
[0024] Further understanding of the invention can be obtained by
reference to the following detailed description, in conjunction
with the associated figures, described briefly below.
BRIEF DESCRIPTION OF THE DRAWINGS
[0025] FIG. 1A is a flow chart depicting various steps of an
exemplary embodiment of a method of the invention for detecting
agents in the presence of background clutter,
[0026] FIG. 1B is a schematic diagram of a system according to one
embodiment of the invention for implementing the method of FIG.
1A,
[0027] FIG. 2 schematically illustrates two different types of
measurements and a method for weighting and combining these
measurements into a single feature vector,
[0028] FIG. 3 graphically illustrates that the principal component
vectors of a specific agent or background constituent at different
concentrations (with the zero concentration at the origin) in a
background-subtracted PC space, indicating that the vectors form a
straight line through the origin,
[0029] FIG. 4 schematically depicts the suppression of a single
background constituent by rotation and projection transformation
according to the teachings of the invention (the resulting PC space
has one less dimension than the space before the rotation and
projection),
[0030] FIG. 5 depicts a geometric interpretation of the
rotate-and-suppress method for suppressing a background
constituent,
[0031] FIG. 6 schematically depicts criteria for indicating
detection of an agent based on a spectral angle separating a test
sample's background-suppressed PC vector from a respective vector
of the agent and a predefined threshold, and
[0032] FIG. 7 schematically depicts an apparatus for performing
fluorescence excitation-emission and lifetime spectroscopy.
DETAILED DESCRIPTION
[0033] The present invention provides generally methods and systems
for detecting an agent in a sample under investigation in the
presence of background clutter (e.g., in the presence of
interfering background constituents). An agent detection method
according to the teachings of the invention can suppress the
contributions of one or more background constituents to spectral
data obtained from the sample so as to facilitate detection of an
agent in that sample. The agents of interest can be, without
limitation, bio-warfare agents, chemical warfare agents,
radiological agents, explosives, or pollutants (e.g., heavy
metals). As discussed in more detail below, a detection method of
the invention is capable of suppressing contributions from
multiple, independent background constituents to measured data
obtained from a sample. Further, a background suppression and
detection method of the invention can adapt to the background in
any location or season, and can suppress both the normal quiescent
contributions of a background as well as those contributions that
are manifested as large, variable spikes. In many embodiments
discussed below the teachings of the invention are illustrated
through their application to excitation-emission spectra and time
histories (lifetime) of UV-induced fluorescence. It should,
however, be understood that these teachings are equally applicable
to any detection problem where a large number of measurements of
agents, backgrounds and samples to be tested can be made.
[0034] With reference to a flow chart 10 of FIG. 1A, an exemplary
embodiment of a detection method of the invention can include
obtaining measurements of samples of ambient background
constituents (i.e., a background in which the detection of an agent
is desired) as well as those of one or more agents of interest
(step A). In many embodiments, these measurements comprise
interrogating the background constituents and the agents with
electromagnetic radiation to obtain spectral data corresponding
thereto.
[0035] The spectral data corresponding to each background
constituent and each agent can be subjected to a Principal
Component Analysis (PCA) to derive its associated principal
components (step B). For each background constituent, a
rotate-and-suppress transformation can derived based on its
principal components that can be applied to spectral data obtained
for the agent(s) or test samples of interest to suppress the
contribution of that background constituent to the data (step C).
These rotate-and-suppress transformations can then be combined,
e.g., via multiplication of matrices representing them, to obtain a
composite background-suppression transformation (step D). The
rotate-and-suppress transformations can be stored in a library to
be retrieved and utilized when measurements are made on test
samples, in a manner discussed below.
[0036] For example, the composite background-suppression
transformation can be applied to principal components of each
agent, that is to the principal component vector corresponding to
each agent, to generate a plurality of background-suppressed
principal component agent vectors (herein also referred to as
reference agent vectors) (step E).
[0037] The composite transformation for suppressing the background
contributions as well as the reference agent vectors can then be
utilized in the following manner to determine whether one or more
of the agents are present in a test sample. A test sample can be
interrogated by employing the same measurement modality as that
utilized for collecting data for the background constituents and
the agents so as to generate spectral data corresponding to that
test sample (step F). The spectral data can then be transformed
into the principal component space (step G). For example, the
spectral data can be subjected to a PCA analysis so as to generate
its principal components, which can be represented as a vector in
the principal component space.
[0038] The aforementioned composite background rotate-and-suppress
transformation can then be applied to the principal component
vector of the test sample so as to suppress contributions from the
background constituents thereto (step H). The application of the
background rotate-and-suppress transformation to the sample's
principal component vector results in generating a
background-suppressed principal component vector corresponding to
the test sample. That background-suppressed principal component
vector can then be compared with any of the background-suppressed
PC agent vectors to determine whether that agent is present in the
test sample (step I). For example, the presence of an agent in the
test sample can be established if an angle between the test
sample's background-suppressed principal component vector and a
background-suppressed PC agent vector lies within a predefined
range and the dot product of the two vectors exceeds a predefined
threshold.
[0039] FIG. 1B schematically depicts an exemplary system according
to one embodiment of the invention that implements the above method
for detecting agent(s) of interest in the presence of background
clutter. The system 10 includes preprocessing modules 12a, 12b, 12c
that receive, respectively, measurement data corresponding to one
or more agents of interest (A), measurement data for a sample under
investigation (S) and measurement data for one or more background
constituents (B). Each preprocessing module generates the principal
components of the data it receives. Typically, the principal
component transformation applied to each of the data streams will
be identical. Although the preprocessing modules are shown as
separate entities, in some embodiments they can constitute a single
logical entity that can perform the functionality of the three.
[0040] Further the preprocessor 12c generates background principal
components and communicates these to a background library 14 for
storage therein. The preprocessor 12a, in turn, transmits the
principal components of the agents of interest to an agent library
16 for storage. These libraries are then utilized in a manner
discussed below to analyze data obtained from a test sample for
detection of one or more of the agents corresponding to those in
the agent library.
[0041] More specifically, the preprocessor 12b communicates the
principal components of the data obtained for a sample under
investigation to a background suppression module 18. The background
suppression module 18, in turn, retrieves the background
constituents (data corresponding to these constituents) from the
background library 14, calculates the respective rotate and
suppress transformations, and applies these transformations to the
principal components of the test sample data, as well as the
principal components of one or more agents of interest, retrieved
from the agent library. An agent detection module 20 then receives
the background-suppressed principal component vectors corresponding
to the test sample and the agents of interest and compares these
vectors in a manner discussed above to determine whether one or
more agents of interest are present in the sample. Upon
establishing detection of one or more agents, the agent detection
module can trigger an alert/alarm processing unit 22 to issue an
alert/alarm.
[0042] In many embodiments, the agent and the background libraries
are compiled prior to obtaining and/or analyzing measurement data
corresponding to a sample under investigation. In other
embodiments, the compilation of the libraries and obtaining and/or
analyzing the sample's measurement data can be performed
concurrently.
[0043] In one embodiment of the invention, fluorescence
EXcitation-EMission spectra and fluorescence Lifetime data are
captured for agents, backgrounds and real-time test samples. These
data are herein referred to as XML measurements. Other types of
measurements can include, without limitation, optical reflectance
or scattering measurements, x-ray fluorescence spectra,
laser-induced breakdown spectroscopy (LIBS) spectra, Raman spectra,
or Terahertz transmission or reflection spectra. XML measurements
can be preprocessed, e.g., prior to interrogating the test samples,
to create agent and background libraries, such as those discussed
above. The agent and background XML measurements can be
preprocessed to combine them into a single data vector for each
measurement (e.g., a single data vector corresponding to the
background XML measurements and a single data vector corresponding
to the agent XML measurements can be formed), and then the
principal components of the set of measurements can be determined
using standard methods, e.g., Principal Component Analysis
(PCA).
[0044] The XML measurements can be converted to a single data
vector for further analysis by using, for example, the method shown
schematically in FIG. 2. For example, in this exemplary embodiment,
excitation-emission spectrum intensities ("I.sub.1" through
"I.sub.29") are "unwrapped" column by column into a single vector.
Time history data (lifetime measurements denoted by "L.sub.1"
through "L.sub.7") can then be appended to that vector, but the
lifetime measurements are re-normalized, as denoted by the primes
in the vector at the bottom of FIG. 2. In this embodiment, the
normalization is chosen so that the area under the lifetime curve
is equal to the steady state intensity from the location in
excitation-emission space where the lifetime measurements are
taken. This normalization allows the two measurements to be
combined on the same footing. The actual number of
excitation-emission spectral points and time history points may
vary from those shown in the figure.
[0045] Each XML measurement taken can be converted to a single data
vector using the above procedure. Subsequently, a set of XML
measurements that comprise, for example, measurements of different
analytes (agents, simulants, or interferents) taken at different
concentrations in multiple replicates, can be analyzed by utilizing
Principal Component Analysis (PCA), which is a data analysis method
known in the art. For example, it is described by Sharaf, et al. in
Chemometrics (Wiley & Sons, New York) 1986, which is herein
incorporated by reference. PCA can provide a standard eigenvector
decomposition of the XML vector space, with the vectors (the
"principal components") arranged in the order of their eigenvalues.
There are generally far fewer meaningful principal components than
nominal elements in the data vector, because neighboring
fluorescent wavelengths are typically very highly correlated. The
number of meaningful principal components is a measure of how many
real independent variables are present in the XML space.
[0046] A striking result is seen when the multiple replicates at
different concentrations are plotted in principal component (PC)
space: each analyte forms a vector as shown in FIG. 3. When the
signature of the substrate or medium containing the agent or
background is subtracted from each vector, the vectors emanate from
the origin and each vector's length depends on concentration of its
associated analyte. This fact permits a vector-based analytical
approach for detecting analytes (agents) of interest in a test
sample. More specifically, each analyte has a characteristic
vector, and how different one analyte is from another can be
measured by how far apart their respective vectors are. Subtraction
of the baseline spectrum of the substrate or medium containing the
agents and backgrounds allows the use of different collection and
concentration methods, including deposition onto a substrate of
various materials (e.g., a metal surface or Teflon tape),
collection and concentration into water or other liquid such as
ethylene glycol, or interrogation of a large volume of air or water
containing unconcentrated agents and backgrounds.
[0047] Normally, measurements of agents will be performed in a
laboratory setting for safety reasons. However, measurement data
for simulants and background may be obtained in either a laboratory
or ambient natural setting. Ideally, replicates of each agent,
simulant, or background constituent will have been measured
separately as described above. In some cases, it may be desirable
to measure the background in a natural setting in which the
background constituents are mixed together in independently varying
amounts. In such a case, techniques well-known in the field of
multispectral analysis can be used to determine the individual
background constituents, typically known as "endmembers." Some
examples of such techniques are described in an article entitled "A
Survey of Spectral Unmixing Algorithms," by N. Keshava published in
Lincoln Laboratory Journal 14 (2003), which is herein incorporated
by reference. The endmembers determined from a reasonably large set
of measurements can become the background library for the local
ambient setting. It should be obvious that either a background
library measured in a laboratory setting or a background library
determined from ambient measurements can be used in the agent
detection and background suppression methods of the invention.
[0048] The principal components of agents and backgrounds may have
very different values; typically there are a few strong components
and a larger number of weaker components. In some cases, it may be
desirable to scale the axes of the principal component space for
display or prior to subsequent computations. For example, the axes
can be scaled so that each PC axis has the same dynamic range. This
would be equivalent to autoscaling the axes in a plot. By way of
example, the axes can be scaled so that measurement noise is equal
on each axis. In this case, measurement noise would form a
hypersphere in the PC-space following scaling.
[0049] The rotate-and-suppress (RAS) method makes use of the fact
that any specific background constituent forms a vector in
PC-space. The method works by rotating the multidimensional
PC-space so that a background constituent PC-vector is contained in
a single component, then suppressing the background by dropping
that component from the PC-space. By way of example and only for
illustrative purposes, FIG. 4 schematically shows how RAS operates
in two dimensions to suppress a single background constituent. Note
in the figure that the background constituent B.sub.1 is
eliminated, and that in the process one dimension of the principal
component space is lost. Typically, the RAS process starts
operating in a much higher dimensional space, and the RAS process
is applied several times in succession to eliminate multiple major
background constituents. It should be noted that collection of more
fluorescence data points, leading to a larger number of principal
components, provides the ability to suppress a larger number of
background constituents.
[0050] A geometric interpretation of a method of the invention will
help clarify the mathematical description which follows. FIG. 5
shows a perspective view of a background constituent vector and an
agent vector in 3-space. Captured data have shown that the length
of the background vector varies over a very large dynamic range,
with spikes of unpredictable length that can be an order of
magnitude larger than the variations in the recent past. If we
assume that the background constituent has been characterized so
that its direction is known, the effect of these large,
unpredictable variations can be almost completely eliminated.
Imagine, in 3-space, that an observer positions himself so that his
line of sight is exactly along the background vector. In this case,
the variation in the background will be unseen by the observer. Any
agent not exactly parallel to the background will have a component
perpendicular to the line of sight which can be seen by the
observer. Moving the observer's point of view to be along the line
of sight of the background vector, and then detecting agent vectors
perpendicular to this line of sight is exactly analogous to the
Rotate and Suppress method to be described further below. Those
having ordinary skill in the art will appreciate how to extend this
geometrical argument in 3-space into higher dimensional spaces.
[0051] In general, in a method of the invention to rotate a
background vector B.sub.1=.left brkt-bot.B.sub.1, B.sub.2, . . .
B.sub.i, B.sub.j, . . . B.sub.N.right brkt-bot. in the principal
component space (herein also referred to as principal component
background vector) so as to set its i.sup.th component, namely,
B.sub.i to zero, the following quantities are computed:
D= {square root over (B.sub.i.sup.2+B.sub.j.sup.2)}
C=B.sub.j/D
S=-B.sub.j/D
The rotated background vector can then be given by:
B.sub.1'=.left brkt-bot.B.sub.1, B.sub.2, . . .
0,-SB.sub.i+CB.sub.j, . . . B.sub.N.right brkt-bot.
The rotation matrix for the rotation described above is given
by:
R 1 = [ 1 0 0 0 1 C S - S C 0 1 ] ##EQU00001##
where the diagonal elements not shown are equal to unity, and the
off-diagonal elements not shown are equal to zero. The "C" entries
are in elements (i,i) and (j,j) and the "S" entries are in elements
(i,j) and (j,i). This method can be performed multiple times to set
all but one of the background components to zero. The
transformation matrices for each rotation can be combined using
standard matrix multiplication techniques to compute the overall
transformation.
[0052] Once there is only a single non-zero component remaining in
rotated background vector B.sub.1', that component can be
eliminated from all rotated library entries for both agents and
background constituents. The combined rotation matrix is given
by:
R C 1 = R N - 1 .times. R N - 2 .times. .times. R 1 = [ r 11 r 1 N
r N 1 r NN ] ##EQU00002##
When the combined rotation matrix is applied to the principal
components of the real-time test samples and the non-zero component
of the background is dropped, any contribution due to background
constituent B.sub.1 will be eliminated. If the row of the combined
rotation matrix corresponding to the non-zero component of the
background is dropped, then the resulting transformation matrix
will perform the entire operation when it multiplies any set of
principal components derived from agent samples, background
samples, or real-time samples for testing. This transformation
matrix is given by
T C 1 = [ r 11 r 1 , N r N - 1 , 1 r N - 1 , N ] ##EQU00003##
[0053] A second background constituent can be eliminated in the
same fashion. In this case, the exemplary background constituent
has one fewer component:
B.sub.2'='B.sub.1, B.sub.2, . . . B.sub.i, B.sub.j, . . .
B.sub.N-1,
as do the resulting rotation matrices. Note that the use of the
prime indicates that the components of the second background vector
to be eliminated should be from the rotated background library.
[0054] The RAS rotations are determined by the agent and background
libraries measured at any particular location and season. The
combined rotation determined by the background library is applied
to the collected samples in real time by a matrix multiplication
which combines all the rotations and dropping of components into a
single matrix calculated ahead of time. This combined
transformation for removal of M background constituents is given
by
T.sub.C=T.sub.CM.times. . . . .times.T.sub.C1
[0055] Subsequent multiplication of the principal components
derived from data (e.g., spectral data) obtained for a sample to be
tested by the combined matrix T.sub.C results in the background
suppression of those components. Because this operation comprises a
single matrix-vector multiplication and the matrix T.sub.C can be
computed ahead of time, it can be done in real-time with no
significant computational burden.
[0056] Agent detection can then performed in the rotated Principal
Component space (PC-space) in the following manner. PC-space with
more than three dimensions can only be visualized using subspace
graphs. Measurements of UV-induced fluorescence intensity at many
excitation-emission wavelength pairs will provide at least seven
principal components, and the addition of fluorescence lifetime
information will provide at least two more dimensions. The number
of principal components will likely increase as more analytes are
added to the libraries. With a seven- to nine-dimensional space, a
large number of subspaces need be explored to determine whether two
analytes are well-separated. The Spectral Angle (SA) is a
one-dimensional metric which can be used to determine the
separation of agents or simulants from backgrounds.
[0057] For two multidimensional vectors a and b, the spectral angle
between them is given by:
SA ( a , b ) = cos - 1 [ a b a b ] . ##EQU00004##
The two vectors are aligned when the normalized dot product (the
quantity in the square brackets) is equal to one. This condition is
met only when the normalized components of each vector are equal.
As the components become unequal (or even of opposite sign) the
quantity in the brackets can approach zero, in which case the
spectral angle approaches 90.degree. and the vectors are
orthogonal.
[0058] A detection method of the invention compares the
background-suppressed test samples with the agents in the library
to determine whether one or more of those agents are present in the
test sample. Note that the agent library has been subjected to the
RAS matrix transformation. FIG. 6 schematically shows the
application of exemplary criteria that can be utilized for
determining a match to an agent, i.e., for determining whether an
agent is present in the test sample. First, a spectral angle (SA)
between the background-suppressed principal component vectors
corresponding to the test sample and the agent of interest is
computed. If the SA is greater than a preset value (e.g.,
30.degree.) no detection occurs. If the SA is within a predefined
range, e.g., the cone defined by the SA threshold shown by the
dotted lines in FIG. 6, then the dot product of the sample and the
agent is computed. When the dot product exceeds a second preset
threshold, detection occurs. By way of example, in FIG. 6, S.sub.1
(on the left) and S.sub.2 (in the middle) fall outside the
detection region, but S.sub.3 (on the right) is a detected
event.
[0059] In some embodiments, when a detection occurs, an Alert/Alarm
is issued, e.g., the Alert/Alarm processor (rightmost box in FIG.
1B) is triggered. By requiring two successive detection events
(Alert then Alarm) all single event detections can be eliminated,
significantly reducing the probability of false alarm without
affecting the probability of detection.
[0060] As noted above, the agent detection methods of the invention
can be utilized in connection with a variety of measurement
modalities, and are not limited for use with fluorescence
excitation-emission and lifetime measurements. As such, they can be
utilized to process data obtained from a variety of measurement
devices. For example, in the above exemplary embodiment, the
processing of XML measurements was discussed for illustrating the
features of an exemplary embodiment of an agent detection method of
the invention. An embodiment of an apparatus for gathering XML
measurements is shown in FIG. 7. The exemplary apparatus includes a
single sample chamber (4) serviced by both a steady-state
spectrofluorometer (top of figure) and a lifetime
spectrofluorometer (bottom of figure). This apparatus is available
commercially from Photon Technology International, Birmingham,
N.J.
[0061] In the steady-state spectrofluorometer, a Xenon lamp (1)
provides illumination that is focused onto a slit (2) and thence
enters an excitation monochromator (3). It then enters a second
slit (2) that defines the wavelength spread, and into a sample
chamber (4). The wavelength-selected excitation light is focused
via optics (7) onto a cuvette, which holds a sample of interest
(not shown), and is held in a cuvette holder (8). Fluorescence
radiation (9), emitted by the sample in response to the excitation
light, is then focused onto slit (2a) and enters an emission
monochromator (11). Passing through the emission monochromator,
emission radiation in the selected waveband is focused onto another
slit (2b) and enters a photomultiplier tube (12), where the
steady-state intensity of the selected emission waveband from the
selected excitation is detected.
[0062] In the lifetime spectrofluorometer, shown at the bottom of
the drawing, the Xenon arc lamp from the steady-state system is
replaced by a short pulse of light delivered by a laser (13). The
arrival of the light pulse is delayed by the fiber optic (14) for a
sufficiently long time to enable gating electronics (not shown) to
set up a moving time gate on the photomultiplier tube output
relative to the excitation pulse. Subsequent to the laser light
pulse arrival, fluorescence photons are emitted by the sample and
pass through the same emission path as before and strike the
photomultiplier. The time gate which moves relative to the arrival
of the excitation pulse enables the decay of fluorescence output
relative to excitation to be directly measured.
[0063] The fluorescence excitation-emission data as well as
fluorescence lifetime data generated by the above apparatus can
then be analyzed in a manner discussed above to determine whether
one or more agents of interest are present in the sample under
investigation.
[0064] To further illustrate the salient features of agent
detection methods of the invention, the following hypothetical
example is provided. It should be understood that the example is
provided only for illustrative purposes and is not intended to
necessarily indicate the dimensionality of the data vectors or that
of the principal component space.
EXAMPLE
[0065] A simple numerical example will further illustrate the
operation of the Rotate and Suppress method according to the
teachings of the invention. In this example, the agent and
background vectors are defined in three-dimensional space, and the
suppression of the background vector results in a two-dimensional
space. Assume that the Principal Components of a background
constituent and two agents of interest are specified by:
B.sub.1=[3,4,5]
A.sub.1=[1,0,2]
A.sub.2=[1,2,0]
Application of the first rotation results in setting the first
component of B.sub.1 to zero. More specifically, the following
rotation matrix rotates all of the first component of the
background into the second component:
R 1 = [ 4 / 5 - 3 / 5 0 3 / 5 4 / 5 0 0 0 1 ] ##EQU00005##
and the resulting rotated background vector is given by:
B 1 ' = R 1 .times. B 1 T = [ 0 5 5 ] . ##EQU00006##
In the second rotation, the second component of B.sub.1' is set to
zero, without reinserting any signal into the first component. This
second rotation matrix is given by:
R 2 = [ 1 0 0 0 2 / 2 - 2 / 2 0 2 / 2 2 / 2 ] ##EQU00007##
The combined rotation matrix is given by the product of R.sub.2 and
R.sub.1 as follows:
R C = R 2 .times. R 1 = [ 4 / 5 3 / 5 0 3 2 10 2 2 5 - 2 / 2 3 2 10
2 2 5 2 / 2 ] ##EQU00008##
The result of applying R.sub.2 to B.sub.1' is:
B.sub.1''=.left brkt-bot.0,0,2 {square root over (5)}.right
brkt-bot..
The transformation matrix T.sub.C is then obtained from R.sub.C by
dropping the third row of R.sub.C:
T C 1 = [ | 4 / 5 - 3 / 5 0 3 2 10 2 2 5 - 2 2 ] . ##EQU00009##
Note that for N Principal Components and M background constituents
to be suppressed, the final transformation matrix will have
dimensions (N-M, N). Applying T.sub.C to the agent and background
Principal Component vectors yields:
T C 1 .times. B 1 = [ 0 0 ] ##EQU00010## T C 1 .times. A 1 = [ 4 /
5 - 7 2 10 ] ##EQU00010.2## T C 1 .times. A 2 = [ - 2 / 5 11 2 10 ]
##EQU00010.3##
[0066] The background constituent has been completely suppressed by
this transformation, and the two agent Principal Component vectors
have been rotated into a new two-dimensional space and comprise the
agent library to be used for detection. Note that the rotation
matrices and thus the transformation matrices are independent of
the magnitude of the background vector and depend only on the
direction of the background vector. Thus, the library of background
constituents may be normalized to unity, or to any other value, or
may be used without normalization.
[0067] The teachings of the following publications are herein
incorporated by reference: [0068] D. Manolakis, D. Marden, and G.
A. Shaw, "Hyperspectral Image Processing for Automatic Target
Detection Applications," Lincoln Laboratory Journal 14 (2003) p.
79. [0069] N. Keshava, "A Survey of Spectral Unmixing Algorithms,"
Lincoln Laboratory Journal 14 (2003) p. 55. [0070] C. A.
Primmerman, "Detection of Biological Agents," Lincoln Laboratory
Journal 12 (2000) p. 3. [0071] T. H. Jeys, "Aerosol Triggers," New
England Bioterrorism Preparedness Workshop (3-4 Apr. 2002). [0072]
J. R. Lakowicz, Principles of Fluorescence Spectroscopy (Kluwer,
New York) 1999. [0073] M. A. Sharaf, D. L. Illman, and B. R.
Kowalski, Chemometrics (Wiley & Sons, New York) 1986. [0074]
Applied Optics, "Laser-Induced Breakdown Spectroscopy," (feature
issue) 20 Oct. 2003. [0075] Existing and Potential Standoff
Explosives Detection Techniques, National Research Council (The
National Academies Press, Washington D.C.) 2004. [0076] L. S.
Powers and C. R. Lloyd, "Method and Apparatus for Detecting the
Presence of Microbes and Determining their Physiological Status,"
U.S. Pat. No. 6,750,006, Jun. 15, 2004. [0077] L. S. Powers,
"Method and apparatus for sensing the presence of microbes," U.S.
Pat. No. 5,968,766, Oct. 19, 1999. [0078] L. S. Powers, "Method and
apparatus for sensing the presence of microbes," U.S. Pat. No.
5,760,406, Jun. 2, 1998. [0079] T. H. Jeys and A. Sanchez,
"Bio-particle fluorescence detector," U.S. Pat. No. 6,194,731, Feb.
27, 2001.
[0080] Those having ordinary skill in the art will appreciate that
various modifications can be made to the above embodiments without
departing from the scope of the invention.
* * * * *