NonGaussian Data Fusion System NGDFS

1
Non-Gaussian Data Fusion System (NGDFS)
Daniel H. Wagner Associates, Inc. -
www.wagner.com
The Non-Gaussian Data Fusion System (NGDFS) uses
advanced multiple hypothesis, Gaussian sum and
Non-Gaussian tracking, non-Gaussian registration,
and Bayesian inference techniques to
significantly improve the ability of United
States forces to conduct search and surveillance.
The Non-Gaussian tracking is accomplished through
the use of Monte Carlo (aka particle filters,
cellular automata, intelligent agents) methods.
Each target distribution is represented by a
collection of thousands of subtargets, each one a
potential realization of the actual target.
Thus, target tactics or other sophisticated types
of non-linear motion can be modeled directly by
assigning an appropriate path to each subtarget.
Each subtarget is given a weight (initially
uniform) and both positive information in the
form of sensor contacts, and negative information
in the form of lack of contact by a given sensor,
can be incorporated into the target distribution
estimate by adjusting the weights using Bayes
Theorem.
SQQ-89 Improved Performance Sonar (IPS) Data
Fusion Functional Segment running on a Low Mass
Workstation
The key advantages of NGDFS over existing data
fusion/mission planning systems derive from its
use of Non-Gaussian techniques. These advantages
include the ability to (1) More accurately
determine which contacts are associated with
which targets, (2) Provide better estimates of
the position and velocity of targets of interest,
(3) Allow the generation of optimal resource
allocation plans, (4) Provide much more accurate
modeling of line-of-bearing data, including the
use of signal propagation information, (5) Allow
for the use of negative information from sensors
that are not detecting the target, which in many
cases can be nearly as valuable as positive
information when estimating target position, and
(6) Allow for the accurate modeling of complex
target motion and tactics such as a submarine
attempting to avoid active prosecution.
NGDFS is used on-board the USS Paul Hamilton, USS
John S. McCain, USS Decatur, and USS Milius to
track submarines, surface ships, and aircraft as
part of the SQQ-89 Improved Performance Sonar
(IPS) Data Fusion Functional Segment (DFFS)
(currently supported in part by a Phase II
ONR/NAVSEA SBIR program). It is also being used
to

• Track personnel and vehicles attempting to cross
  the U.S./Mexican border (for Department of
  Homeland Security)
• Track vehicles and dismounted personnel (for an
  AFRL Phase II SBIR program)
• Manage Contacts for the Littoral Combat Ship
  (LCS) Anti-Submarine Warfare (ASW) Mission
  Package (for NAVSEA PMS-420)
• Classify, localize, and generate an optimal
  response to torpedo attacks (for ONR/NAVSEA)
• Generate a collision avoidance fused tactical
  picture for Unmanned Surface Vehicles (USVs) (for
  NAVSEA PEO Ships)

Distribution A Approved for Public Release
Distribution is unlimited.
2
Near-Real-Time Data Fusion (NRTDF) Testbed
Daniel H. Wagner Associates, Inc. -
www.wagner.com
NRTDF is a full-capability data fusion engine
that utilizes both linear (Kalman Filter) and
non-linear (Monte Carlo) tracking models. The
Kalman filter is used for all tracks and is
supplemented by the Monte Carlo model for
high-interest and/or low reporting rate targets.
NRTDF has been used to fuse MTI radar data and
SIGINT reports as well as air radar data and
undersea warfare data. The Fusion Testbed has
been created to test and demonstrate the NRTDF
for different scenarios. The testbed has an
easy-to-use GUI and database (yellow) that allows
an analyst to quickly and easily create, modify,
store, and run scenarios with the correlator. It
also has high-fidelity sensor simulations for
radar, ESM, acoustic and other sensors, along
with a special modules to evaluate the
performance of the correlator (red).
The components of the NRTDF correlator are the
input processor, the multiple-hypothesis
correlator and Kalman Filter tracker, the
Non-Gaussian tracker, a sensor optimizer, and an
output process (green).
The core of the NRTDF is a Multi-Hypothesis
correlator, called MATCH, that can handle
multiple sensor types, multiple platforms,
out-of-sequence reports, and both kinematic and
attribute-based sensors. The importance of
multiple hypothesis method is important because
it allows MATCH to correct improper associations
quickly. When an association is ambiguous, MATCH
creates multiple hypotheses and holds the most
likely ones in memory. The most likely
hypothesis is displayed to the operator but, when
new data indicate a change, the correct
hypothesis is quickly substituted. The example
shown below, created in the testbed, demonstrates
this capability. There are two friendly ships
with ESM capability both reporting to a common
correlator. There are two hostile ships with
emitting radars, beyond radar range of either
friendly. Both friendly ships gain passive
contact on both enemies.
When tracking begins, the correlator picks the
wrong pair of intersections.
Because both lines of bearing cross at two
locations, there are two hypotheses for which
pair of intersections reflects the two real
targets.
After a short time, the motion information makes
the other pair of intersections more likely and
MATCH switches the scenarios immediately.
NRTDF was supported by a Phase II SBIR from NSWC
Dahlgren Division. Distribution A Approved for
Public Release Distribution is unlimited.
3
Operational Route Planner (ORP)
Daniel H. Wagner Associates, Inc. -
www.wagner.com
ORP was initially designed to find near-optimal,
environmentally responsive, search paths for
anti-submarine warfare. ORP can design paths for
either acoustic or non-acoustic sensors. The
criterion for optimality is the cumulative
detection probability (CDP) against a target
obeying a given, probabilistic, motion model.
The problem of allocating search effort against a
moving target was essentially solved in 1980,
under the simplifying assumption that effort
could be spread arbitrarily across the search
region (Brown, 1980). ORP addresses the much
harder problem where the searcher is constrained
to a physically realizable path. The
searcher-path problem is known to be NP-hard
(i.e., at least as hard as the traveling salesman
problem), and no computationally feasible
solutions are known to any NP-hard problem (many
researchers believe such solutions to be
impossible). ORP uses a genetic algorithm (GA)
to find near-optimal plans. A GA mimics the
process of evolution. The first step is to
define an encoding of trial solutions into
chromosomes (strings of tokens, or genes) and a
fitness function. Each chromosome defines an
individual. The fitness function is some
transformation of the objective function (CDP in
this case). The initial population consists of
many randomly generated individuals (trial
solutions). Succeeding generations are formed by
selecting parents (with probability proportional
to their fitness), randomly interchanging
portions of their chromosomes to form children,
and randomly mutating some of the childrens
genes. As in nature, the overall fitness of the
population tends to grow with successive
generations.
Mannually Designed
ORP-Generated
CDP 0.472
CDP 0.475
A moderately simple problem Red 10 nm
range Blue 1 nm range Fixed starting point
ORP evaluates the fitness function (CDP) by a
Monte Carlo simulation of the targets motion
model. The model can be a simple random tour,
or it can include sophisticated patrol and
transit tactics. For acoustic searches, ORP
obtains signal excess from detailed acoustic
predictions provided by Navy standard acoustic
models. The models can account for
inhomogeneous, non-isotropic environments. For
magnetic anomaly detection (MAD), ORP uses a
detailed internal model for the MAD signal, given
the Earths local magnetic field and the magnetic
characteristics of the submarine. ORP combines
the signal calculation with atmospheric noise
values (ap index) and geomagnetic noise maps
provided by NRL/SSC to obtain a probability of
detection on each search leg. These are combined
to yield CDP. A representative (relatively
complex) example is illustrated below.
Colors indicate sweep width (2 times mean
detection range) Fixed starting point Blue
standard ladder search Green best ladder path
of 500 random orientations and spacings Black
ORP track
ORP development has been supported by the Naval
Research Laboratory, Stennis Spacecenter
(NRL-SSC) and the Office of Naval Research and is
currently supported by NAVAIR and the NAVSEA
Undersea Warfare-Decision Support System
(USW-DSS) program. Distribution A Approved for
Public Release Distribution is unlimited.