Fault-Detection Tool Has Companies
‘Mining’ Own Business
Computer Technology
Originating Technology/ NASA Contribution
A successful launching of NASA’s Space Shuttle
hinges heavily on the three Space Shuttle Main
Engines (SSME) that power the orbiter. These critical
components must be monitored in real time, with
sensors, and compared against expected behaviors
that could scrub a launch or, even worse, cause
in-
flight hazards.
Since 1981, SSME faults have caused 23 scrubbed
launches and 29 percent of total Space Shuttle
downtime, according to a compilation of analysis
reports. The most serious cases typically occur
in the last few seconds before ignition; a launch
scrub that late in the countdown usually means
a period of investigation of a month or more. For
example, during the launch attempt of STS-41D in
1984, an anomaly was detected in the number three
engine, causing the mission to be scrubbed at T-4
seconds. This not only affected STS-41D, but forced
the cancellation of another mission and caused
a 2-month flight delay.
In 2002, NASA’s Kennedy Space Center, the Florida
Institute of Technology, and Interface & Control
Systems, Inc., worked together to attack this problem
by creating a system that could automate the detection
of mechanical failures in the SSMEs’ fuel control
valves.
Partnership
Indialantic, Florida-based Interface & Control
Systems was awarded a Kennedy Small
Business Technology Transfer (STTR) contract to develop the
failure-detection system. Using data-mining techniques,
the company is able to extract behavioral characteristics
from past performance data and automate real-time
monitoring and controlling with a high degree of
accuracy.
|
A remote camera
captures a close-up view of a Space Shuttle
Main Engine during a test firing at the Stennis
Space Center. In addition to pre-flight testing,
NASA is capable of monitoring the performance
of Space Shuttle Main Engines during flight,
allowing for real-time, autonomous detection
of potential mechanical failures. |
What resulted from the STTR was the combination
of two cross-cutting technologies: Florida Institute
of Technology developed adaptive machine learning
algorithms which interact with graphical front-end
tools and an artificial intelligence tool developed
by Interface & Control Systems’s. This tool,
known as Spacecraft Command Language (SCL) software,
is a rule-based expert software system used for
real-time monitoring and control. Utilizing the
adaptive machine learning algorithms, Interface & Control
Systems can mine historical data to define relationships
and discover operational signatures for groups
of related sensors.
With the SCL expert technology, the company can
build data-processing control systems that run
on embedded systems for autonomous applications
in a wide variety of environments. These systems
are capable of monitoring thousands
of rapidly changing sensors and directing hundreds
of actuators based on complex, machine-generated
logic.
Prior to the STTR partnership, there was no evidence
of cooperation between a rule-based expert system
and adaptive machine learning technology within
the production, launch, or flight environments
of the aerospace industry, according to Interface & Control
Systems. This new offering vastly improves NASA’s
abilities to autonomously conduct SSME analysis
in real time, while still keeping humans “in the
loop”—minus the hassle of having to carry out mundane,
manual tasks that can easily be handled by the
failure-detection technology.
Product Outcome
Built on top of the SCL architecture, Interface & Control
Systems’s SensorMiner is a time-series data-mining
tool that uses past performance data to build human-
readable models for real-time fault detection.
SensorMiner “learns” from the past performance
data, develops a “signature” for nominal operation,
and automatically generates a fault-monitoring
system. Conventional technology has made this process
both a labor-intensive and error-prone task that
is often misunderstood, even by experts, according
to Interface & Control Systems. SensorMiner,
on the contrary, will automatically correlate seemingly
unrelated data and provide rich graphical feedback
to the user.
|
Interface & Control
Systems, Inc.’s SensorMiner is a time-series
data-mining tool that uses past performance
data to build human-readable models for real
time fault detection. |
As part of a complete toolset, SensorMiner uses
a combination of time warp, cluster, rule induction,
Euclidean error, and state machine technologies.
The result is a sophisticated temporal machine
learner for real-time anomaly detection. Using
time-series anomaly detection, the toolset trains
a model on known good data, estimates the probability
distribution, and assigns a likelihood-based score
to new sensor data. The score is used to determine
if real-time data have deviated from the signature
for these data points. Errors, or “out of family”
characteristics, are detected by the SCL expert
system and corrective measures can be automatically
invoked.
“The ability to automatically make predictions
or help people make decisions faster and more accurately,
in real time, has far-reaching implications that
spread across industry boundaries,” said Brian
Buckley, vice president of marketing for Interface & Control
Systems. “We designed SensorMiner to identify anomalies
in a wide variety of systems.”
SensorMiner’s toolset also consists of data-mining
tools and a real-time monitoring system. The data-mining
tools interact with a graphical front end that
allows a user to display data signatures discovered
by the data-mining strategies. Once a signature
is captured, additional data sets can be played
through a built-in simulator to exercise the monitoring
system and reveal graphical results. Using SensorMiner,
data-mining algorithms can ultimately be fine-tuned
to increase the likelihood of detecting anomalies
in the real-time system.
Interface & Control Systems software engineer
Walter Schiefele has contributed nearly 2 years
of research and development to the SensorMiner
product. “Many artificial intelligence programs
yield models that are not understandable by humans,”
according to Schiefele. “Our approach was to generate
models that are easily readable and graphically
displayed. People can visualize the rules. More
importantly, domain experts can manually trim or
adjust rules in the final product.”
The company is introducing the commercially available
data-mining tool to existing control system and
test set markets that require advanced, real-time
anomaly detection and analysis systems. SensorMiner
will then be introduced to supervisory control
and data acquisition and process-control system
fields. The company is also inviting inquiries
for partnerships and value-added resellers, for
targeted commercial markets.
|