Data Mining in Space
from Gerhard Holtkamp, 13. March 2010, 00:29
Special statistical techniques in the so called field of data mining which are designed to find patterns in large volumes of data can be applied to detect anomalies in space missions...
Although statistical methods for analyzing multiple parameters simultaneously have been around for much longer than the Internet the large amount of data being collected nowadays and stored electronically has given rise to something called data mining. Rather than picking out single data items data mining tries to look at the overall data set available. This can lead to unexpected patterns and can help to better characterize systems, persons or whatever else you try to investigate.
Every active spacecraft sends so called housekeeping telemetry about the status of its various systems to the ground in addition to its actual payload data. These data are usually archived for at least the duration of the overall mission so that in case of anomalies possible causes can be investigated.
Mission controllers typically follow certain housekeeping parameters to check the health of systems and sub-systems and to ensure that those systems are operating properly. For instance a rise in temperature or an increase in power consumption of a particular system might indicate a potential problem which needs to be addressed.
But sometimes problems can manifest themselves in more subtle ways affecting parameters which one would not normally associate with the respective system. It is here that a data mining technique called Distance-Based Anomaly Detection can lend a helping hand.
The idea behind this technique is that you combine N different housekeeping parameters to form vectors which can be treated as points in an N-dimensional vector space. The Euclidean distance metric (or some other suitable metric) can then be employed to calculate the distance between different points. (If these technical terms have confused you - this is similar to how we measure distances in our everyday three-dimensional world except that instead of X, Y, Z space coordinates we use something like temperature, pressure or power.)
But before you can do that you have to normalize the different parameter values in some way because the various parameters are measured in different units resulting in quite different numbers. Just think of temperature, pressure and power as an example. One method of normalization is to use the mean and standard deviation calculated for each parameter then subtracting the mean from the current value and dividing the difference by the standard deviation. In addition to that you could apply some weight to each parameter due to its significance.
When everything operates as planned the points (in this N-dimensional space) tend to cluster in certain regions. A point outside these regions would indicate that at least some of the parameters have values outside their usual range. This could be due to some planned event like the firing of an engine to change the orbit of the satellite but it could also be a sign of trouble which needs investigation.
In 2002 a Control Moment Gyroscope (CMG)
of the International Space Station (ISS) failed. Such gyroscopes can be found in many satellites for attitude control. If a CMG fails unexpectedly before it can be shut down in an orderly way the ISS will drift out of its nominal attitude. Although the attitude can be restored by firing of thrusters and the spinning up of a spare CMG the event can severely disrupt ISS operations for at least a day. (Some news media tend to mix up "attitude" with "altitude" and invariably will report that the ISS has been thrown out of its regular orbit!)
Other CMGs of the ISS have also failed since and NASA decided that the above mentioned data mining technique could be promising for detecting early symptoms of CMG degradation. 13 parameters were selected for monitoring including CMG vibration, temperatures, rotation speed, electrical current and ISS rotation rates. 10 months of archived data were used to "train" the system. Then as a test the known failure in 2002 was checked.
As it turned out the data mining monitoring system showed the beginning of an anomalous behaviour more than 14 hours in advance of the actual CMG failure.
More subsystems of the ISS are now regularly checked via these data mining tools. Other satellites stand to benifit as well particularly constellations like the system of navigation satellites where you have a large number of identical satellites deployed which need to be operated in a highly reliable way.
In the early days of spaceflight nearly everything needed in space was taylor-made while some of these custom-designed inventions found their way to applications on Earth. But with space having become an integral part of our life these days the tide has turned and ever more COTS (Commercial Off The Shelf) applications find their way into space as the above example shows.
Although statistical methods for analyzing multiple parameters simultaneously have been around for much longer than the Internet the large amount of data being collected nowadays and stored electronically has given rise to something called data mining. Rather than picking out single data items data mining tries to look at the overall data set available. This can lead to unexpected patterns and can help to better characterize systems, persons or whatever else you try to investigate.
Every active spacecraft sends so called housekeeping telemetry about the status of its various systems to the ground in addition to its actual payload data. These data are usually archived for at least the duration of the overall mission so that in case of anomalies possible causes can be investigated.
Mission controllers typically follow certain housekeeping parameters to check the health of systems and sub-systems and to ensure that those systems are operating properly. For instance a rise in temperature or an increase in power consumption of a particular system might indicate a potential problem which needs to be addressed.
But sometimes problems can manifest themselves in more subtle ways affecting parameters which one would not normally associate with the respective system. It is here that a data mining technique called Distance-Based Anomaly Detection can lend a helping hand.
The idea behind this technique is that you combine N different housekeeping parameters to form vectors which can be treated as points in an N-dimensional vector space. The Euclidean distance metric (or some other suitable metric) can then be employed to calculate the distance between different points. (If these technical terms have confused you - this is similar to how we measure distances in our everyday three-dimensional world except that instead of X, Y, Z space coordinates we use something like temperature, pressure or power.)
But before you can do that you have to normalize the different parameter values in some way because the various parameters are measured in different units resulting in quite different numbers. Just think of temperature, pressure and power as an example. One method of normalization is to use the mean and standard deviation calculated for each parameter then subtracting the mean from the current value and dividing the difference by the standard deviation. In addition to that you could apply some weight to each parameter due to its significance.
When everything operates as planned the points (in this N-dimensional space) tend to cluster in certain regions. A point outside these regions would indicate that at least some of the parameters have values outside their usual range. This could be due to some planned event like the firing of an engine to change the orbit of the satellite but it could also be a sign of trouble which needs investigation.
In 2002 a Control Moment Gyroscope (CMG)
of the International Space Station (ISS) failed. Such gyroscopes can be found in many satellites for attitude control. If a CMG fails unexpectedly before it can be shut down in an orderly way the ISS will drift out of its nominal attitude. Although the attitude can be restored by firing of thrusters and the spinning up of a spare CMG the event can severely disrupt ISS operations for at least a day. (Some news media tend to mix up "attitude" with "altitude" and invariably will report that the ISS has been thrown out of its regular orbit!)Other CMGs of the ISS have also failed since and NASA decided that the above mentioned data mining technique could be promising for detecting early symptoms of CMG degradation. 13 parameters were selected for monitoring including CMG vibration, temperatures, rotation speed, electrical current and ISS rotation rates. 10 months of archived data were used to "train" the system. Then as a test the known failure in 2002 was checked.
As it turned out the data mining monitoring system showed the beginning of an anomalous behaviour more than 14 hours in advance of the actual CMG failure.
More subsystems of the ISS are now regularly checked via these data mining tools. Other satellites stand to benifit as well particularly constellations like the system of navigation satellites where you have a large number of identical satellites deployed which need to be operated in a highly reliable way.
In the early days of spaceflight nearly everything needed in space was taylor-made while some of these custom-designed inventions found their way to applications on Earth. But with space having become an integral part of our life these days the tide has turned and ever more COTS (Commercial Off The Shelf) applications find their way into space as the above example shows.
