Data Mining at the Service of Space Operations

1
Data Mining at the Service of Space Operations

• J. A. Martínez-Heras, Black Hat, Spain
• A. Donati, ESA/ESOC, Germany

2
Agenda

• Innovative Operations Concepts Unit
• Data Mining
• Data Mining Operational Prototypes
• Leakage of propellant through thrusters valve
  seals
• Gyroscope average laser intensity variation
• Conclusions
• Future Work

3
Mission Control Technologies Unit

• Investigate of feasible innovative operation
  concepts
• Promote the application of new technologies for
  ESOCs core business.
• Prime Objectives
• Enhancement of overall system performances (meet
  increasing demands)
• Reduction of cost
• Mitigation of risk

4
Data Mining Process
5
Data Mining Process Examples

• Selection of data from a bigger dataset
• Pre-processing deal with null or empty values,
  harmonize time format
• Transformation re-sampling, removal of noisy
  data, feature extraction
• Data Mining Algorithms clustering, association
  rules
• Interpretation of results and Evaluation of the
  performance of the model

6
Leakage of propellant

• CLUSTER ESA Mission
• Problem Case leakage of propellant through the
  thrusters valve seals (discovered by anomalous
  thermal indications)

7
Leakage of propellant

• Objectives
• Characterize the signature of maneuvers
  generating propellant leakage
• Build a model able to recognize these maneuvers
  automatically
• Data (after sensibility analysis)
• Type of anomaly (orbit, attitude, spin-rate)
• Spacecraft orientation with respect to the Sun
• Spacecraft acceleration
• Fuel tank pressures and temperatures
• Reaction Control Subsystem temperatures

8
Leakage of propellant

• Solving Approach
• Preprocessing
• Clean data remove file headers, deal with
  missing data, timestamp homogenization
• Re-sampling
• Data Transformation
• Feature extraction max, min, mean, stdev, 5
  first coefficients of Fourier power spectrum
• Advantages dimensionality reduction and
  meaningful representation

9
Leakage of propellant

• Data Mining
• Clustering

Clustering
10
Leakage of propellant

• Data Mining
• Association Rules
• PatternA,1? PatternB,2(support, confidence)
• Meaning if parameter A is in cluster 1, then
  parameter B is likely to be in cluster 2
• Support percentage of transactions containing
  A,1 and B,2
• Confidence percentage of transactions containing
  B,2 among the transactions containing A,1
• Focus on rules where B,2 means anomaly

11
Leakage of propellant

• Evaluation Results
• Tests with blind data (new maneuvers)
• 100 of the cases verify the necessity rules
• 65 of the cases verify the sufficiency rules
• 35 could not be classified neither as nominal
  nor as failure maneuvers.
• Additional training data is required to build a
  more accurate model
• This limitation is due limited occurrences of
  anomalies

12
Leakage of propellant

• Conclusions
• It can be connected to telemetry to
• Automatically detect faulty maneuvers
• Verify that a maneuver was nominal
• It can be used to understand the complex behavior
  of the spacecraft.
• This is useful to avoid risky situations

13
Gyro Laser Intensity Variation

• ROSETTA ESA Mission
• Problem Case model the Laser Intensity of the
  Inertial Measurement Packages (IMP) with respect
  to temperature.

14
Gyro Laser Intensity Variation

• Objectives
• Find rules to explain the interaction between the
  temperatures and the laser intensities of the
  IMPs
• DAMTAM data mining tool applied to mission
  operations
• Automatically detects anomalies in telemetry
• Find possible causes for the detected anomalies
• Data
• IMPs Laser Intensities
• IMPs on / off status
• IMPs temperatures
• Sun Azimuth
• Sun Elevation

15
Gyro Laser Intensity Variation

• Solving Approach
• Preprocessing
• Get data from MUST
• Re-sampling
• Data Transformation
• Sliding window technique
• Feature extraction
• By range max, min, mean, stdev
• By variation slope, (max slope), (min
  slope), (stdev slope)

16
Gyro Laser Intensity Variation

• Data Mining
• Clustering by range and variation for every
  time-window.
• Temporal Association Rules
• PatternA,1?(T) ? PatternB,2(support, confidence)
• Meaning if parameter A is in cluster 1, then
  parameter B is likely to be in cluster 2 within T
  transactions
• Useful when the consequence is visible only after
  some time.
• When T 0, it is the normal association rules
  technique

17
Gyro Laser Intensity Variation

• Evaluation Results
• DAMTAM can correctly identify the steepest
  variations.
• Most correlations suspected by the Rosetta Flight
  Control Team were found
• Among temperatures of the IMPs
• Between Sun attitude and IMPs temperatures
• IMPs temperature, on / off status and laser
  intensities
• IMPs temperature, on / off status and laser
  intensities of other IMPs
• Some desired correlations could not yet be found
• Current DAMTAM is good at detecting strong
  variations

18
Gyro Laser Intensity Variation

• Conclusions
• DAMTAM is the first step into developing a
  generic system that allows the detection of
  correlations among parameters
• DAMTAM saves time effort to the FCT by going
  through a huge amount of data and suggesting
  correlations
• Future work improve the technique to also detect
  smaller variations

19
Conclusions

• Data Mining is ready for Space Operations
• Anomaly characterization
• Behavioural modelling
• Benefits
• Increased understanding of the spacecraft
  behaviour
• Automatic anomaly detection and identification
• Capability of doing what-if analysis
• Degradation detection and anomaly anticipation
• Get prepared ahead of time

20
Future Work

• Connection to telemetry flow
• Verification in near real time
• Alarming on forecasted anomaly
• DAMTAM evolution
• Improved clustering technique
• It will allow the detection of small and big
  variations

21
ContactFor more Information

• Jose.Antonio.Martinez.Heras_at_esa.int
• Alessandro.Donati_at_esa.int
• European Space Operations Centre
• Darmstadt, Germany