Technique for novelty detection of status & sensor data
A Belgian university research group frequently collaborating with ESA/NASA with a large body of knowledge in data processing algorithms, classification algorithms has developed advanced data mining techniques for automatic spacecraft status characterisation.
The development of novelty detection tools is a result of extensive research in data mining techniques that are useful for the health monitoring of spacecraft.
The underlying principle is the identification of anomalies that are defined as telemetry behaviour that is so different from its historical behaviour that it raises suspicion that it is caused by a different mechanism.
The developed algorithms can be used in environments where a large number of housekeeping parameters are monitored and while providing a wealth of diagnostic information, these vast numbers exceed the possibilities of the human brain to oversee all telemetry measures. Automatic checks and computer-aided data analysis are needed to help operators and operation engineers keep the overview.
The tools were developed for novelty detection in large sets of telemetry parameters and derived parameters to be used in environments where automatic checks and computer-aided data analysis are needed to help operators and operation engineers keep an overview.
These sets of data mining techniques have the possibility to assist in
- Detecting anomalies/sudden change in behaviour
- Detecting trends or change in trends
- Finding correlations based on Command and Effect Analysis
Innovations & Advantages
The main advantage of the complete set of techniques is the ability to quickly adapt towards/ implement to a different environment or industry. The set of techniques gives meaningful results at relatively low computational cost and little effort in learning and tuning the algorithms in comparison with competing techniques such as neural networks.
The technology consists of 3 tracks of techniques which enables a broad range of applications possibilities:
1. Kernel Density Estimate (KDE) techniques strengths:
- Gives an Accurate and robust description of the data distribution
- Best used for Outlier detection / novelty in single data values
- Detecting change in data distribution / epochs in data distribution
- Dynamic definition of soft limits
2. Poincaré based technique strengths:
- Complementary potential for novelty detection
- Particularly suited for enumerated / status data
- But also for continuous data
- Detection of new state transitions
3. Command effect analysis provides:
- Powerful tool to assist in the monitoring of correct command execution with computationally efficient algorithms
- Unsupervised characterisation of ‘command footprint’
Will be provided by the tech provider upon request.
Current and Potential Domains of Application
Potential applications are to be sought in the health monitoring of complex systems, e.g. automated cockpits (aviation), industrial machinery, Windmills.