Health Monitoring of Spacecraft
[wtm_use_case_cats]
An alternative data-driven approach has been suggested for anomaly detection by Takehisa Yairi and collaborators (Yairi et al.). The process takes into account five characteristics of telemetry data: High dimensionality, multimodality, heterogeneity, temporal dependence, missing data and trivial outliers
Benefits for the company
In a spacecraft, there can be 1000+ monitored telemetry variables, that on-earth engineers analyse to diagnose abnormalities. AI helps in saving time and resource, deliviring output with great precision.
Feasability
Medium
Type of expertise/ AI domain
Principal Component Analysis, Support Vector Machine, AI
Internal data required
Telemetry Data
After the telemetry variables are normalised, they are used to create an integrated model called a mixture of probabilistic principal components analysers can categorical distributions (MPPCACD). Yairi et al. tested the model on telemetry data from the SDS-4, a demonstration satellite used by JAXA. A total of 89 continuous and 365 status variables were modelled. The performance of MPPCACD was compared to state of the art anomaly detection algorithms such as one-class support vector machine (OCSVM) and support vector data description (SVDD). The three algorithms function by producing an anomaly score graph, in which anomalies are denoted by graph spikes. In the graph out of MPPCACD, it was noted that in normal systems the anomaly scores are typically low (0–10) whereas anomalies are easily identifiable with scores of 100–200. As OCSVM and SVDD are general purpose method, the anomaly scores were erratic even in normal system operation making it difficult to distinguish anomalies.