Applying Data Mining for Detecting Anomalies in Satellites

Telemetry data is the only source for identifying/predicting anomalies in artificial satellites. Human specialists analyze these data in real time, but its large volume, makes this analysis extremely difficult. In this experience paper we study the hypothesis of using clustering algorithms to help operators and analysts to perform telemetry analysis. Two real cases of satellite anomalies in Brazilian space missions are considered, allowing assessing and comparing the effectiveness of two clustering algorithms (K-means and Expectation Maximization), which showed to be effective in the case study where several telemetry channels tended to deliver outlier values and, in these cases, could support the satellite operators by allowing the anticipation of anomalies. However for silent problems, where there was just a small variation in a single telemetry, the algorithms were not as efficient.