ML for Satellite Anomaly Detection

Acceleration data showing three anomalies. Orange: Real values, Green: Predicted values, Blue: Expected limits
Acceleration data without anomalies. Orange: Real values, Green: Predicted values, Blue: Expected limits
Periodic (Temperature) data with anomalies. Note that unusual behavior is still within the normal range of values.
Periodic (Temperature) data without anomalies. These results are similar to current research.

Problem

Spacecraft cannot be manually inspected, so instead humans must look at graphs of sensor output to identify problems. Humans cost money and miss details.

What it does

Given example data of nominal spacecraft behavior, abnormalities are detected in real time and users are alerted.

How it works

Simlualted telemetry is developed for both periodic and mode-dependent variables.
Predictions are framed as a time-series forecasting model, with regression on a sliding window using a multilayer perceptron neural network on one set of training data.
Bounds of uncertainty are developed using a random forest regressor on a second set of training data.
Simulated real time data, with anomalies, is displayed in the COSMOS command and telemetry system
After training, the system produces "blue limit" bounds in real time for incoming data, and alerts the user when an anomaly deviates from these bounds.

Biggest challenges

Dynamically determining levels of uncertainty for data with variable noise
Using interprocess communication to integrate with COSMOS

Results

Unlike other research in this area, this project:

Handles multiple modes of operation, with different levels of uncertainty. For example, acceleration should be zero if the spacecraft is not maneuvering, but will be highly variable during maneuvers
Is capable of analyzing many different variables at the same time
Automatically detects anomalies and produces real-time alerts based on live data

This creates a "plug and play" system that can be used by operators without machine learning experience.