Information-Theoretic Channel Selection and Spatiotemporal Deep Learning for Early Fault Detection in Microsatellite...

Area of Science:

• Spacecraft engineering
• Artificial intelligence in aerospace
• Satellite thermal management

Background:

• Microsatellite thermal control systems (TCS) face challenges in early fault detection due to high-dimensional data, complex periodicities, and limited downlink.
• Existing methods struggle with scarce labeled data and fail to capture inter-sensor spatial correlations.

Purpose of the Study:

• To develop a robust framework for early fault detection in microsatellite TCS.
• To address limitations of existing data-driven approaches by integrating novel feature selection and deep learning techniques.

Main Methods:

• Utilized Generalized Maximum Information Coefficient (GMIC) for information-theoretic feature selection, reducing dimensionality by 82%.
• Employed a dual-level Seasonal Trend Decomposition (STL) to disentangle multi-scale thermal dynamics.
• Applied Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) networks for spatiotemporal modeling and prediction.
• Implemented an adaptive threshold-based weighted error fusion for anomaly detection within a single day's data.

Main Results:

• Achieved high-precision fault detection across multiple fault types using a minimal set of temperature channels.
• Demonstrated significant improvements in prediction accuracy and detection reliability compared to existing benchmarks.
• Successfully identified anomalies within a single day's telemetry data.

Conclusions:

• The proposed hybrid framework effectively overcomes challenges in microsatellite TCS fault detection.
• The integration of GMIC, STL, and CNN-LSTM offers a powerful solution for reliable and timely anomaly identification.
• This approach enhances satellite operational safety and mission success through early fault diagnosis.