Multi-Label Fault Detection in Small Satellite ADCS Using Hybrid Random Forest and LSTM Algorithm

The Satellite Technology and Research Centre (STAR) has built 3 small satellites for formation flying mission, namely Lumelite-1, Lumelite-2 and Lumelite-3. The satellites are flying in a closed-pendulum formation. One challenge is the simultaneous and continuous monitoring of all satellites’ health. Any anomalies must be immediately detected. Therefore, this paper proposes an integrated framework that combines Long Short-Term Memory (LSTM) neural networks with Random Forest (RF) classifiers to enable advanced fault prediction and multi-label fault classification in small satellite systems, particularly the attitude determination and control subsystem. The LSTM network captures temporal dependencies in sequential telemetry data to forecast potential faults, while the RF classifier identifies specific fault types through multi-label classification. We validate the approach using satellite telemetry datasets and evaluate performance using metrics such as Accuracy, precision, recall, and F1 Score. Experimental results confirm the robustness, accuracy, and reliability of the proposed method in fault forecasting and diagnosis tasks.