Small-Satellite System Fault Diagnosis via a Temporal-Spatial 3D-CNN with Imbalanced-Aware Training.

Area of Science:

• Spacecraft engineering
• Artificial intelligence
• Data science

Background:

• Autonomous satellite operations require robust onboard fault detection and diagnosis (FDD).
• Satellite telemetry data presents challenges due to high dimensionality, time correlation, and severe class imbalance, hindering the recognition of rare critical faults.

Purpose of the Study:

• To develop an imbalance-aware spatiotemporal diagnostic framework for reliable FDD in autonomous small-satellite constellations.
• To enhance the detection of rare fault modes within complex satellite telemetry streams.

Main Methods:

• Conversion of multivariate telemetry into structured spatiotemporal volumes using sliding-window segmentation and grid-based embedding.
• Development of a lightweight residual three-dimensional convolutional neural network (3D-CNN) for end-to-end multi-class classification.
• Introduction of a class-balanced focal objective function to address data imbalance and improve sensitivity to minority fault types.

Main Results:

• The proposed 3D-CNN framework achieved the highest overall accuracy and Macro-F1 score compared to benchmark models (LSTM-RF, SVM, 2D-CNN, CNN-LSTM, ResNet).
• The algorithm demonstrated superior Recall for low-frequency, critical faults.
• Computational complexity analysis confirmed the algorithm's potential for real-time satellite health monitoring.

Conclusions:

• The imbalance-aware spatiotemporal diagnostic framework effectively addresses the challenges of FDD in autonomous satellite constellations.
• The proposed 3D-CNN approach offers a promising solution for enhancing the reliability and safety of satellite operations through improved fault detection.