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

Reliable onboard fault detection and diagnosis (FDD) is essential for autonomous small-satellite constellation operations. The satellite telemetry streams are typically high-dimensional, strongly time-correlated, and severely imbalanced. These characteristics make rare but critical faults hard to recognize. To address these issues, this paper proposes an imbalance-aware spatiotemporal diagnostic framework based on three-dimensional convolutional neural networks (3D-CNNs). Multivariate telemetry is first converted into structured spatiotemporal volumes via sliding-window segmentation and grid-based embedding. This enables the model to jointly learn temporal evolution and cross-parameter coupling patterns. A lightweight residual 3D-CNN is developed to enable end-to-end multi-class classification. In addition, a class-balanced focal objective function is introduced to mitigate class-imbalance issues and enhance sensitivity to minority fault modes. The Lumelite series satellite telemetry dataset, comprising 23 fault types, is constructed for training and evaluation. The proposed lightweight residual 3D-CNN is benchmarked against long short-term memory-random forest (LSTM-RF), support vector machine (SVM), 2D-CNN, CNN-LSTM, and residual neural network models. Experimental results show that the proposed algorithm has the highest overall accuracy and Macro-F1 score. It also obtains higher Recall for low-frequency faults. The computational complexity studies indicate that the proposed algorithm has promising potential for real-time satellite health monitoring.