Self-supervised Real NVP for on-board satellite fault detection

In the space sector, the harsh environmental conditions and limited accessibility make robust fault detection systems crucial for ensuring mission success. Many existing approaches are computationally heavy, making them unsuitable for real-world applications where satellites have limited computational capabilities. We introduce a lightweight and effective normalizing flow model based on Real-valued Non-Volume Preserving (Real NVP), to accomplish the task of high robustness in fault detection with low computational power. A self-supervision technique is implemented in the framework to improve its robustness. Extensive experimentation was conducted on three publicly available datasets, benchmarking different configurations, including pre training with self-supervision, multi-task learning, and standalone self-supervised training. We show good generalization capabilities with different data, including the recently released dataset of the European Space Agency. An ablation study is performed to analyze performance changes when introducing faults during both self-supervision and main loss stages of training. Finally, systems-on-chip and artificial intelligence accelerators are used to perform latency tests, verifying our framework’s feasibility and adaptability in real satellite scenarios. The proposed self-supervised model based on Real NVP achieves better results with respect to the state-of-the-art in terms of detection performance and inference time over multiple benchmark datasets of the field, also confirming its consistency on low-power edge devices.