Physics-Informed Real NVP for Satellite Power System Fault Detection

TL;DR

This work tackles satellite power-system fault detection under space-environment constraints by integrating physics with a generative normalizing-flow model. The authors introduce a physics-informed Real NVP (PI-Real NVP) trained in a semi-supervised fashion, with a learned physics loss that enforces circuit-level relationships, and demonstrate superior fault detection performance on NASA's ADAPT EPS dataset compared with GRU and Autoencoder baselines. Key findings include lower false-alarm rates (FPR95) and competitive or higher F1-scores, aided by the Lagrangian-dual optimization of the physics-informed weight and interpretable feature maps. The approach offers a practical, robust, and energy-efficient fault-detection solution for space missions, with plans for ablations and cross-dataset validation.

Abstract

The unique challenges posed by the space environment, characterized by extreme conditions and limited accessibility, raise the need for robust and reliable techniques to identify and prevent satellite faults. Fault detection methods in the space sector are required to ensure mission success and to protect valuable assets. In this context, this paper proposes an Artificial Intelligence (AI) based fault detection methodology and evaluates its performance on ADAPT (Advanced Diagnostics and Prognostics Testbed), an Electrical Power System (EPS) dataset, crafted in laboratory by NASA. Our study focuses on the application of a physics-informed (PI) real-valued non-volume preserving (Real NVP) model for fault detection in space systems. The efficacy of this method is systematically compared against other AI approaches such as Gated Recurrent Unit (GRU) and Autoencoder-based techniques. Results show that our physics-informed approach outperforms existing methods of fault detection, demonstrating its suitability for addressing the unique challenges of satellite EPS sub-system faults. Furthermore, we unveil the competitive advantage of physics-informed loss in AI models to address specific space needs, namely robustness, reliability, and power constraints, crucial for space exploration and satellite missions.

Figures (6)

• Figure 1: Training step for the proposed fault detection pipeline with Real NVP. Nominal data is fed into the model and a loss composed of two terms, $-L_{log\_prob}$ and $L_{phys\_inf}$, is evaluated. The former, which is the standard loss used with these models, is computed starting from the output of the forward pass. The latter is a physics-informed loss that boosts the performance and interpretability of the model, generated by the inverse propagation of Real NVP.
• Figure 2: Computational graph for propagation in both directions in a Coupling Layer realnvp_paper. In this kind of model, s is related to the scaling function while t is related to the translation one. These functions assume the shape of neural networks with several layers that must be designed properly as a function of the application.
• Figure 3: Circuit of the testbed used to generate the ADAPT dataset. This circuit is used to handle the constraints related to the physics-informed loss composition.
• Figure 4: Training method for GRU and Autoencoders. The models are tested with standard and physics-informed loss.
• Figure 5: Physics-informed loss composition (6).