A Self-Supervised Task for Fault Detection in Satellite Multivariate Time Series

TL;DR

The paper tackles fault detection in satellite multivariate time series under limited labeled data by introducing a permutation-based self-supervised task integrated into a Physics-Informed Real NVP framework. By permuting sensor channels and training the network to predict the permutation index, the approach leverages unlabeled data to learn meaningful representations, with configurations including multi-task, pre-training, and standalone self-supervision. The study shows that the self-supervised objective improves fault-detection performance across settings, with standalone self-supervision on the full dataset delivering the best results, highlighting data efficiency for space-domain datasets. This work advances data-efficient fault detection for spacecraft and points to broader applicability in other domains with scarce labels.

Abstract

In the space sector, due to environmental conditions and restricted accessibility, robust fault detection methods are imperative for ensuring mission success and safeguarding valuable assets. This work proposes a novel approach leveraging Physics-Informed Real NVP neural networks, renowned for their ability to model complex and high-dimensional distributions, augmented with a self-supervised task based on sensors' data permutation. It focuses on enhancing fault detection within the satellite multivariate time series. The experiments involve various configurations, including pre-training with self-supervision, multi-task learning, and standalone self-supervised training. Results indicate significant performance improvements across all settings. In particular, employing only the self-supervised loss yields the best overall results, suggesting its efficacy in guiding the network to extract relevant features for fault detection. This study presents a promising direction for improving fault detection in space systems and warrants further exploration in other datasets and applications.

Figures (2)

• Figure 1: Self-supervised loss: the dataset's columns are permuted and the Real NVP model is trained to predict the correct permutation. As shown in the image patch extracted from the ADAPT circuit, each column is associated with a given sensor of the testbed.
• Figure 2: $R$, $F$, and $G$ represent the model, a fully connected and a Gaussian distribution layer. $p$ is the permutation applied to the input and used in the self-supervised loss, ($L_{\text{self\_sup}}$). For the self-supervised task both nominal, $x_{p}^{nom}$, and fault data, $x_{p}^f$, were used, while for the main loss only nominal data, $x^{nom}$, was given as input. The main loss is composed by the default loss for Real NVP, ($-L_{\text{log\_prob}}$), and a physics-informed loss specific to ADAPT, ($L_{\text{phys\_inf}}$).