In the space sector, the harsh environmental conditions and limited accessibility make robust fault detection systems crucial for ensuring mission success and protecting valuable assets. This work analyzed the use of Real NVP neural networks for on-board fault detection in satellites, along with a self-supervised task based on sensor data permutation. The approach is aimed at improving fault detection in satellite multivariate time series data. We conducted extensive experiments on multiple datasets involving various configurations, including pre-training with self-supervision, multi-task learning, and standalone self-supervised training. Our key contributions include an evaluation of the methods across multiple datasets, including a recently published dataset provided by the European Space Agency (ESA), achieving state-of-the-art results, and an ablation study investigating the impact of introducing faults during both the self-supervision and main loss stages of training. Finally, we performed latency tests on relevant systems-on-chip and artificial intelligence accelerators, ensuring practical deployment feasibility. Results demonstrate significant performance gains across all tested configurations, with self-supervised learning proving especially effective in feature extraction for fault detection on unlabeled samples, highlighting its potential for broader application in space systems.