Satellite Failure Prognosis with Cascaded Temporal Convolution and Transformer Network for Multi-Scale Features

Failure prognosis provides critical decision-making support for Integrated System Health Management (ISHM), ensuring the operational safety of satellites in orbit. Temporal Convolutional Networks (TCNs), known for their capability in processing time-series data, have become an important approach for failure prognosis. The gradual performance degradation of satellites, combined with multi-physics coupling effects, gives rise to multi-scale features. However, existing TCN based failure prognosis methods remain limited in their ability to simultaneously capture both local and global features, posing challenges when processing such multi-scale features. To address this issue, a Cascaded Temporal Convolution and Transformer Network (CTCTN) framework is proposed for satellite failure prognosis and uncertainty quantification. The CTCTN first adaptively aligns the feature dimensions of the Depthwise Separable Temporal Convolution (DS-TC) block and the Transformer module through Adaptive Average Pooling (AAP), enabling both local feature extraction and global dependency modeling. A heteroscedastic Huber loss function is then designed to optimize the mean and variance of the CTCTN output. Finally, epistemic and aleatoric uncertainties are separately estimated and used to construct probabilistic prediction intervals. A satellite model is developed, and a run-to-failure dataset is constructed to validate the proposed CTCTN framework using a performance degradation scenario caused by damage to the Solar Array Paddle (SAP) of a Low Earth Orbit (LEO) satellite. Experimental results demonstrate that the proposed CTCTN method not only achieves more accurate Remaining Useful Life (RUL) predictions but also effectively quantifies uncertainty arising from multi-scale features. This work provides a reference case for failure prognosis in LEO satellites and offers decision support for ISHM.