LSCFNet: A Lightweight Network for BCDR Fault Diagnosis

Abstract

Reliable fault diagnosis is critical for the on-orbit operation of the satellite electrical power system (EPS). Because onboard platforms have limited computing power, storage, and energy, a fault diagnosis model must be accurate while also being small and fast. This paper proposes a lightweight neural network, LSCFNet (Lightweight Satellite Circuit Fault Network), for recognizing several types of circuit faults. The network uses a sliding window and dimensional reshaping to map one-dimensional time-series signals into two-dimensional feature matrices so as to capture cross-channel multi-variable interactions and localized temporal dynamics, and its backbone adopts depthwise separable convolution and an inverted residual structure, together with an SE (Squeeze-and-Excitation) attention module that adaptively strengthens the key fault-feature channels. Due to the scarcity of fault data, we develop a simulation model of a satellite power system with a fully regulated bus. Four typical short-circuit faults are introduced at different locations within the power distribution network, and seven key electrical variables are collected and processed using sliding-window segmentation and dimensional reshaping to construct a highly dynamic transient fault dataset. The results show that LSCFNet achieves the highest classification accuracy (99.48%) while greatly reducing computational and storage cost, achieving very high hardware efficiency in number of parameters, weight file size, computation, and inference speed. The proposed model therefore provides a low-complexity candidate architecture with high theoretical feasibility for future deployment on satellite onboard embedded platforms.