Abstract
Aiming at the urgent need for autonomous fault diagnosis in on-orbit intelligent perception and understanding of space power devices such as liquid rocket engine turbopumps, the problem that the early weak fault of its key components is difficult to perceive in strong noise environment is solved. This paper proposes a physical-data hybrid method based on a Lock-in Amplification Encoding Network (LIAE-Net). This method injects physical priors into the modeling of fault frequency uncertainty interval, guides adaptive Gaussian filtering to suppress noise, and uses signal statistical volatility as a fault significance index to finally realize fault self-diagnosis. The strong noise experiment on the turbopump data set of the liquid rocket engine shows that the diagnostic accuracy of LIAE-Net reaches 82.35% ± 1.42% when the signal-to-noise ratio is as low as 0 dB, and reaches 99.42% ± 0.28% at 10 dB, which is significantly better than the comparison models such as ResNet18 and Transformer, and shows better stability. This study provides a new idea of fault diagnosis with high reliability and strong anti-interference for on-orbit intelligent operation and maintenance and autonomous health management of spacecraft and space power equipment, which has important reference value for improving the reliability of China's aerospace transportation and weapon equipment.