基于长短时神经网络的卫星陀螺仪故障检测

Fault Detection of Satellite Gyroscope by Using LSTM Network

针对卫星陀螺仪故障检测中存在的冗余依赖、微小故障覆盖问题,提出一种基于长短时神经网络(LSTM)的故障检测方法。首先对卫星陀螺仪建模,考虑到卫星姿态控制回路对陀螺仪微小故障覆盖影响,利用半物理仿真平台采集陀螺仪正常与故障数据;然后使用部分正常数据训练LSTM神经网络,使得网络具有预测陀螺仪输出的能力,并将另一部分正常数据输入到训练好的网络模型,得到预测误差,进一步设定故障阈值;最后,将测试数据输入提出的故障检测模型,仿真验证其时效性和准确性。结果表明,在采样频率为10Hz时,对于陀螺仪的卡死、噪声以及偏差故障,基于LSTM神经网络的故障检测模型能在故障发生2s内检测出故障,并达到了98.9%的准确率。

Aiming at the problems of redundancy dependence and small fault coverage in satellite gyroscope fault detection, a fault detection method based on long short-term neural network(LSTM) is proposed. Firstly, the satellite gyroscope is modelled, and the gyroscope normal and fault data are collected by using a semi-physical simulation platform, under consideration of the influence of the satellite attitude control loop on the gyroscope fault coverage. Then, the LSTM neural network is trained by using part of the normal data to make the network capable of predicting the gyroscope output, and another part of the normal data is fed into the trained network model to obtain the prediction error and set the fault threshold. Finally, the test data is input into the proposed fault detection model and its timeliness and accuracy are simulated and verified. The simulation results show that the fault detection model based on LSTM neural network can detect faults within 2s of fault occurrence and achieve 98.9% accuracy for jamming, regarding noise and deviation faults of the gyroscope at a sampling frequency of 10Hz.