一种改进滤波器与LSTM融合的飞轮故障预测算法

Flywheel Fault Prediction Algorithm Based on Improved Filter and LSTM Fusion

针对基于滤波器的预测方法由于依赖模型精度而导致长周期预测精度低的问题,提出了一种基于滤波器和长短期记忆(LSTM)网络的融合故障预测算法,实现航天器飞轮执行器缓变故障的预测。首先,分别设计了小批量标准化LSTM网络和趋势识别模块,二者串联组成神经网络预测器提高对时序预测的准确性。然后,利用递归最小二乘(RLS)参数估计原理改进卡尔曼滤波器更新过程,以增强对时序预测误差的鲁棒性。在此基础上,将神经网络预测器输出的预测值与改进后的滤波器相融合,获得未来时刻的预测残差项实现迭代更新和预测,克服了滤波器算法对模型的依赖,提高了预测精度。最后,设计仿真实验比较了三种神经网络预测器的时序预测性能,并考虑飞轮轴承性能退化故障,利用所提融合预测算法判断出飞轮在856 s时性能退化达到阈值,预测时间误差为36 s,从而验证了预测算法对缓变故障的有效性。

To solve the problem that the poor accuracy of long-term prediction relying on model accuracy for the filter-based prediction method,a fusion fault prediction algorithm based on filter and long short-term memory(LSTM)network is proposed to achieve the prediction of spacecraft flywheel slowly growing faults.Firstly,a mini-batch normalization LSTM network and a trend recognition module are designed,which are connected in series to form a neural network predictor to improve the time series prediction accuracy.Then,the Kalman filter update process is improved by the recursive least square(RLS)parameter estimation principle to enhance the robustness for time series prediction error.On this basis,the predicted values output by the neural network predictor are fused with the improved filter.Prediction residual can be obtained for iterative updating and prediction,overcoming the dependence of the filter algorithm on the model and improving prediction accuracy.Finally,the time series predictive performance of three neural network predictors are analyzed by simulation experiments.And considering the degradation fault of flywheel bearing performance,the proposed fusion prediction algorithm is used to determine that the flywheel degrades to a threshold at 856 s,with a prediction time error of 36 s.The simulation results verify the effectiveness of the proposed algorithm for slowly growing fault.