基于改进LSTM的组合式飞轮故障诊断方法

Combined Flywheel Fault Diagnosis Method Based on Improved LSTM and Fault Tree

考虑难以获取飞轮精确数学模型及星上算力受限问题,提出了一种基于改进LSTM与故障树相结合的故障诊断方法。首先,从种群初始化、距离控制参数及α狼位置更新等角度改进传统灰狼算法(GWO),使其拥有更好的收敛性能;然后在网络训练过程中引入优化算法,对超参数空间开展寻优,克服传统手动调整方法或网格搜索法导致的超参数选取效率低的问题;进一步考虑故障树分析的工程实用性和神经网络的自主性,设计二者组合的故障诊断框架;最后,建立飞轮故障树模型并进行仿真实验,仿真证明了改进GWO出色的收敛性以及组合式诊断算法对飞轮故障检测和识别的有效性。

Under consideration of the difficulty in obtaining an accurate model of the flywheel and the limitation of computing power,a combined fault diagnosis method based on improved LSTM and fault tree is proposed.Firstly,the traditional grey wolf optimizer algorithm(GWO)is improved by population initialization,distance control parameters andαwolf position updates to achieve better convergence performance.Then,during the network training process,the improved GWO is introduced to optimize the hyper-parameter space,so the low efficiency of hyper-parameter selection caused by traditional manual adjustment method or grid search method is overcome;Further,due to considering the engineering practicality of fault tree analysis and the autonomy of neural network,a fault diagnosis framework that combines with those two ways is designed;Finally,a flywheel fault tree model is established and simulation experiments are conducted,which demonstrate the excellent convergence of the improved GWO and the effectiveness of the combined diagnosis algorithm for flywheel fault detection and recognition.