Abstract: Aiming at the single time-frequency domain index cannot explain the life cycle degradation characteristics of harmonic drive,a remaining useful life prediction method combining Long Short Term Memory (LSTM) and Geometric Mean Optimizer Particle Filter (GMOPF) was proposed for crack extension behaviors in different parts of flexspline.The degradation characteristics of harmonic drive were difficult to be characterized by a single time-frequency domain index.On the basis of local mean decomposition based vibration signal processing,LSTM was used to obtain multiple time-frequency domain index relationships,and the mapping between signal characteristics and degradation state was realized.Then,the state equation of three degradation periods of early,middle and late stages for harmonic drive were constructed based on Paris model and Foreman model.To alleviate the problem of particle weight degradation,the double fitness index was introduced,and an iterative method of state equation parameter updating based on GMOPF was proposed.The validity of the method was verified through flexspline crack extension experiments,and the prediction accuracy was increased by 16.2% compared with the current popular method,which provided a basis for the reliability design of the harmonic drive.

Key words: harmonic drive, residual life prediction, long short term memory, particle filter

摘要： 针对单一时频域指标未能较好表征谐波减速器全寿命周期退化特性问题,并考虑柔轮裂纹在不同部位的扩展速率差异性对剩余寿命预测精度的影响,提出一种基于长短期记忆网络和几何平均优化的粒子滤波(LSTM- GMOPF)的谐波减速器剩余寿命预测方法。谐波减速器退化特性难以通过单一时频域指标表征,在基于局部均值分解的振动信号处理的基础上,利用LSTM获取多个时频域指标关系,实现信号特征与退化状态之间的映射;考虑裂纹扩展速率的差异性,基于Paris及Foreman模型构建了谐波减速器前中后3个退化模型状态方程;为缓解粒子权值退化问题,引入双适应度指标,提出了基于GMOPF的状态方程参数更新迭代方法。通过谐波减速器加速寿命试验验证了方法的有效性,对比目前流行方法预测准确性最大提高了16.2%,为谐波减速器可靠性设计提供依据。

关键词: 谐波减速器, 剩余寿命预测, 长短期记忆网络, 粒子滤波