Abstract: The classical sequence recurrent neural networks (SRNNs) have the defect of long-term dependence in the prediction of time series;meanwhile,they increase time complexity by traversing the whole training data set to realize supervised learning,which makes SRNNs show low prediction accuracy and computation efficiency in the residual life prediction of space rolling bearings.In view of this,a novel SRNN named variational eligibility trace meta-reinforcement recurrent network (VETMRRN) is proposed.In the proposed VETMRRN,a new sequence recurrent network structure is constructed to increase the memory of historical information,and a meta-learning hyperparameter self-initialization network based on oracle gate mechanism is designed to quickly seek the optimal review sequence length of VETMRRN.Moreover,a variational auto-encoding meta policy gradient learning algorithm with an eligibility trace operator is designed to improve the training speed and global optimization effect for VETMRRN parameters.On the basis of the above advantages of VETMRRN,a new residual life prediction method of space rolling bearings based on VETMRRN is proposed as follows.Firstly,the performance degradation features of space rolling bearings are extracted by Shapely-value feature fusion.Then,the performance degradation features are input into VETMRRN to predict the performance degradation feature trends of space rolling bearings.Finally,a Weibull-distribution reliability model is established to predict the residual lives of space rolling bearings.The proposed VETMRRN-based residual life prediction method can obtain higher prediction accuracy,better generalization performance and higher computation efficiency.The effectiveness of the proposed method is proved by a residual life prediction example of space rolling bearings.

Key words: meta-learning sequence recurrent network, Oracle gate mechanism, meta policy gradient, space rolling bearing, residual life prediction

摘要： 针对经典时间循环神经网络（SRNNs）在预测时间序列中存在长时依赖的缺陷,同时由于遍历整个训练数据集进行现有监督式学习增加了时间复杂度,造成SRNNs在预测空间滚动轴承剩余寿命中存在预测精度和计算效率较低的问题,提出一种变分资格迹元强化循环网络（VETMRRN）。在VETMRRN中,构建新型时间循环网络结构以增大历史信息记忆量,并设计基于神谕门机制的元学习超参数自初始化网络来加速搜索VETMRRN的最优回顾序列长度;设计含有资格迹算子的变分自编码元策略梯度学习算法,以提高对VETMRRN参数的训练速度和全局优化效果。在此基础上提出基于VETMRRN的空间滚动轴承剩余寿命预测方法。首先采用Shapely值特征融合提取空间滚动轴承的性能退化特征;然后将性能退化特征输入VETMRRN中预测性能退化特征趋势;最后建立威布尔分布可靠度模型来预测空间滚动轴承的剩余寿命。该方法具有较高的预测精度、较好的泛化性能和较高的计算效率。最后通过空间滚动轴承剩余寿命预测实例证明了该方法的有效性。

关键词: 元学习时间循环网络, 神谕门机制, 元策略梯度, 空间滚动轴承, 剩余寿命预测