In industry, the fatigue failure of bearings will lead to unexpected shutdown of mechanical equipment. Therefore, it is necessary to predict the remaining useful life (RUL) for guiding the maintenance process. The precision of prediction depends on the construction of degradation features. However, the selection and construction of feature set is complex and changeable, and the selection of degradation index is highly subjective. In the article, the bearing RUL prediction of fatigue degradation process based on dynamic features construction is proposed to solve these two problems. In this method, the degradation process is divided into two stages by the frequency domain analysis. The improved convolution neural network (CNN) by deep mutual learning (DML) is used to extract features automatically for the former stage. The prediction results of the latter stage can be achieved through the long short-term memory (LSTM) network. The average percentage error of prediction on the two data sets is 5.05%. The experimental results show the effectiveness of the proposed method.

在工业中，轴承的疲劳失效会导致机械设备意外停机。因此，有必要预测剩余使用寿命（RUL），以指导维修过程。预测的精度取决于退化特征的构建。然而，特征集的选择和构建是复杂多变的，退化指标的选择具有很强的主观性。针对这两个问题，本文提出了基于动态特征构造的轴承疲劳退化过程RUL预测。在该方法中，劣化过程通过频域分析分为两个阶段。通过深度互学习（DML）改进的卷积神经网络（CNN）用于前阶段自动提取特征。后期的预测结果可以通过长短期记忆（LSTM）网络来实现。两个数据集的平均预测误差百分比为 5.05%。实验结果表明了所提方法的有效性。