Abstract: To address the problems of individual variability of bearing degradation trajectories and artificial subjectivity of degradation stage division, a self-monitoring data-driven extraction method of independent bearing degradation trajectory and an autonomous segmentation technique of degradation stages were proposed. A multiscale residual deep convolutional autoencode was developed herein to autonomously extract the bearing performance degradation features by unsupervised learning of bearings own historical monitoring data, and then combined with support vector data description model to construct single bearing independent degradation trajectory. The de-trending super-threshold waveform method was introduced to automatically detect the starting degradation point, and the failure threshold was set autonomously using logistic regression-based failure probability statistics method, thus the bearing independent degradation trajectory was adaptively segmented. Driven by degradation stage index obtained from trajectory segmentation, the accurate prediction of bearing life was achieved by combining full time power gray prediction model. The experimental results show that the multiscale residual deep convolutional autoencode network proposed herein may construct a degradation trajectory reflecting the degradation law of the bearing itself according to the respective working conditions of the bearings, and the adaptive degradation trajectory segmentation method proposed herein may detect the starting degradation point and the failure threshold of the bearing without references. The results may improve the scientific objectivity of bearing degradation assessment and the engineering operability of life prediction.

Key words: self-driven degradation trajectory, multi-scale residual convolution, failure threshold, gray prediction

摘要： 针对轴承退化轨迹的个体差异性及退化阶段划分的人为主观性问题，提出了自监测数据驱动的轴承独立退化轨迹的提取与退化阶段的自主分割方法。该方法采用多尺度残差深度卷积自编码器对轴承自身历史监测数据进行无监督学习，实现了轴承性能退化特征的自主提取，并结合支持向量数据描述模型构建单个轴承的独立退化轨迹。引入基于去趋势化超阈波峰法的退化起始点自动检测方法与基于逻辑回归失效概率的失效阈值自主设定方法，对轴承独立退化轨迹进行了自适应分割。以轨迹分割得到的退化阶段数据为驱动，结合全阶时间幂灰色预测模型实现了轴承寿命的准确预测。试验结果表明，提出的多尺度残差深度卷积自编码器能够根据轴承各自工况环境构建反映轴承自身服役性能变化规律的退化轨迹，提出的退化轨迹自适应分割方法能够无参考地检测出轴承的起始退化点与失效阈值，从而提高轴承退化评估的科学客观性与寿命预测的工程可操作性。

关键词: 自驱式退化轨迹, 多尺度残差卷积, 失效阈值, 灰色预测