基于深度学习的精密滚珠丝杠副服役性能退化过程状态评估及预测方法研究

Prognostics and health management (PHM) is an important way to improve the serviceability and maintainability of mechanical systems, which are being increasingly complex, integrated, intelligent. The development trend of high-end CNC machine tool of being high reliability and smart health management requires a high precision retaining ability of ball screw. Based on the idea of PHM, this project carries out the health monitoring, status assessment and performance prognostics of the precision degradation process of ball screw under the actual working conditions. Specifically, the vibration and temperature signals during the operation of ball screw are acquired. The empirical mode decomposition and wavelet decomposition methods are employed to extract the energy entropy, intrinsic mode function entropy, statistical features, etc. from the raw signals. The auto-associative kernel regression method is used to fuse the multiple features to reconstruct the health indicators. Three metrics are constructed to quantify the goodness of the health indictors. The optimization algorithm, multi-objective binary differential evolution is used to select the optimal health indictor, which will be used to characterize the precision degradation of ball screw. The long short-term memory neural network is proposed to assess the status of degradation process while the nonlinear autoregressive with external inputs neural network is employed to perform the prognostics of the ball screw degradation. The distribution of the remaining useful life of ball screw is predicted. The implementation of this project is expected to provide support for the smart health management of the high-end CNC machine tool.

预测及健康管理(PHM)技术是提高日趋复杂、综合化和智能化机械系统服役性和维修保障性的重要手段。本项目针对高档数控机床高可靠性及整机智能化健康保障的发展方向对精密滚珠丝杠副日益增长的高精度保持性需求，基于PHM思想开展滚珠丝杠副实际工况下精度退化过程的在线监测、状态评估及性能预测相关方法研究。采集丝杠副运行中振动和温升信号，基于经验模态分解、小波包分解等方法提取能量熵、本征模函数熵在内的多维特征，提出自相关内核回归多维重构方法对多维特征进行融合构建特征空间。设计考察特征优劣性的多维量化指标，并以上述指标为目标，基于多目标二元差分演化方法在特征空间中选择最优解，作为描述丝杠副性能退化的健康因子。设计深度学习中的长短期记忆神经网络对健康因子时序序列进行分类实现对精度退化状态的评估。设计带外部输入非线性自回归网络实现对未来精度退化水平的预测。本项目的实施有望为数控机床整机智能化健康保障提供支持。

预测及健康管理(PHM)技术是提高日趋复杂、综合化和智能化机械系统服役性和维修保障 性的重要手段。本项目针对高档数控机床高可靠性及整机智能化健康保障的发展方向对精密滚珠丝杠副日益增长的高精度保持性需求，基于PHM思想开展滚珠丝杠副实际工况下精度退化过程的在线监测、状态评估及性能预测相关方法研究。经历3年研究，本项目进行了滚珠丝杠副预紧力、润滑退化实验，连续采集了三轴加速度信号与温度信号，提取信号中包括均值、方差、小波包能量、瞬时谱熵、瞬时频率等在内的15中特征；采用斯皮尔曼等级相关系数对上述特征以单调性为评价指标进行排序，选出单调性好的若干种特征；采用主成分分析法将选出的特征进行融合构成表征退化的健康因子；再采用贝叶斯更新理论对丝杠预紧力退化趋势实现预测。最终实现以传感器信号间接表征丝杠预紧力退化的研究目标。.此外，为了进一步探索深度学习方法在其他工业设备如旋转机械关键部件轴承、齿轮状态监测与诊断方面的应用前景，申请人带领课题组研究生赴江苏无锡厚德自动化仪表有限公司开展若干组轴承、行星减速度不同故障状态下的数据采集，还收集了公开轴承故障、齿轮故障、刀具磨损、轴承全寿命数据等PHM Society论坛公开数据集和历届PHM数据挑战赛的公开数据，深入研究探索了基于深度学习的数据驱动类故障诊断与剩余寿命预测方法，一定程度上解决了变工况变设备下的迁移诊断、针对诊断模型面对新诊断任务时复用性差、需要重新设计诊断模型、依赖人工经验、手动调参、试错迭代等劳动密集工作，提出了基于强化学习与神经结构搜索的诊断模型自动化建模方法，实现诊断模型的快速设计。.在完成上述项目基础上，基于本课题研究内容与成果，基于深度学习PHM建模技术在航空航天兵器等领域科研项目上得到应用。