智能化滚珠丝杠副退化状态的评估方法  

Intelligent Evaluation Method for Ball Screw Degradation State

作　　者：张江泉 高宏力[1] 向守兵 郭亮[1,2] 谭咏文[1] ZHANG Jiangquan;GAO Hongli;XIANG Shoubing;GUO Liang;TAN Yongwen(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China;Laboratory of Science and Technology on Integrated Logistics Support,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031 [2]国防科技大学装备综合保障技术重点实验室,湖南长沙410073

出　　处：《西南交通大学学报》2022年第4期813-820,共8页Journal of Southwest Jiaotong University

基　　金：国家自然科学基金(51775452)。

摘　　要：现有滚珠丝杠副退化状态评估方法通常假设已有充足且带标签的数据集,但实际工程应用中故障成本过高、获取标签难度过大,难以在特定工况下获得大量带标签数据集.针对上述问题,提出一种基于多尺度对抗域对抗学习的智能化状态评估方法,结合注意力卷积神经网络模块和域对抗学习模块,利用不同工况下采集的传感器信号建立深度学习模型,从而自适应地学习域不变特征并实现高效的知识复用和特征迁移;利用多工况下采集的滚珠丝杠副退化信号构建试验数据集来验证方法的有效性.研究结果表明:本文方法在6个标签缺失跨工况条件下的滚珠丝杠副退化状态识别子任务中均取得了高于89.02%的识别准确率;能够充分迁移带标签数据的关键特征,实现了标签样本缺失条件下目标工况退化状态识别.The existing ball screw degradation assessment method usually assumes that sufficient labeled data sets are available.However,it is difficult to obtain massive labeled data sets under practical projects due to excess failure cost and difficulty of obtaining labels.To solve the above problems,an intelligent state evaluation method based on multi-scale adversarial domain adversarial learning is proposed.Combining an attention convolution neural network module and a domain adversarial learning module,a deep learning model is established by using sensor signals collected under different working conditions,so as to learn domain invariant features adaptively and realize efficient knowledge reuse and feature migration.The experimental data sets are constructed by using the ball screw degradation signals collected under multiple working conditions to verify the effectiveness of the method.The results show that the proposed method achieves a recognition accuracy higher than 89.02%in six sub-tasks of degradation state identification of ball screw under cross-working conditions with missing labels.The proposed method can fully migrate key features with labeled data and achieve the degradation state identification of target operating conditions under missing label samples.

关 键 词：滚珠丝杠副 退化状态评估 深度学习 域对抗学习 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置] TH117.1[自动化与计算机技术—控制科学与工程]