深度对比一类分类在轴承异常检测中的应用  

Unsupervised Anomaly Detection for Bearings Using Deep Contrastive One⁃Class Classification

作　　者：严如强 朱启翔 李亚松 周峥 YAN Ruqiang;ZHU Qixiang;LI Yasong;ZHOU Zheng(School of Mechanical Engineering,Xi'an Jiaotong University Xi'an,710049,China)

机构地区：[1]西安交通大学机械工程学院,西安710049

出　　处：《振动.测试与诊断》2025年第1期1-9,196,共10页Journal of Vibration,Measurement & Diagnosis

基　　金：国家自然科学基金资助项目(U23A20620)。

摘　　要：针对机械设备的异常数据难以获取和传统异常检测方法容易误报的问题,提出了一种基于深度对比一类分类的无监督异常检测框架,用于检测轴承等关键部件产生异常的时间点。所提出的框架分为2部分。第1部分,针对深度一类分类方法的分布未知与模型坍缩问题,提出一种改进的深度对比一类分类损失,修改了相似度度量方式,并添加了增强样本对之间的相似度约束。在训练过程中,选取4种备选的数据增强方案进行实验和分析,并选取了最佳的数据增强组合,使模型学习得到了更加均匀的正常数据分布。第2部分,采用极值理论在检测过程中不断拟合分布尾部的极值分布,动态更新异常样本阈值进而避免误报。最后,在辛辛那提轴承寿命数据集上验证了提出的异常检测框架在特征分布的均匀性、异常样本的分类准确性与故障起始点检测的精准性方面都具有优越性。Since the anomaly data of mechanical systems is difficult to obtain and the traditional anomaly detec⁃tion methods are easy to produce false alarms,an unsupervised anomaly detection framework based on deep con⁃trastive one-class classification is proposed for detecting the failure start point of bearings.The proposed frame⁃work is divided into two parts.In the first part,to solve the problem of unknown distribution of normal data and model collapse,an improved deep contrastive one-class classification loss is construct through modifying the similarity function and adding constraint term between augmented samples.In the training process,four data augmentation methods are selected to experiment and analyze,and the best combination of data enhancement is selected,results in a more uniform and compact normal data distribution.In the second part,the extreme value theory is used to dynamically calculate more accurate anomaly thresholds to avoid false alarms.Through valida⁃tion on the bearing dataset from intelligent maintenance systems of the university of Cincinnati benchmark,the proposed method achieves superiority over other state-of-the-art methods in terms of uniformity of feature distri⁃bution,classification accuracy of abnormal samples and precision of failure start point detection.

关 键 词：深度一类分类 异常检测 极值理论 对比学习 

分 类 号：TH165.3[机械工程—机械制造及自动化] TN911.72[电子电信—通信与信息系统]