基于自步学习的刀具加工过程监测数据异常检测方法  

Abnormal Detection Method of Tool Machining Monitoring Data Based on Self-Paced Learning

作　　者：张建 胡小锋[1] 张亚辉[2] ZHANG Jian;HU Xiaofeng;ZHANG Yahui(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;Institute of Marine Equipment,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240 [2]上海交通大学海洋装备研究院,上海200240

出　　处：《上海交通大学学报》2023年第10期1346-1354,共9页Journal of Shanghai Jiaotong University

基　　金：国防基础科研计划项目(JCKY2021110B048);国家重点研发计划资助项目(2018YFB1700502)。

摘　　要：针对零件加工过程的监控数据异常导致刀具剩余寿命预测准确性下降的问题,提出一种基于自步学习的数据异常检测方法.首先建立多层感知机模型关联刀具加工过程监测数据和所对应的刀具剩余寿命;其次在模型权重更新过程中,先固定模型权重参数,预测损失拟合高斯分布得到异常样本的损失阈值,然后构建基于自步学习方法的损失函数,迭代更新模型参数.在模型训练结束后,根据损失阈值划分出异常样本.最后利用汽轮机转子轮槽的实际加工监测数据进行验证,并与局部异常因子算法、基于密度的聚类算法、K均值算法、孤立森林算法、一分类支持向量机等方法进行对比分析,验证本方法的有效性.Aimed at the problem that the accuracy of tool remaining life prediction was reduced due to the abnormal monitoring data in machining process,a data anomaly detection method based on self-paced learning was proposed.First,a multi-layer perceptron model was established to correlate the tool processing monitoring data with the tool remaining life.Next,in the process of updating model weight,the model weight parameters were fixed first,and the loss threshold of abnormal samples was obtained by predicting loss fitting Gaussian distribution.Then,the loss function based on the self-paced learning method was constructed to update model parameters iteratively.At the end of the model training,abnormal samples were divided according to the loss threshold.Finally,the actual processing monitoring data of turbine rotor groove were used to verify the proposed method,and compare with the local anomaly factor algorithm,the density-based clustering algorithm,the K-means algorithm,the isolated forest algorithm,and the one-class support vector machines.

关 键 词：刀具加工监测 数据质量 异常检测 自步学习 

分 类 号：TH166[机械工程—机械制造及自动化]