液压与气动2025，Vol.49Issue(11)：47-55,9.DOI:10.11832/j.issn.1000-4858.2025.11.006
改进解耦对比学习的多种极端长尾数据故障诊断
Improving Decoupled Supervised Contrastive Learning for Fault Diagnosis of Multiple Extreme Long-tail Dataset
摘要
Abstract
Due to the limitations of factors such as on-site space and economic,the sample acquisition size of certain fault type of each component is very small,and thus a fault dataset of extreme long-tail distribution is formed,which makes the traditional decoupled supervised contrastive learning model unable to conduct effective diagnosis.Therefore,an improved decoupled supervised contrastive learning model is proposed,namely contrastive distillation type equilibrium decoupled supervised contrastive learning mode.Firstly,the synthetic minority oversampling method is introduced to generate tail samples appropriately to alleviate the problem of dataset imbalance;secondly,the parameter contrastive learning is introduced to construct a double contrast mechanism to increase the contribution of the tail and the diagnosis accuracy,solving the problem of sparse feature distribution of the tail type;finally,the type balanced self-distillation is introduced to solve the problem of insufficient representation of tail features through knowledge transfer.The experimental analysis of two extreme distribution forms of measured fault samples from the hydraulic pump,the gear and the rolling bearing shows that the proposed model can effectively solve the problem of extreme long-tail distribution,with diagnostic accuracies reaching up to 90.93％and 98.61％,respectively.In addition,the accuracy of the proposed method is 40.99％higher than that of the original method,and 76.97％and 35.83％higher than those of the traditional wide parameter contrastive learning and balanced contrastive learning methods.关键词
对比学习/过采样/类别平衡自蒸馏/极端长尾分布/故障诊断Key words
contrastive learning/oversampling/type balanced self-distillation/extreme long-tail distribution/fault diagnosis分类
机械工程引用本文复制引用
林帅恒,王彦凤,郑直,朱占辉,赵文博..改进解耦对比学习的多种极端长尾数据故障诊断[J].液压与气动,2025,49(11):47-55,9.基金项目
河北省自然科学基金资助项目(E2022209086) （E2022209086）
河北省级研究生专业学位教学案例(库)项目(KCJSZ2025050) （库）
