基于自适应边际损失的小样本故障诊断方法  

Fault diagnosis method for small sample based on adaptive margin loss

作　　者：熊鹏伟 李志农[1] 刘晨宇 冯博 谷丰收[2] XIONG Pengwei;LI Zhinong;LIU Chenyu;FENG Bo;GU Fengshou(Key Laboratory of Non-Destructive Testing Technology of Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China;Centre for Efficiency and Performance Engineering,University of Huddersfield,Huddersfield HD13DH,UK)

机构地区：[1]南昌航空大学无损检测技术教育部重点实验室,南昌330063 [2]哈德斯菲尔德大学效率与性能工程中心,英国哈德斯菲尔德HD13DH

出　　处：《兵器装备工程学报》2024年第9期253-260,共8页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(52075236);江西省自然科学基金重点项目(20212ACB202005)。

摘　　要：针对基于度量学习的小样本故障诊断方法,在模型优化过程中对边际影响的忽略导致模型对训练数据过于敏感,进而产生过拟合。为此,构造了一种自适应边际损失函数,帮助模型学习样本之间的相对距离,以获得足够的距离信息,提高对新样本的泛化能力。另外,根据训练数据的分布和模型的性能,自动调整边际的大小,使其自适应地区分不同的故障类别。为更好地解决小样本问题,提出了元度量学习框架,采用元学习片段式训练模式。在度量模块中,引入余弦相似性以提高方法的表达能力,并指导模型的优化和训练,使其更好地适应小样本数据。为验证所提方法的有效性,使用带故障的无人机飞行日志数据构建了数据集,并将所提方法与传统的度量学习的方法进行了对比。实验结果表明,所提方法在无人机小样本故障诊断中表现出良好的诊断性能和稳定性,为小样本故障诊断提供了一种有效方案。Existing metric learning-based fault diagnosis methods lack consideration of margins while optimized by the loss function,making the model overly sensitive to the training data.To this end,an adaptive margin loss function is constructed to reduce overfitting and improve the model performance on new samples.This margin loss helps the model to learn the relative distances between samples and thus gain enough distance information to generalize to new samples.It automatically adjusts the margin size according to the distribution of training data and the model performance to adaptively distinguish different fault categories.To better meet the small sample,a small sample fault diagnosis method based on a meta-metric network with adaptive margin loss is proposed by combining the meta-learning episode training mode and the flexibility of the metric module in metric learning.The metric module introduces the cosine similarity to improve the expressiveness,then guides the optimization and model training to fit the small sample better.To validate the effectiveness of the proposed method,a dataset consisting of UAV flight log data with faults is used,and the proposed method is compared with the traditional methods based on metric learning.Experiment results show that the proposed method performs well in small sample fault diagnosis for UAV with good diagnostic performance and stability.The proposed method provides an effective solution to ensure the safe and reliable operation of UAV.

关 键 词：故障诊断 小样本 元度量学习 自适应边际损失 无人机 

分 类 号：V279[航空宇航科学与技术—飞行器设计] TP183[自动化与计算机技术—控制理论与控制工程]