文献类型：期刊文献

中文题名：基于1D-SE-CDM的1DCNN-KAN小样本故障诊断方法研究

英文题名：Research on 1 DCNN-KAN Small Sample Fault Diagnosis Method Based on 1 D-SE-CDM

作者：王子云 杨旭东 孙栋

第一作者：王子云

机构：[1]贵州大学机械工程学院,贵州贵阳550025;[2]贵州理工学院机械工程学院,贵州贵阳550025

第一机构：贵州大学机械工程学院,贵州贵阳550025

年份：2026

卷号：54

期号：4

起止页码：25-30

中文期刊名：机床与液压

外文期刊名：Machine Tool & Hydraulics

收录：;北大核心:【北大核心2023】；

基金：贵州省科技计划项目(黔科合成果[2023]一般812);2023年度省工业和信息化发展专项资金科技创新项目(202308)。

语种：中文

中文关键词：轴承故障诊断;小样本学习;条件扩散模型;神经网络

外文关键词：bearing fault diagnosis;small sample learning;conditional diffusion model;neural network

摘要：针对工业场景中轴承故障数据稀缺导致诊断模型训练不足的问题,提出一种面向小样本条件基于1D-SE-CDM的1DCNN-KAN故障诊断方法,旨在提高轴承故障的诊断效率。采用改进的条件扩散模型1D-SE-CDM生成新的一维振动数据,并引入多尺度香农熵作为损失函数,以充分利用小样本数据,获得高质量且具备多样性的数据。将生成的数据输入1DCNN-KAN网络,利用其双通道结构和嵌入注意力机制进行故障诊断,从而进一步提升诊断准确性。最后,通过CWRU滚动轴承数据集进行实验验证。结果表明:1D-SE-CDM生成的信号与原始数据高度相似,加入生成数据后诊断准确率显著提升,仅添加每类10个生成样本即从89.8%提高至99.2%,样本增至40个时可达100%;1D-SE-CDM在相同条件下优于其他生成模型,诊断准确率最高且波动小;1DCNN-KAN诊断网络同样优于对比模型;极端小样本测试中,即使每类仅用1个原始样本生成数据,加入生成样本后最高准确率仍可达95%,较仅使用原始样本提升5.2%。该方法对一维振动信号的小样本故障诊断具有一定的理论价值与应用潜力。
To address the issue of insufficient model training stemming from scarce bearing fault data in industrial scenarios,a 1DCNN-KAN fault diagnosis method based on 1D-SE-CDM(1D-Shannon entropy conditional diffusion model)for small-sample conditions was proposed,aiming to improve the efficiency of bearing fault diagnosis.An improved conditional diffusion model,1D-SE-CDM,was employed to generate new 1D vibration data.Additionally,multi-scale Shannon entropy was incorporated as the loss function to fully leverage small-sample data and obtain high-quality and diverse datasets.The generated data were input into the proposed 1DCNN-KAN network,which employed its dual-channel structure and embedded attention mechanism to perform fault diagnosis,further improving diagnostic accuracy.Finally,experiments were conducted using the CWRU rolling bearing dataset.The results show that the signals generated by 1D-SE-CDM are highly similar to the original data,and adding the generated data significantly improves diagnostic accuracy-using only 10 generated samples per class raises the accuracy from 89.8%to 99.2%,and increasing the samples to 40 per class achieves 100%accuracy.Under identical conditions,1D-SE-CDM is proven to be superior to other generative models,achieving the highest diagnostic accuracy with minor fluctuations.The 1DCNN-KAN diagnostic network is also shown to be superior to the compared models.In the extreme small-sample test,even when only 1 original sample per class is used to generate data,the highest accuracy is still achieved at 95%after the generated data are incorporated,which is 5.2%higher than that obtained using only the original samples.This method is considered to possess certain theoretical value and application potential for the small-sample fault diagnosis of 1D vibration signals.