摘要
尽管工业条件下可获取大量轴承状态监测数据,但其价值密度低且多为正常状态,可利用的不同类型故障数据较少。针对少样本条件下难以实现高准确率轴承故障诊断的问题,提出一种基于多尺度卷积关系网络的轴承故障诊断方法。该方法首先利用关系网络建立已标记样本之间的对比关系模型;其次,在网络的第一层利用多个大小不同卷积核提取特征并进行特征融合,以增强模型在数据稀缺的条件下对丰富性和互补性故障特征的提取能力;此外,考虑交叉熵损失函数,以提升模型对不同故障类型中判别性特征的提取能力。在帕德博恩大学轴承数据集下,仅利用50条样本训练模型,所提方法相较于WDCNN、SECNN、孪生网络、原型网络和关系网络对1000条无标记样本的平均测试准确率分别提升33.66%,28.63%,7.62%,7.82%和4.21%。此外,对机车轴承数据集添加SNR为-1 dB的高斯白噪声以模拟强噪声干扰环境,所提方法仅利用20条训练样本对1200条测试样本达到89.83%的较高诊断精度。实验结果显示,在小样本训练条件下,所提方法能够有效提升模型的泛化、抗噪和辨识能力。
Although a large amount of bearing condition monitoring data can be obtained under industrial conditions,its value density is low and mostly normal.There are fewer different types of fault data available.A rolling bearing fault diagnosis method based on a multi-scale convolutional relation network is proposed to address the difficult problem of achieving high accuracy with limited samples.Firstly,a frame based on a relation network is designed to establish a comparative relationship between the labeled samples.Secondly,the first layer of the designed frame is improved by multiple convolutional kernels with different sizes to enhance the extraction and fusion of rich and complementary fault features under data scarcity.Thirdly,the crossentropy loss function is considered to enhance the model’s extraction capability for discriminative features in different fault types.With the Paderborn University bearing data,the proposed method improves the average test accuracy by 33.66%,28.63%,7.62%,7.82%and 4.21%compared to WDCNN,SECNN,Siamese network,Prototypical network,and Relation network for 1000 unlabeled samples using only 50 samples to train the model,respectively.Additionally,the proposed method achieves a high diagnostic accuracy of 89.83%for 1200 test samples using only 20 training samples by adding white Gaussian noise with SNR of−1 dB to the locomotive bearing dataset to simulate a strong noise interference environment.The experimental illustrates show that the proposed method can effectively improve the model’s ability of generalization,anti-noise and discrimination under limited training samples.
作者
郝伟
丁昆
暴长春
贺婷婷
陈仰辉
张楷
HAO Wei;DING Kun;BAO Changchun;HE Tingting;CHEN Yanghui;ZHANG Kai(CRRC Qingdao Sifang Co.,Ltd.,Qingdao 266111,China;School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China;Tangshan Research Institute,Southwest Jiaotong University,Tangshan 063000,China)
出处
《中国测试》
CAS
北大核心
2024年第3期160-168,共9页
China Measurement & Test
基金
中央高校基本科研业务费(2682022CX006)
国家重点研发计划项目(2021YFB3400700)。
关键词
轴承故障诊断
小样本
关系网络
多尺度卷积网络
bearing fault diagnosis
limited samples
relation network
multi-scale convolutional network
