Page 163 - 《振动工程学报》2026年第3期
P. 163
第 39 卷第 3 期 振 动 工 程 学 报 Vol. 39 No. 3
2026 年 3 月 Journal of Vibration Engineering Mar. 2026
自编码器在机械设备未知故障检测中的应用
孟祥恒 , 许方敏 , 张永军 , 车天翊 , 卫云龙 , 刘 汭 , 吕 鹏 3
3
3
1
1
2
1
(1. 北京邮电大学信息与通信工程学院,北京 100876; 2. 北京邮电大学电子工程学院,北京 100876;
3. 宁德时代新能源科技股份有限公司,福建 宁德 352100)
摘要: 在实际的工业环境中,随着机械设备运行时间的增加,往往会出现一些未知故障。此时,传统的故障诊断方法并不能及
时发现这些故障,从而影响正常生产进度,造成生产损失。因此,本文提出了一种基于多解码器的自编码分类(multi⁃decoders
AE classifier, MDAEC)模型,在对已知样本进行分类的同时保持对未知故障样本的检测。利用自编码器的特征提取能力,将
中间的隐藏特征输入一个分类网络中,使用已知样本对自编码器以及分类网络进行训练。根据已知样本类别训练各自的解码
器,以提高模型对未知样本的检测能力。在检测时,通过自编码器的重构误差以及阈值可判断输入样本是否为未知故障,如果
判断输入样本属于已知样本,则将隐藏层的特征输入分类网络中即可得到具体故障类型。试验结果表明,本文所提出方法在
对未知故障样本进行准确识别的同时保持着对已知故障样本较高的分类准确率。
关键词: 故障诊断; 自编码器; 深度神经网络; 特征提取; 未知故障检测
+
中图分类号: TH165 .3; TP183; TH133.33 文献标志码: A DOI:10.16385/j.cnki.issn.1004-4523.202403068
Application of autoencoder in unknown fault detection
of mechanical equipment
2
1
1
1
MENG Xiangheng , XU Fangmin , ZHANG Yongjun , CHE Tianyi ,
3
3
WEI Yunlong , LIU Rui , LYU Peng 3
(1.School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876,
China; 2.School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China;
3.Contemporary Amperex Technology Co., Ltd., Ningde 352100, China)
Abstract: In recent years, the rapid development of artificial intelligence has led to the successful application of numerous machine
learning algorithms in mechanical equipment fault diagnosis. However, most of these algorithms need to be trained based on known
normal and fault samples, and the detection results of new data can only be one of these known categories. In practical industrial
settings, however, previously unknown faults often emerge as equipment operational time accumulates. Consequently, the tradi⁃
tional fault diagnosis method cannot find these new faults in time, which affects the normal production schedule and causes produc⁃
tion loss. Therefore, this paper proposes a multi-decoders autoencoder classifier (MDAEC) model based on multi-decoders,
which classifies known samples while maintaining the detection of unknown fault samples. By using the feature extraction capability
of the autoencoder, the intermediate hidden features are input into a classification network, and the autoencoder and classification
network are trained by using known samples. Respective decoders are trained according to the known sample categories to improve
the model’s detection ability of unknown samples. During detection, the reconstruction error and threshold of the autoencoder can
be used to determine whether the input sample is a unknown fault. If the input sample is known, the specific sample type can be ob⁃
tained by inputting the features of the hidden layer into the classification network. Experimental results demonstrate that the pro⁃
posed method can accurately identify unknown fault samples while maintaining a high classification accuracy for known samples.
Keywords: fault diagnosis;auto-encoder;deep neural network;feature extraction;unknown fault detection
[1]
在故障诊断方面,传统的机器学习方法,比如支 出智能诊断 ,但是这些方法在处理复杂、大规模、
[2]
持向量机(support vector machine, SVM)以及随机 高维度的数据时很难达到理想的效果 。深度学习
森林(random forest, RF),虽然能够对机械故障作 由于能够通过不断学习自动提取数据特征,并且具
收稿日期: 2024-03-29; 修订日期: 2024-10-23

