PDF(12376 KB)
Sample Imbalanced Fault Diagnosis Method Based on Multi-channel Data Double Augmentation
GUOYiming, TONGYifei, HEFei, XIEZhongqu, SONGShida, HUANGJing
PDF(12376 KB)
Sponsored by: China Association for Science and Technology (CAST)
Editor-In-Chief: Xu Yida
ISSN 1000-1093
Hosted By: China Ordnance Society
Published By: Acta Armamentarii
CN 11-2176/TJ
PDF(12376 KB)
Sample Imbalanced Fault Diagnosis Method Based on Multi-channel Data Double Augmentation
In complex manufacturing processes,it is crucial to collect and analyze the multi-channel data for condition monitoring and fault diagnosis.The existing methods are used to difficultly handle the problems of complex spatial-temporal correlation and sample imbalance of the multi-channel data.To solve these problems,a sample imbalance fault diagnosis method based on multi-channel data double augmentation is developed.The proposed method has the advantages of two-stage data augmentation and global optimization.It first learns the fault features,and then converts them into the multi-channel data for the data augmentation.The distribution difference evaluation mechanism is introduced to effectively describe the correlation between different channels,and a multi-objective global optimization strategy is designed to improve the quality of generated data.The effectiveness of the proposed method is verified by studying a real-world case.The experimental results show that the data double augmentation method can effectively expand the multi-channel data with small samples,and the global optimization strategy can improve the performance of generated data in the fault diagnosis.Compared with existing methods,the proposed method has higher fault diagnosis accuracy in various sample imbalance scenarios.
multi-channel data / sample imbalance fault diagnosis / data augmentation / global optimization {{custom_keyword}} /
Table 1 MSE value between generated and real data表1 生成数据和真实数据间MSE值 |
| 数据通道 | 传统GAN | 所提方法 |
|---|---|---|
| 1 | 0.0331 | 0.0039 |
| 2 | 0.0361 | 0.0065 |
| 3 | 0.0531 | 0.0048 |
| 4 | 0.0540 | 0.0054 |
Table 2 MSE value between different channels表2 不同通道数据间MSE值 |
| 生成数据 | 真实数据 | |||
|---|---|---|---|---|
| 通道1 | 通道2 | 通道3 | 通道4 | |
| 通道1 | 0.0039 | 0.0167 | 0.0177 | 0.0109 |
| 通道2 | 0.0193 | 0.0065 | 0.0182 | 0.0159 |
| 通道3 | 0.0237 | 0.0185 | 0.0048 | 0.0103 |
| 通道4 | 0.0133 | 0.0178 | 0.0115 | 0.0054 |
Table 3 Diagnosis accuracy when Faults 1-5 are minority samples表3 故障1-5分别为少数样本时的诊断准确率 % |
| 对比方法 | 诊断准确率 | ||||
|---|---|---|---|---|---|
| 故障1 | 故障2 | 故障3 | 故障4 | 故障5 | |
| 随机过采样 | 99.25 | 98.88 | 100 | 89.14 | 98.88 |
| SMOTE | 99.63 | 99.63 | 100 | 89.89 | 93.26 |
| ADASYN | 99.63 | 99.63 | 100 | 90.26 | 91.76 |
| 传统GAN | 99.63 | 99.63 | 100 | 97.75 | 99.25 |
| 所提方法 | 99.63 | 99.63 | 100 | 99.25 | 100 |
Fig.11 Fault diagnosis result when Fault 4 samples are minority图11 故障4为少数样本时的诊断结果 |
Table 4 Diagnosis accuracy when Faults 4-5 are minority samples %表4 故障4-5均为少数样本时的诊断准确率 |
| 状态 | 对比方法 | ||||
|---|---|---|---|---|---|
| 随机过采样 | SMOTE | ADASYN | 传统GAN | 所提方法 | |
| 故障1 | 100 | 100 | 100 | 100 | 100 |
| 故障2 | 100 | 100 | 100 | 100 | 100 |
| 故障3 | 100 | 100 | 100 | 100 | 100 |
| 故障4 | 10.34 | 10.34 | 17.24 | 79.31 | 93.10 |
| 故障5 | 13.79 | 48.28 | 41.38 | 93.10 | 100 |
| 正常 | 100 | 100 | 100 | 90.98 | 100 |
| 整体 | 80.90 | 84.64 | 84.64 | 92.88 | 99.25 |
| [1] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [2] |
雷亚国, 贾峰, 孔德同, 等. 大数据下机械智能故障诊断的机遇与挑战[J]. 机械工程学报, 2018, 54(5):94-104.
机械故障是风力发电设备、航空发动机、高档数控机床等大型机械装备安全可靠运行的“潜在杀手”。故障诊断是保障机械装备安全运行的“杀手锏”。由于诊断的装备量大面广、每台装备测点多、数据采样频率高、装备服役历时长，所以获取了海量的诊断数据，推动故障诊断领域进入了“大数据”时代。而机械智能故障诊断有望成为大数据下诊断机械装备故障的“一把利器”。与此同时，大数据给机械智能故障诊断的深入研究和应用提供了新的机遇：“数据为王”的学术思想有望成为主流、诊断整机或系统级对象成为可能、全面解析故障演化过程成为趋势等；但也遇到了新的挑战：数据大而不全呈“碎片化”、故障特征提取受制于人为经验、浅层诊断模型诊断精度低等。阐述了机械智能故障诊断大数据的特点；从信号获取、特征提取、故障识别与预测三个环节，综述了机械智能故障诊断的国内外研究进展和发展动态；指出了机械智能故障诊断理论与方法在大数据背景下的挑战；最后讨论了应对这些挑战的解决途径与发展趋势。
Faults are a potential killer of large-scale mechanical equipment, such as wind power equipment, aircraft engines and high-end CNC machine. And fault diagnosis plays an irreplaceable role in ensuring the health operation of such equipment. Since the amount of the equipment diagnosed is great and the number of the sensors for the equipment is large, massive data are acquired by the high sampling frequency after the long-time operation of equipment. Such massive data promote fault diagnosis to enter the era of big data. And machinery intelligent fault diagnosis is a promising tool to deal with mechanical big data. In the big data era, new opportunities have been brought to intelligent fault diagnosis. For instance, data-centric academic thinking will become mainstream, it makes fault diagnosis in the system level possible, and a comprehensive analysis of faults becomes a trend. Meanwhile, new challenges have also been brought:the data are big but fragmentary, the fault feature extraction relies on much prior knowledge and diagnostics expertise, and the generalization ability of the shallow diagnosis model is weak. The characteristics of big data in intelligent fault diagnosis are discussed, and the inland and overseas research advances are reviewed from the three steps of intelligent fault diagnosis. The existing key problems of the current research in the era of big data are pointed out, and the approaches and research directions to these problems are discussed in the end.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [3] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [4] |
邵海东, 李伟, 刘翊, 等. 时变转速下基于双阈值注意力生成对抗网络和小样本的转子-轴承系统故障诊断[J]. 机械工程学报, 2023, 59(12):215-224.
可用故障数据的匮乏给时变转速下转子-轴承系统的端到端故障诊断带来严重挑战，生成对抗网络为解决小样本故障诊断问题提供新思路，但其仍存在梯度消失、全局关联特征学习能力较弱和训练效率较低等缺点。因此，提出一种双阈值注意力生成对抗网络，用于生成高质量的红外热成像图片，以解决时变转速下转子-轴承系统的小样本故障诊断难题。首先，结合Wasserstein距离和梯度惩罚设计新型对抗损失函数，避免训练过程中的梯度消失。其次，构建注意力嵌入的生成对抗网络以指导学习红外热成像图片的全局热力关联特征。最后，开发双阈值训练机制进一步提高生成样本质量和训练效率。将所提方法用于分析转子-轴承系统的实测红外热成像图片，结果表明，所提方法能辅助准确诊断时变转速及小样本下的不同故障模式，性能优于目前常用的生成对抗网络方法。
End-to-end fault diagnosis of rotor-bearing system under time-varying speeds using a few samples is challenging. Despite generative adversarial network (GAN) provides a way to solve the problem of small-sample fault diagnosis, it still has some limitations, such as gradient vanishing, weak extraction of global correlation features, and low training efficiency. Therefore, a dual-threshold attention-embedded GAN is proposed for generating high-quality infrared thermal (IRT) images to solve small-sample fault diagnosis of rotor-bearing system under time-varying speeds. First, Wasserstein distance and gradient penalty are combined to design the new adversarial loss function to avoid gradient vanishing. Second, attention-embedded GAN is constructed to guide learn global thermal-correlation features of the IRT images. Finally, dual-threshold training mechanism is developed to further improve the generation quality and training efficiency. The proposed method is used to analyze the collected small IRT images of a rotor-bearing system, and the results show that the proposed method can accurately diagnosis different fault modes using small samples under time-varying speeds, which is superior to other popular GANs.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [5] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [6] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [7] |
吴春志, 冯辅周, 吴守军, 等. 一种有效的不均衡样本生成方法及其在行星变速箱故障诊断中的应用[J]. 兵工学报, 2019, 40(7):1349-1357.
针对实际运行中行星变速箱故障数据较少、各个状态样本不均衡的问题，提出了由Wasserstein生成式对抗网络(WGAN)样本生成模型和卷积神经网络(CNN)分类模型组合的WGAN-CNN故障诊断分类模型。该模型对故障数据的频谱信号进行过采样，以扩展故障样本数量，从而更好地对故障状态进行分类。采用加州大学欧文分校人工数据集对WGAN生成模型以及经典过采样方法进行对比，并在行星变速箱故障试验台上进行验证。结果表明，样本不均衡会严重影响分类结果，而WGAN-CNN模型可以很好地扩充故障样本集，提高在故障样本稀少情况下的诊断准确率。
A fault diagnosis classification model based on WGAN-CNN is proposed for few fault data of planetary gearbox and the imbalanced samples of each state in actual operation. The proposed model is a combination of Wasserstein generative adversarial network (WGAN), a sample generation model, a convolutional neural network (CNN), and a sample classification model. The model is used to oversample the spectral signals of fault data and expand the number of fault samples, thus classifying the fault states better. UCI artificial datasets were used to compare WGAN generation model and classical oversampling methods, and verified on a planetary gearbox fault test rig. The results show that the imbalanced samples seriously affect the classification results, and the WGAN-CNN model can well expand the fault sample datasets and improve the diagnostic accuracy in the case of rare fault samples. Key
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [8] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [9] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [10] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [11] |
徐鹏, 皋军, 邵星. 基于AMCNN-BiGRU的滚动轴承故障诊断方法研究[J]. 振动与冲击, 2023, 42(18):71-80.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [12] |
何强, 唐向红, 李传江, 等. 负载不平衡下小样本数据的轴承故障诊断[J]. 中国机械工程, 2021, 32(10):1164-1171.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [13] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [14] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [15] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [16] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [17] |
张玉彦, 张永奇, 孙春亚, 等. 不平衡样本下基于生成式对抗网络的风机叶片开裂状态识别[J]. 计算机集成制造系统, 2023, 29(2):532-543.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [18] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [19] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [20] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [21] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [22] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [23] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [24] |
郭伟, 邢晓松. 基于改进卷积生成对抗网络的少样本轴承智能诊断方法[J]. 中国机械工程, 2022, 33(19):2347-2355.
Few bearing samples usually led to inadequate learning and low diagnosis accuracy. To solve this problem， an improved convolutional generative adversarial network was constructed. It made full use of data generation ability of the GAN and learning ability of deep CNN， so that the intelligent fault diagnosis might be conducted for the bearings with few samples under varying working conditions. First， a deep convolutional GAN was constructed. The deep features in few real data were mined through adversarial learning between the generator and discriminator of GAN， and then the generator might generate simulated data exactly like real one to make up for the lack of very few samples. Then， the dense block and dilated convolutions were combined with the CNN， named as DDCNN， to improve the learning ability by extending the network depth and perception range. As a result， the DDCNN may identify tiny differences in multi-class datasets and enhance feature extraction. Finally， the proposed method was verified by using bearing datasets with few samples under fixed conditions and varying rotating speeds， and was compared with other frameworks. The experimental results indicate that DDCNN has higher diagnosis accuracy for bearings with few samples and noisy conditions.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [25] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [26] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [27] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [28] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [29] |
李晓雄, 张淑宁, 赵惠昌, 等. 基于1DC-CGAN和小波能量特征的引信小样本地形目标识别[J]. 兵工学报, 2022, 43(10):2545-2553.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [30] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [31] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [32] |
REZVANI012-11 S,
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [33] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| [34] |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
| {{custom_ref.label}} |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
PDF(12376 KB)
Fig.1 The overall structure of the proposed modelFig.2 Fault feature learning network structureFig.3 Multi-channel data generation network structureFig.4 Global optimization strategyFig.5 Multi-step forging machine and technological processFig.6 Four-channel pressure dataFig.7 Fault diagnosis results of imbalanced sampleFig.8 Multi-channel data generated by two methods
Table 1 MSE value between generated and real dataTable 2 MSE value between different channelsFig.9 Generated fault featureFig.10 Visualization of fault features based on the t-SNETable 3 Diagnosis accuracy when Faults 1-5 are minority samplesFig.11 Fault diagnosis result when Fault 4 samples are minorityFig.12 Fault diagnosis result when Fault 5 samples are minorityTable 4 Diagnosis accuracy when Faults 4-5 are minority samples %/
| 〈 |
|
〉 |
