| 孙哲*,金华强* **,顾江萍**,黄跃进*,王新雷***,郑爱武****,沈希*.基于生成对抗网络的热泵系统故障诊断数据集扩充方法研究[J].高技术通讯(中文),2021,31(12):1280~1292 |
| 基于生成对抗网络的热泵系统故障诊断数据集扩充方法研究 |
| Data augmentation for heat pump systems fault diagnosis based on generative adversarial networks |
| |
| DOI：10.3772/j.issn.1002-0470.2021.12.006 |
| 中文关键词: 生成对抗网络（GAN）；深度学习；数据集扩充；热泵系统；故障诊断 |
| 英文关键词: generative adversarial network (GAN), deep learning, data augmentation, heat pump system, fault diagnosis |
| 基金项目: |
| 作者 | 单位 | | 孙哲* | | | 金华强* ** | | | 顾江萍** | | | 黄跃进* | | | 王新雷*** | | | 郑爱武**** | | | 沈希* | |
|
| 摘要点击次数: 2033 |
| 全文下载次数: 1523 |
| 中文摘要: |
| 随着深度学习的发展，越来越多的热泵系统故障诊断方法引入深度学习技术并取得了较好的效果。基于深度学习的故障诊断技术需要依赖大量带有标记的故障数据，而现实中这类数据获取较为困难，这限制了智能诊断技术的应用。针对这一问题，本文提出利用生成对抗网络（GAN）学习故障数据的分布，并生成更多的标记数据，实现故障数据集的扩充。针对热泵系统运行数据结构复杂且不同故障间差异小给模型学习带来较大困难这一问题，本文提出利用热泵系统基准模型将运行数据转化为残差数据并作为训练数据，降低数据复杂度，增加差异性。利用MMD指标和1-NN指标对生成的数据进行分析，发现生成数据分布和真实数据接近，且利用残差数据训练的GAN模型质量更高。利用故障诊断方法对引入不同比例生成数据的模型训练结果进行分析，发现生成数据的引入可以提高数据量不足条件下的故障诊断精度。实验结果表明，基于GAN的数据扩充方法可有效降低智能诊断对标记数据的依赖，是一种应用前景广阔的技术。 |
| 英文摘要: |
| With the development of deep learning, more and more heat pump system fault diagnosis methods use deep learning technology and get good results. The fault diagnosis technology based on deep learning needs to rely on a large number of labeled fault data, but in reality, such data is very difficult to obtain, which limits the application of intelligent diagnosis technology. Aiming at this issue, the generative adversarial network (GAN) is proposed to learn the distribution of fault data and generate more labeled data to achieve the augmentation of the fault data set. For the complex operation data structure of the heat pump systems and the small difference of data value between different faults bring great difficulty to model learning, this paper proposes to use the heat pump system benchmark model to convert the operation data into residual data and use it as training data to reduce data complexity and increase the difference of data value. Using the MMD and 1-NN indicator to analyze the generated data, it is found that the distribution of the generated data is close to the real data, and the GAN model trained with residual data is of higher quality. Using the method of fault diagnosis to analyze the training results of models that use different amounts of generated data, it is found that the introduction of generated data can improve the accuracy of fault diagnosis under insufficient data conditions. The experimental results prove that the GAN-based data augmentation method can effectively reduce the dependence of intelligent diagnosis on labeled data, and has broad application prospects. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
