| 柏受军,成志新,袁啸林,王静,左桂忠,余耀伟.时间序列生成对抗网络架构下的分子泵退化数据生成研究[J].电子测量与仪器学报,2025,39(6):195-203 |
| 时间序列生成对抗网络架构下的分子泵退化数据生成研究 |
| Research on molecular pump degradation data generation under timeseries generative adversarial network architecture |
| |
| DOI： |
| 中文关键词: 数据生成 分子泵 时间序列生成器对抗网络 退化预测 |
| 英文关键词:data generation molecular pump time series generator adversarial network degradation prediction |
| 基金项目:国家重点研发计划（2024YFE03250300）、中国国家自然科学基金（11905254）、安徽省自然科学基金（2108085QA38）、中国博士后科学基金（2021000278）、安徽省高等学校自然科学研究重点项目（2023AH052105）、安徽工程大学安徽省电力传动与控制重点实验室开放项目（DQKJ202207）、安徽省教育厅重点项目（2023AH050924）资助 |
|
|
| Author | Institution |
| Bai Shoujun | Key Laboratory of Advanced Perception and Intelligence Control of High-end Equipment，Anhui Polytechnic University,
Wuhu 241000, China |
| Cheng Zhixin | Key Laboratory of Advanced Perception and Intelligence Control of High-end Equipment，Anhui Polytechnic University,
Wuhu 241000, China |
| Yuan Xiaolin | Institute of Plasma Physics, Chinese Academy of Sciences, Hefei 230031, China |
| Wang Jing | Chaohu University, Chaohu 238024,China |
| Zuo Guizhong | Institute of Plasma Physics, Chinese Academy of Sciences, Hefei 230031, China |
| Yu Yaowei | Institute of Plasma Physics, Chinese Academy of Sciences, Hefei 230031, China |
|
| 摘要点击次数: 818 |
| 全文下载次数: 483 |
| 中文摘要: |
| 为了确保托卡马克实验的安全运行,对关键真空获取设备分子泵的可靠性评估至关重要。然而,有限的退化数据导致现有的预测方法准确性较低。针对这一挑战,提出了一种基于时间序列生成对抗网络（TGAN）的分子泵退化数据生成方法,旨在通过生成数据来扩充数据集,进而提高预测模型的准确性和可靠性。该方法创新性地结合了Transformer网络和TGAN,并通过引入威布尔分布提高了生成数据的质量，再利用长短期记忆网络对生成的退化数据进行退化预测。实验结果表明,TGAN Transformer能有效生成满足分子泵退化预测需求的数据,显著提升了预测的准确性和可靠性,为分子泵的可靠性分析和安全运行提供了可靠支持。通过对比实验，TGAN-Transformer在均方根误差（RMSE）指标上相较于生成对抗网络（GAN）、深度卷积生成对抗网络（DCGAN）、递归条件生成对抗网络（RCGAN）、变分自编码器（VAE）和条件变分自编码器（CVAE）分别提升51%、48% 、36%、40%、30%；在平均绝对误差（MAE）指标上，分别提升52%、49%、38%、42%、33%，证明了其在分子泵退化预测中的有效性。未来的研究可进一步优化生成网络结构,探索更多生成对抗网络的变种,以提高生成数据的多样性和真实性,从而进一步提升退化预测的准确性和可靠性。 |
| 英文摘要: |
| To ensure the safe operation of Tokamak experiments, the reliability assessment of key vacuum acquisition equipment, specifically molecular pumps, is crucial. However, limited degradation data has resulted in low accuracy of existing predictive methods. To address this challenge, a degradation data generation method for molecular pumps based on time series generative adversarial networks (TGAN) has been proposed, aimed at augmenting the dataset through generated data to enhance the accuracy and reliability of predictive models. This method innovatively combines Transformer networks with TGAN and improves the quality of the generated data by incorporating Weibull distribution. Furthermore, long short-term memory networks are utilized for degradation prediction of the generated data. Experimental results demonstrate that TGAN-Transformer can effectively generate data that meets the needs of molecular pump degradation prediction, significantly enhancing prediction accuracy and reliability, thereby providing solid support for the reliability analysis and safe operation of molecular pumps. Through comparative experiments, TGAN Transformer has improved RMSE indicators by 51%, 48%, 36%, 40%, and 30% compared to GAN, DCGAN, RCGAN, VAE, and CVAE, respectively. On the MAE index, they increased by 52%, 49%, 38%, 42%, and 33% respectively, demonstrating their effectiveness in predicting molecular pump degradation. Future research may further optimize the structure of the generation network and explore more variants of generative adversarial networks to improve the diversity and authenticity of generated data, thereby further enhancing the accuracy and reliability of degradation predictions. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
