| 陈凯诺,张福光,韩建立,尹延涛,杜光传.基于Transformer小样本多源数据融合的装备剩余寿命预测评估[J].装备环境工程,2024,21(11):65-73. CHEN Kainuo,ZHANG Fuguang,HAN Jianli,YIN Yantao,DU Guangchuan.Equipment Residual Life Prediction and Evaluation with Transformer Small Sample Multi-source Data Fusion[J].Equipment Environmental Engineering,2024,21(11):65-73. |
| 基于Transformer小样本多源数据融合的装备剩余寿命预测评估 |
| Equipment Residual Life Prediction and Evaluation with Transformer Small Sample Multi-source Data Fusion |
| 投稿时间：2024-07-15 修订日期：2024-08-01 |
| DOI：10.7643/issn.1672-9242.2024.11.009 |
| 中文关键词: 贮存延寿工程 剩余寿命 Transformer 多源数据融合 迁移学习 源域 目标域中图分类号:TJ760.1 文献标志码:A 文章编号:1672-9242(2024)11-0065-09 |
| 英文关键词:life extension engineering residual life transformer multi-source data fusion transfer learning source domain target domain |
| 基金项目:国家自然科学基金(62371465)；山东省青创团队(2022kj084)；山东省自然科学基金(ZR2020QF010) |
|
|
| Author | Institution |
| CHEN Kainuo | Naval Aviation University, Shandong Yantai 264001, China |
| ZHANG Fuguang | Naval Aviation University, Shandong Yantai 264001, China |
| HAN Jianli | Naval Aviation University, Shandong Yantai 264001, China |
| YIN Yantao | Naval Aviation University, Shandong Yantai 264001, China |
| DU Guangchuan | Naval Aviation University, Shandong Yantai 264001, China |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 解决某类高可靠性航空装备在贮存延寿过程中因失效数据稀缺、样本量不足导致剩余寿命预测精度不高的问题。方法 提出一种基于Transformer的迁移学习和多源数据融合方法。该方法利用多头注意力机制，对装备在贮存、使用和延寿科研等不同阶段获取的多源异构数据进行融合，挖掘数据内在联系，提高信息综合利用水平。在此基础上，引入迁移学习策略，通过在相关领域数据上预训练模型，并采用特征对齐和语义对齐技术，缓解源域和目标域的分布差异，从而提高模型在目标任务上的适应性和判别能力。结果 与传统方法相比，在剩余寿命预测的准确性和实用性方面均取得了显著提升，证明了模型在小样本情况下具有理想的预测精度和鲁棒性。结论 该方法为高可靠性装备剩余寿命的预测提供了一个有效的解决方案，具有重要的实际应用价值。 |
| 英文摘要: |
| The work aims to address the low accuracy in residual life assessments caused by the scarcity of failure data and the small sample size in storage life extension of certain high-reliability aviation equipment. A method utilizing Transformer architecture for transfer learning and multi-source data fusion was proposed. This method effectively integrated sensor data from various stages (such as storage, use and life extension) and types of equipment tests using a multi-head attention mechanism, to explore the internal relation of data and improve the level of comprehensive utilization of information. On this basis, the transfer learning strategy was introduced to mitigate the distribution difference between the source domain and the target domain by pre-training the model on relevant domain data and using feature alignment and semantic alignment techniques, so as to improve the adaptability and discrimination ability of the model on target tasks. Compared with traditional methods, this method significantly improved the accuracy and practicability of the residual life prediction, which proved that the model had ideal prediction accuracy and robustness in the case of small-sample scenarios. This method provides an effective solution for predicting the residual life of high reliability equipment and has important values for practical application. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
