-
摘要:
针对航空发动机轴承故障诊断过程中预测精度不足以及过拟合的问题，提出基于迁移学习的半监督集成学习器（SSIT）用以发动机轴承故障预测。首先，训练改进的基于迁移学习的极限学习机（TELM）以及基于迁移学习的支持向量机（TSVM），分别迁移不同目标空间的高相似度样本加入到源样本空间进行训练。然后，与对应的基学习器集成同簇学习器来识别未标记样本，构成半监督学习器不断调整初始基学习器权重，并继续集成半监督基学习器的识别结果到SSIT中。通过此学习机识别提取特征后的，用以故障识别。实验结果清楚地表明：此种方法可以有效降低迁移学习中的负迁移效果，提升迁移精度10%左右，降低机器学习中的过拟合效果，提高半监督学习稳定性，与现有的预测方法相比可以提高精度9%以上。
Abstract:Aimed at the problems of insufficient prediction accuracy and over-fitting in the fault diagnosis process of aero-engine bearing, a semi-supervised integrated learning device based on transfer learning (SSIT) is proposed to predict engine bearing fault. First, transfer learning based improved extreme learning machine (TELM) and support vector machines (TSVM) were trained by adding the high-similarity sample of different target space to the original sample space, which is integrated to identify the tag sample with the corresponding learning. Then integrate the same cluster learner with the corresponding base learner to identify the unlabeled samples, Next, the constituted semi-supervised learning device constantly adjusts the initial learning unit weight, and continues to integrate semi-supervised learning recognition results into SSIT, which will be used to identify faults after feature identification and extraction by this learning machine. The experimental results clearly show that this algorithm can effectively reduce the negative transfer effect in transfer learning, improve the transfer accuracy by about 10%, reduce the over-fitting effect in machine learning, and improve the stability of semi-supervised learning. Compared with the existing prediction method, this algorithm can improve the accuracy by more than 9%.
-
Key words:
- aevo-engine /
- fault diagnosis /
- semi-supervision /
- transfer learning /
- integrated learning
-
图 12 发动机转子实验平台内部结构
Figure 12. Internal structure of engine rotor experimental platform
表 1 数据集属性
Table 1. Data set properties
数据集 字母 样本数 字母数 L A~E 100 5 U A~E 400 5 L1-L6 A~E 30 5 U1-U3 A~E 120 5 D1 F，G，H 400 3 D2 I，G，K 400 3 D3 L，M，N 400 3 合计 A~N 1850 29 
下载: 导出CSV
表 2 不同数据集各方法预测精度对比
Table 2. Comparison of prediction accuracy of different methods on different data sets
% 测试集 SSIT CNN TELM SL 测试集1 90.5 81.3 68 69.7 测试集2 90.6 82.1 68 69.6 测试集3 90.6 82.2 72 68.5 测试集4 90.5 79.2 71 70.4 测试集5 90.6 79.5 73 69.5 平均值 90.6 80.9 70 69.5
下载: 导出CSV
表 3 最终迁移样本百分比
Table 3. Percentage of final migrated sample
% 测试集 D1 D2 D3 测试集1 63.2 40.1 22.3 测试集2 69.4 43.9 20.3 测试集3 66.7 39.6 25.1 平均值 66.4 41.2 22.6
下载: 导出CSV
-
[1] 房晓南.基于半监督和集成学习的不平衡数据特征选择和分类[D].济南: 山东师范大学, 2016. http://cdmd.cnki.com.cn/Article/CDMD-10445-1016086762.htmFANG X N.Unbalanced data feature selection and classification based on semi-supervised and integrated learning[D]. Jinan: Shandong Normal University, 2016(in Chinese). http://cdmd.cnki.com.cn/Article/CDMD-10445-1016086762.htm [2] 杨印卫.基于异态集成学习的刀具状态监测技术研究[D].天津: 天津大学, 2014.YANG Y W.Research on tool condition monitoring technology based on alien integration learning[D]. Tianjin: Tianjin University, 2014(in Chinese). [3] 张伟.基于卷积神经网络的轴承故障诊断算法研究[D].哈尔滨: 哈尔滨工业大学, 2017. http://cdmd.cnki.com.cn/Article/CDMD-10213-1017864225.htmZHANG W.Research on bearing fault diagnosis algorithm based on convolutional neural network[D]. Harbin: Harbin Institute of Technology, 2017(in Chinese). http://cdmd.cnki.com.cn/Article/CDMD-10213-1017864225.htm [4] 郭勇.基于单源及多源的迁移学习方法研究[D].西安: 西安电子科技大学, 2013. http://cdmd.cnki.com.cn/Article/CDMD-10701-1013304243.htmGUO Y.Research on transfer learning method based on single source and multiple sources[D]. Xi'an: Xidian University, 2013(in Chinese). http://cdmd.cnki.com.cn/Article/CDMD-10701-1013304243.htm [5] ZHOU Z H, LI M.Tri-training:Exploiting unlable data using three classfiers[J]. IEEE Transactation on Knowledge and Data Engineering, 2005, 17(11):1529-1541. doi: 10.1109/TKDE.2005.186 [6] 卞则康, 王士同.基于相似度学习的多源迁移算法[J].控制与决策, 2017, 32(11):1942-1948. http://d.old.wanfangdata.com.cn/Periodical/kzyjc201711003BIAN Z K, WANG S T.Multi-source transfer algorithm based on similarity learning[J]. Control and Decision, 2017, 32(11):1942-1948(in Chinese). http://d.old.wanfangdata.com.cn/Periodical/kzyjc201711003 [7] 谭建平.基于半监督的SVM迁移学习文本分类方法[D].广州: 广东工业大学, 2016. http://cdmd.cnki.com.cn/Article/CDMD-11845-1016139239.htmTAN J P.A semi-supervised SVM transfer learning text classification method[D]. Guangzhou: Guangdong University of Technology, 2016(in Chinese). http://cdmd.cnki.com.cn/Article/CDMD-11845-1016139239.htm [8] 张倩, 李海港, 李明.基于多源动态TrAdaBoost的实例迁移学习方法[J].中国矿业大学学报, 2014, 43(4):713-720. http://d.old.wanfangdata.com.cn/Periodical/zgkydxxb201404023ZHANG Q, LI H G, LI M.An example transfer learning method based on multi-source dynamic TrAdaBoost[J]. Journal of China University of Mining & Technology, 2014, 43(4):713-720(in Chinese). http://d.old.wanfangdata.com.cn/Periodical/zgkydxxb201404023 [9] PAN J, WANG X S, CHENG Y H, et al.Multi-source transfer ELM-based Q learning[J]. Neurocomputing, 2014, 137:57-64. doi: 10.1016/j.neucom.2013.04.045 [10] 王雪松, 潘杰, 程玉虎, 知识迁移学习方法及应用[M].北京:科学出版社, 2016.WANG X S, PAN J, CHENG Y H.Knowledge transfer learning method and application[M]. Beijing:Science Press, 2016(in Chinese). [11] CHENG Y H, WANG X S.Multi source tri-training transfer learning[J]. Transactions on Information and System, 2014, 97(6):1668-1672. http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=J-STAGE_2198442 [12] 邓万宇, 屈玉涛.基于ELM-AE的迁移学习算法[J].计算机与数字工程, 2018, 46(5):854-860. http://d.old.wanfangdata.com.cn/Periodical/jsjyszgc201805002DENG W Y, QU Y T.Transfer learning algorithm based on ELM-AE[J]. Computer and Digital Engineering, 2018, 46(5):854-860(in Chinese). http://d.old.wanfangdata.com.cn/Periodical/jsjyszgc201805002 [13] 莫志军, 基于复杂网络的航空发动机故障传播特性研究[D].长沙: 湖南科技大学, 2016. http://cdmd.cnki.com.cn/Article/CDMD-10534-1016792678.htmMO Z J.Research on fault propagation characteristics of aeroengine based on complex network[D]. Changsha: Hunan University of Science and Technology, 2016(in Chinese). http://cdmd.cnki.com.cn/Article/CDMD-10534-1016792678.htm [14] 郭艳平, 龙涛元.振动信号模型在滚动轴承故障诊断中的应用[J].机械设计与制造, 2018, 30(17):270-272. http://d.old.wanfangdata.com.cn/Periodical/jxsjyzz201801078GUO Y P, LONG T Y.Application of vibration signal model in fault diagnosis of rolling bearings[J]. Machinery Design & Manufacture, 2018, 30(17):270-272(in Chinese). http://d.old.wanfangdata.com.cn/Periodical/jxsjyzz201801078 [15] LIU X B, LIU Z T.Ensemble transfer learning algorithm[J]. IEEE Access, 2018, 32(6):2389-2396. http://d.old.wanfangdata.com.cn/Periodical/jsjgcyyy201012037 [16] IQBAL M S, LUO B, KHAN T.Heterogeneous transfer learning techniques for machine learning[J]. Iran Journal of Computer Science, 2018, 1(1):31-46. doi: 10.1007/s42044-017-0004-z [17] BLUM A, MITCHELL T.Combining labeled and unlabeled data with co-training[D]. Pittsburgh: Carnegie Mellon University Madison, 1998: 92-100. -
下载:
点击查看大图
计量
- 文章访问数: 1269
- HTML全文浏览量: 182
- PDF下载量: 621
- 被引次数: 0
