| 魏能,周立成,张舸*,刘泽佳,刘逸平,蒋震宇,杨宝,汤立群.基于移动时间窗和时空窗主成分分析组合特征的损伤识别方法[J].实验力学,2023,38(1):17~27 |
| 基于移动时间窗和时空窗主成分分析组合特征的损伤识别方法 |
| Damage identification method based oncombined features by moving time window and spatiotemporal-window principal component analysis |
| 投稿时间：2022-04-07 修订日期：2022-05-10 |
| DOI：10.7520/1001-4888-22-087 |
| 中文关键词: 移动主成分分析 时空窗主成分分析 组合特征 机器学习 损伤识别 |
| 英文关键词:moving principal component analysis double-window principal component analysis combined feature machine learning damage identification |
| 基金项目:国家自然科学基金(11972162)；中国博士后科学基金(2021M700886)资助 |
|
| 摘要点击次数: 849 |
| 全文下载次数: 95 |
| 中文摘要: |
| 近年来，机器学习算法在结构损伤识别领域得到了广泛应用。已有研究表明，设计良好的输入特征能够适应多种算法，从而减少调参时间和计算成本，并提高算法的识别准确率。本文引入移动主成分分析和时空窗主成分分析方法提取损伤敏感的特征组合作为机器学习输入，使用双跨平面梁有限元模型生成的模拟应变数据对组合特征的损伤识别效果进行验证。结果表明，良好的特征向量组合对损伤更加敏感，不仅有助于识别在传感器周围发生的损伤，而且能有效地定位结构边缘和远离传感器位置处的损伤，从而提高算法的损伤识别效果。此外，多种噪声强度下的实验表明，组合特征具有良好的鲁棒性，能较好地适应外界环境的变化。 |
| 英文摘要: |
| Nowadays, machine learning algorithms have been widely applied in structural damage identification. Previous research has proved that machine learning algorithms with damage-sensitive features as inputs have better performance in damage detection, as compared to original responses as input. It is also proven that damage-sensitive features can be adaptive to various machine learning algorithms with satisfying performance. Besides, damage-sensitive features are capable to reduce the time consumption of parameter adjustment as well as computational cost. In this paper, moving principal component analysis (MPCA) and double-window principal component analysis (DWPCA) methods are applied to extract damage-sensitive features as inputs for machine learning algorithms. The simulated results based on a two-span planar beam model show that machine learning algorithms with combined features which are extracted by MPCA and DWPCA methods have the best performance among various kinds of inputs. The methods are demonstrated to be capable to detect the damages located close to the sensors. Additionally, it can localize damage at the edges of the structure and the locations far away from the sensor effectively. Finally, experiments under various noise intensities show that the combined features have good robustness. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
