| 摘 要: 针对桥梁监测领域中损伤识别精度较低的问题,提出了一种基于特征工程和深度自动编码器的识别方案。首先采用快速傅里叶变换分析原始数据的特征和规律,其次通过滑动窗口从频谱图中提取表现出损伤差异的模态频率,最后将经过主成分分析法选择的保留损伤信息量最大的敏感特征作为深度自动编码器的输入。实验结果表明,经过特征工程处理后的新指标提高了模型的识别能力和计算效率,在仅占原始数据集14.9%的特征维度的情况下,模型的识别精确率从81.12%提升到98.67%。 |
| 关键词: 损伤识别;特征工程;特征提取;数据降维;深度学习 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 基于人工智能的动态监测系统关键技术开发及应用研究(黔科合支撑[2019]2886);贵州省科技计划资助项目(黔科合平台人才[2019]5802) |
|
| Research on Bridge Damage Identification Based on Feature Engineering and Deep Auto-encoder |
|
HOU Yi1, QIAN Songrong2, LI Xuemei1
|
(1.School of Mechanical Engineering, Guizhou University, Guiyang 550025, China;
2.State Key Laboratory of Public Big Data, Guizhou University, Guiyang 550025, China)
yihou_gzu@163.com; qiansongr_gzu@163.com; lixuemei_gzu@163.com
|
| Abstract: Aiming at the issue of low damage recognition accuracy in the field of bridge monitoring, this paper proposes a recognition solution based on feature engineering and deep auto-encoder. Firstly, the features and patterns of the original data are analyzed using fast Fourier transform. Then, modal frequencies that show the damage differences are extracted from the spectrogram by a sliding window. Finally, the most sensitive features that retain the largest amount of damage information selected by principal component analysis are used as inputs to the deep auto-encoder. Experimental results show that the new indexes processed by feature engineering improve the model 's recognition capability and computational efficiency, and the recognition accuracy of the model improves from 81.12% to 98.67% with only 14.9% of the feature dimensions of the original dataset. |
| Keywords: damage identification; feature engineering; feature extraction; data dimensionality reduction; deep learning |
