基于ICEEMDAN与1D-CAE的无监督斜拉索异常识别研究
Research on Unsupervised Abnormality Identification of Stay Cables Based on ICEEMDAN and 1D-CAE
摘要
针对复杂环境下斜拉索振动信号的强非平稳特性及传统方法对损伤标签的依赖问题,采用融合改进的自适应噪声完备集合经验模态分解(ICEEMDAN)与一维卷积自编码器(1D-CAE)的无监督异常识别方法,研究斜拉索早期隐蔽异常识别问题.结果表明:1)ICEEMDAN 分解精度优于 CEEMDAN,重构误差降低 4.6%,信噪比提升 0.22 dB,正交性指标改善 11%;2)1D-CAE 模型泛化能力优良,训练集、验证集重构误差符合正态分布;3)该组合方法对模拟典型异常识别准确率超过 95%,无需损伤先验标签,可适应实桥复杂环境,实现斜拉索早期隐蔽异常的有效识别.
Aiming at the strong non-stationary characteristics of stay cable vibration signals in complex environments and the dependence of traditional methods on damage labels,an unsupervised anomaly identification method fusing Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(ICEEMDAN)and One-Dimensional Convolutional Autoencoder(1D-CAE)was adopted to study the problem of early hidden anomaly identification of stay cables.The results show that:1)The decomposition accuracy of ICEEMDAN is superior to that of CEEMDAN,with the reconstruction error reduced by 4.6%,the signal-to-noise ratio is improved by 0.22 dB,and the orthogonality index is improved by 11%;2)The 1D-CAE model has excellent generalization ability,and the reconstruction errors of the training set and validation set conform to the normal distribution;3)The identification accuracy of the proposed hybrid method for simulated typical anomalies exceeds 95%,requires no prior damage labels,adapts to complex in-service bridge environments,and can effectively identify early concealed anomalies of stay cables.
年,卷(期):
2026, 42(4)
分类号:
U443.38
机标分类号:
TH133.33
TN911.4
TM74
在线出版日期:
2026年9月17日
基金项目:
CQJT-CZKJ2025-01:重庆市交通科技项目,