为了提高现有方法对结构微小损伤的识别准确率,本文基于移动主成分分析(Moving Principal Component Analysis,MPCA)方法提出了 一种新型损伤敏感特征——多阶PCA(Principal Component Analysis,PCA)损伤敏感特征.本文将这一新型敏感特征作为机器学习的输入,预测受损位置和损伤程度.首先,对监测数据做MPCA分析,并基于方差累积率确定特征向量阶数;其次,通过对各阶特征向量进行内积运算,获得多阶敏感特征(Multi-Order Sensi-tive Features,MOSF);最后,将多阶敏感特征作为机器学习算法的输入,对结构开展损伤位置和损伤程度的损伤识别研究.基于双跨连续梁的数值实验结果发现:相比于直接采用传统的PCA特征向量作为损伤特征,本文所提出的新型敏感特征对损伤具有更好的灵敏度.结合新型敏感特征与机器学习的损伤识别方法,在损伤位置和损伤程度任务中均具备较高的准确率和稳健的鲁棒性.即使在信噪比为10dB的高强度噪声环境下,本文所提出的损伤识别方法在损伤位置和损伤程度识别任务中也具备高于80％的预测准确率.
A novel intelligent damage identification method utilizing sensitive features extracted by moving principal component analysis
Novel damage-sensitive features,termed Multi-Order Principal Component Analysis(PCA)features,have been developed based on Moving Principal Component Analysis(MPCA)to enhance the accuracy of existing methods for identifying minor structural damage.These features were used as inputs for machine learning algorithms to predict both the location and severity of damage.Firstly,the monitoring data was analyzed using MPCA,and the optimal number of feature vector orders was determined by the cumulative variance ratio.Next,Multi-Order Sensitive Features(MOSF)were generated through inner product operations on these feature vectors.These MOSF were then used as inputs for machine learning algorithms in studies focused on structural damage identification,specifically targeting damage localization and severity assessment.Based on the results of numerical experiments conducted on a two-span continuous beam,we found that the proposed sensitive features exhibit enhanced sensitivity to damage.When combined with machine learning-based damage identification methods,these features demonstrate high accuracy and robust performance in both damage localization and severity assessment.Even in high-noise environments,with a signal-to-noise ratio of 10dB,the proposed method maintains a prediction accuracy exceeding 80％for both damage localization and severity assessment.
structural health monitoringmoving principal component analysisdamage sensitive featuredamage identificationmachine learning
万方数据
维普
