PSI - Issue 52

Yuhang Pan et al. / Procedia Structural Integrity 52 (2024) 699–708 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

703

5

Fig. 4. The eigenvalue and the cumulative percentage obtained from PCA: (a) All of points, (b) First thirty-five points.

As shown in Fig.4, the cumulative contribution percentage of the first nine principal components is more than 70.14%, which indicates that these nine principal components represent most of the information of FRFs in the range 500-2000H. Therefore, these nine principal components are selected as input features for the vibration-based method in this study. 3.2. Model development After obtaining the features, the next step is to find a model to develop the relationship between these features and the structure ’s health. Assessing the state of a structure as damaged or undamaged represents a standard binary classification problem. Consequently, neural pattern recognition techniques are employed for detecting structural damage. Moreover, to facilitate damage localization, a set of extracted features is provided as input, and the output is the location of the damage, including both x and y coordinates. Therefore, a Backpropagation Neural Network (BPNN) is employed for damage localization process (Pan et al. 2020). The developed model is described in Fig. 5.

Fig. 5. The developed model for the damage detection and localization.