Damage Sensitive PCA-FRF Feature in Unsupervised Machine Learning for Damage Detection of Plate-Like Structures
Pei Yi Siow, Zhi Chao Ong, Shin Yee Khoo, Kok‐Sing Lim
Abstract
Damage detection is important in maintaining the integrity and safety of structures. The vibration-based Structural Health Monitoring (SHM) methods have been explored and applied extensively by researchers due to its non-destructive manner. The damage sensitivity of features used can significantly affect the accuracy of the vibration-based damage identification methods. The Frequency Response Function (FRF) was used as a damage sensitive feature in several works due to its rich yet compact representation of dynamic properties of a structure. However, utilizing the full size of FRFs in damage assessment requires high processing and computational time. A novel reduction technique using Principal Component Analysis (PCA) and peak detection on raw FRFs is proposed to extract the main damage sensitive feature while maintaining the dynamic characteristics. A rectangular Perspex plate with ground supports, simulating an automobile, was used for damage assessment. The damage sensitivity of the extracted feature, i.e. PCA-FRF is then evaluated using unsupervised [Formula: see text]-means clustering results. The proposed method is found to exaggerate the shift of damaged data from undamaged data and improve the repeatability of the PCA-FRF. The PCA-FRF feature is shown to have higher damage sensitivity compared to the raw FRFs, in which it yielded well-clustered results even for low damage conditions.