Font Size: 

Frequency response function (FRF) estimation in ambient vibration analysis for vibration based damage detection in large-scale structures has been proven effective due to no information leakage during data extraction as compared to modal domain data. One of the most explored computation approaches in frequency response function (FRF) estimation is artificial neural networks (ANNs) which are proven effective in damage detection based on its capability to relate the nonlinear relationship between vibration parameters and damage location and severity. To this regard, ANN is chosen as a medium to interpret damage detection model from frequency domain data. However, the main concerns in using FRFs for damage detection is the size of the FRF data. A full-size of FRF data will result in wide configuration range of the ANN input layer. Thus, it will affect the iteration divergence in training and the computational inefficiency. Therefore, data reduction method is used to reduce the size of the FRFs and to filter the measurement noise. The most common method used is principal component analysis (PCA). Based on PCA, several other reduction techniques has been developed including proper orthogonal decomposition (POD), singular value decomposition (SVD), eigenvalue decomposition (EVD) and discrete Karhunen–Loève theorem (KLT). Therefore, this study compares stated data reduction methods on PCA using non-probabilistic ANN for damage detection. The input data for the network are the compressed FRFs and Young’s modulus (E values) which acting as the elemental stiffness parameter (ESP) are used as the output. To establish the relationship between the input parameters and output parameters, Possibility of Damage Existence (PoDE) is designed for the undamaged and damaged states. Stiffness Reduction Factor (SRF) is also used in order to represent the changes of the stiffness parameter. To compare the results obtain from different data reduction method, damage measure index (DMI) is used as the assessment medium. A numerical of two-storey steel frame is used as an example.