Non-linear Principal Component Analysis of Response Spectra

ABSTRACT

The present work aims at exploring the application of nonlinear principal component analysis in dimensionality reduction and prediction of response spectra. The evaluation is performed based on log10 scaled response spectra at 91 spectral periods corresponding to 13552 records available in the NGA-West2 database. The non-linear principal component analysis performed on the data showed that 91 spectral periods can be addressed with just 3 principal components. Further, an artificial neural network (ANN) model is developed to predict these three principal components with magnitude, distance, shear wave velocity and focal mechanism as input. The inter- and intra-event residuals obtained for the response spectra predicted using the developed model are comparable with the existing ground motion prediction equations (GMPEs) from the same database. The developed model is also observed to capture all the prominent attenuation features of ground motions. Hence, the study indicates that the response spectra can be described with just three uncorrelated variables.

KEYWORDS:

• Non-Linear Principal Component Analysis
• Auto-Associative Network
• Response Spectra
• Dimensionality Reduction

Acknowledgments

We thank the reviewer for his careful evaluation of our work which helped in improving the presentation of the inferences from the study. We would also like to acknowledge the PEER group for the efforts made by the team in processing and compiling the various ground motion databases and providing open access through https://ngawest2.berkeley.edu/