STUDY ON NEURAL NETWORK BASED ON RANDOM SEISMIC RESPONSE HISTORY

Abstract

 In the previous study, focusing on the self-organizing ability of the neural network, it was shown that it is possible to model the nonlinear historical restoring force characteristics without using existing mathematical models when the load-displacement relationship obtained by giving a regular wave that gradually increase and decrease as a forced displacement is used as the training data. Therefore, in this study, we performed dynamic analysis using more complicated and random artificial waves as acceleration data. It was confirmed that the same modeling is possible by using the load-displacement relationship obtained in this way as training data. We also examined a method that improves estimation accuracy by using 10 unlearned artificial waves and 12 designed seismic motions as verification waves and performing dynamic analysis via a trained neural network.