抄録
Floor response spectrum is essential for the seismic design of nonstructural components in buildings. Standardized floor response spectrum enables to evaluate seismic design forces without time history analysis. However, this approach is not often applicable to structures with a wide variety of damping ratio. This paper proposes a novel evaluation method of floor response spectra based on machine learning with neural networks. The proposed method transforms target response spectra into the corresponding floor response spectra with dynamic characteristics of buildings and nonstructural components.
Both buildings and nonstructural components are assumed to be linear elastic. Simulated ground motions generated from a specified acceleration response spectrum are employed. Time history analysis creates floor response spectra for training data regarding various natural periods and damping ratio.
Firstly, buildings with a single degree of freedom are assumed to design the architecture of a neural network. A simple feedforward network underestimates the dynamic amplification factor for systems with low damping ratio. For floor response spectrum in these buildings, we propose a neural network with two subnetworks. One is a network evaluating resonance characteristics. The input values to this subnet are damping ratio and natural frequency. Another is a network representing response characteristics over a wide range of natural period. The two subnetworks are joined each other before the output layer. The neural network is trained with backpropagation algorithm.
Parametric studies investigate the adequate number of the hidden layers and the units. As a result of the investigation, the subnetwork regarding resonance characteristics should have more hidden layers compared to the other subnetwork. With the best neural network, the predicted dynamic amplification factors coincide with theoretical values in a wide variety of damping factor and natural period. The predicted floor response spectrum obtained from the neural network also has good agreement with the results from time history analysis.
Next, the framework of the evaluation method of the floor response spectra is extended to cover multi-degree of freedom systems. This paper considers three vibration modes at most. A proper way of the modal combination is discussed regarding floor response spectra. Through the discussion, the neural network is extended to consider multiple modal responses.
Finally, numerical examples are demonstrated for steel buildings having five or ten stories. Predicted floor response spectra have the same shape as the ground truth on any floor. The neural network infers quite instantly, and this advantage is useful to implementation of interactive software.
