A semi-active control method for a seismically excited nonlinear benchmark building equipped with a magnetorheological (MR) damper is presented and evaluated. A linear quadratic Gaussian (LQG) controller is designed to estimate the optimal control force. The required voltage for the MR damper to produce the control force estimated by LQG controller is calculated by a neural network predictive control algorithm (NNPC). The LQG controller and the NNPC are linked to control the structure. The coupled LQG and NNPC system are then used to train a semi-active neuro-controller designated as SANC, which produces the necessary control voltage that actuates the MR damper. The effectiveness of the NNPC and SANC is illustrated and verified using simulated response of a 3-story full-scale, nonlinear, seismically excited, benchmark building excited by several historical earthquake records. The semi-active system using the NNPC algorithm is compared with the performances of passive as well as active and clipped optimal control (COC) systems, which are based on the same nominal controller as is used in the NNPC algorithm. The results demonstrate that the SANC algorithm is quite effective in seismic response reduction for wide range of motions from moderate to severe seismic events, compared with the passive systems and performs better than active and COC systems. of the reactive velocity and the issued voltage as described in Equations (2)-(5). The damper velocity is the same as the relative velocity of the floors the damper is connected to. This neural network model is denoted as NNMR ( Figure 5) and is trained ( Figure 6) using input-output data generated analytically using the simulated MR model based on Equations (2)-(5). The NNMR calculates the damper forces based on the current and few previous histories of measured velocity, voltage signals, and damper forces.Training the NNMR requires the compilation of input-output data. To completely identify the underlying MR system model, the data must contain information about the entire operating range of the system. Here, in this study, the velocity and voltage are generated randomly using band limited white Gaussian noise. The generated forces are results of the MR model described in Equations (2)-(5). The sampling rate of the training data was 200 Hz for 30 s period, which resulted in 6000 patterns for training, testing, and validation (Figure 7). Next step is to select the network architecture. To do so, it is required to determine the numbers of inputs, outputs, hidden layers, and nodes in the hidden layers which is usually done by trial and error. The most suitable input data in our case were found to be the current and the four previous histories for Figure 8. Neural network controller (SANC).
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.