260 likes | 268 Views
This symposium presentation proposes a semi-active neuro-control system using MR dampers for a base-isolated benchmark problem in structural dynamics and vibration control. The proposed method involves training a neural network to minimize a cost function, eliminating the need for presetting desired responses. The system offers high adaptability, no requirement for large power supplies, and comparable performance to active base isolation systems.
E N D
The 18th KKCNN Symposium 18-20, Dec., 2005 Semi-active Neuro-Control Using MR Damper for Base-Isolated Benchmark Problem Heon-Jae Lee*: Ph.D. Candidate, KAIST, Korea Hyung-Jo Jung: Professor, Sejong University, Korea Woo-Hyun Yun: Professor, Kyungwon University, Korea In-Won Lee: Professor, KAIST, Korea
OUTLINE • Introduction • Benchmark Problem • Proposed Method • Training Neuro-Control System • Performance Evaluation • Conclusion Structural Dynamics & Vibration Control Lab., KAIST, Korea
Kodiak, Alaska San Francisco Introduction • Base Isolation • One of the most widely implemented seismic protection system • Mitigates the effects of an earthquake by isolating the structure from ground motion Structural Dynamics & Vibration Control Lab., KAIST, Korea
Lead Rubber Bearing Friction Pendulum Bearing Introduction • Nonlinear Devices for Base Isolation • Benefit • Restoring force and adequate damping capacity can be obtained in one device • Drawbacks • Strongly nonlinear • Poor adaptability for a wide range of ground motion • Especially strong impulsive ground motions generated at near-source location Structural Dynamics & Vibration Control Lab., KAIST, Korea
Introduction • Active Base Isolation System • Consisting of a base isolation system combined with actively controlled actuators • Advantages • High performance in reducing vibration • High adaptability for different ground excitation • Ability to control of multiple vibration modes • Drawbacks • Requirement of a large external power supply • Active systems may have the risk of instability Structural Dynamics & Vibration Control Lab., KAIST, Korea
Introduction • Semi-Active Base Isolation System • Base isolation system that employs semi-active control devices (e.g. MR dampers, controllable friction dampers) • Similar high adaptability to the active system • No requirement of large power supplies • In this presentation, semi-active neuro-controller with base isolation system will be proposed. Structural Dynamics & Vibration Control Lab., KAIST, Korea
Benchmark Problem • Purpose • To provide systematic and standardized means • By making direct comparisons between competing control strategies, including devices, algorithms, sensors, etc. • To allow researchers in structural control to test their algorithms and devices • This benchmark problem is about a smart base isolation system. Structural Dynamics & Vibration Control Lab., KAIST, Korea
Benchmark Problem • Benchmark Structure • A base-isolated eight-story, steel-braced framed building • Similar to existing building in Los Angeles, California • In this investigation, only the linear elastomeric isolation system which consists of 92 bearings is considered. • Modeled using three master degrees of freedom per floor • 16 MR dampers (8 along the x-axis and 8 along the y-axis) are also installed Structural Dynamics & Vibration Control Lab., KAIST, Korea
Benchmark Problem • Sample Earthquakes • Both the fault-normal (FN) and fault-parallel (FP) components of Newhall, Sylmar, El Centro, Rinaldi, Kobe, Ji-ji, Erzinkan • Control Cases • Case I : x-direction (FP), y-direction (FN) • Case II: x-direction (FN), y-direction (FP) Structural Dynamics & Vibration Control Lab., KAIST, Korea
STRUCTURE Clipped Algorithm MR Damper Proposed Method Neural Network Clipped Neuro-Controller Block diagram of the proposed method Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Structure of the neural network Output layer Input layer Hidden layer Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Conventional Neuro-Controller • In training neural network, inputs and outputs are needed • Inputs: structural responses and ground acceleration • Outputs: control force • In these neuro-controllers, presetting the desired structural response or control force is needed. • Presetting the desired response or control force now becomes another difficulties. Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Improved Neuro-Controller • New training algorithm (Kim et al., 2000, 2001) • The neuro-controller is trained by minimizing a cost function • If the neuro-controller is trained by minimizing the cost function, there is no need to predetermine the desired response. where, is specific state, is control signal and are weighting matrices Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Improved Neuro-Controller • The update of weights and biases between the output layer and hidden layer can be simply expressed as where, :learning rate :sensitivity :unit vector :activation function of the output layer Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Improved Neuro-Controller • The update of those between the hidden layer and input layer can be obtained as where, Structural Dynamics & Vibration Control Lab., KAIST, Korea
Proposed Method • Clipped Algorithm • Desired force (by neural network): • Generated force (by MR damper): Structural Dynamics & Vibration Control Lab., KAIST, Korea
where, Training Neuro-Controller • Training Data • Filtered artificial earthquake record • Magnitude: scaled to match the maximum acceleration of the given earthquakes • Shaping filter Ground acceleration (m/sec2) Time (sec) Structural Dynamics & Vibration Control Lab., KAIST, Korea
Training Neuro-Controller • Structure of Neuro-Controller • Two neuro-controllers are employed for x and y-directions, respectively. Structure of each direction’s neuro-controller Structural Dynamics & Vibration Control Lab., KAIST, Korea
Training Neuro-Controller • Cost Function • In cost function, are included. • Optimal weighting matrices Structural Dynamics & Vibration Control Lab., KAIST, Korea
Training Neuro-Controller • Cost function vs. epoch • The cost function converges in both x and y-directions, which means the training is successful. x-direction Cost function y-direction epoch Structural Dynamics & Vibration Control Lab., KAIST, Korea
Performance Evaluation • Evaluation Criteria • Based on both maximum and RMS responses of the building J1 - Peak base shear J2 - Peak structure shear J3 - Peak base displacement J4 - Peak inter-story drift J5 - Peak absolute floor acceleration J6 - Peak generated force J7 - RMS base displacement J8 - RMS absolute floor acceleration J9 - Total energy absorbed Structural Dynamics & Vibration Control Lab., KAIST, Korea
Performance Evaluation • Performance Comparison (Case I) Sylmar El Centro Newhall Rinaldi Kobe Jiji Erzinkan Passive (Vmax) Clipped optimal Proposed Structural Dynamics & Vibration Control Lab., KAIST, Korea
Performance Evaluation • Performance Comparison (Case II) Sylmar El Centro Newhall Rinaldi Kobe Jiji Erzinkan Passive (Vmax) Clipped optimal Proposed Structural Dynamics & Vibration Control Lab., KAIST, Korea
Performance Evaluation • Corner drift (normalized by uncontrolled values) Structural Dynamics & Vibration Control Lab., KAIST, Korea
Conclusion • A new semi-active control strategy for seismic response reduction using neuro-controller and MR dampers is proposed. • The proposed strategy was applied to a benchmark building installed with linear elastomeric isolation system. • In numerical simulation results, the proposed strategy can significantly reduce the floor acceleration, base shear and building corner drift with only a slight increase of the base displacement. Structural Dynamics & Vibration Control Lab., KAIST, Korea
Thank you for your attention!!! Structural Dynamics & Vibration Control Lab., KAIST, Korea