doi:10.3850/978-981-08-7619-7_P035

ABSTRACT

Trained lattice pattern probabilistic neural network (LPPNN) was proposed to reduce the structural responses under earthquakes more accurately and efficiently. The lattice form of the trained LPPNN was composed of the control force and state vector, and trained by an iterative process using the gradient descent method under earthquakes. The trained LPPNN calculates the control force using the adjacent information of input only on lattice pattern, thus its corresponding calculation process is fast. In order to get the control force, calculations of distances between input and the adjacent patterns, and their weights are essential. In this study, the Euclidian distance and the probability density function were used for them, respectively. Three-story building under El Centro earthquake was trained for numerical verification. The maximum displacement and velocity of the structure were used to construct the corresponding lattice pattern limits. The mean value and standard deviation of the responses were used for the probability density function. California and Northridge earthquakes were used to prove the performance of the proposed method. In the numerical simulation, the responses by the trained LPPNN showed that the proposed algorithm can effectively reduce more than a lattice pattern control method the response of structures under earthquakes.

Keywords: Active control, Vibration, Earthquake, Training, Gradient descent method, State space.

© 2011 RPS