The relationships among the static and dynamic parameters of a civil structure are normally highly non-linear. Neural network provides a new method for models of a dynamic system, model modification and damage diagnosis. However, due to some problems like noise and testing error of input data, and sensitivity of model parameters and input parameters, further application of neural network in the fields of model modification and damage identification of structures is limited. In this paper, an intelligent model modification method is discussed which uses the neural network model based on selected parameters and sample selection. It is suggested that mode frequencies, mode shapes and mode flexibility are appropriate as input parameters of neural network modification. A test design optimum method is used for selecting samples, with special attention to the redundancy questions of the samples. When training the network, an optimum method is used which depends on the principles of convergence. It is also used to optimize the hidden nodes and to establish the topological structure of the network. The LM algorithm is used to train the network. It is also used for identification and modification. After comparing with the different modification accuracy of input parameters, numerical simulation for model modification of a cantilever beam in the Badong Yangtze River Bridge proves that the structural model can be modified effectively with the mode flexibility as inputs, the model modification method is effective.
