Research On Damage Identification Of Truss Structure Based On Neural Network

Damage identification technology is a hot issue of difficulty in current civil engineering area. The correlative theory and technologies which are researched by many scholars are on development. According to the dynamic characteristic difference between a prototype structure and a damage structure, utilizing robustness, fault toleration and generalization ability of neural network, taking the plane truss and the three-dimensional truss as an example, three different neural networks are applied to evaluate the localization and quantification of the structure damage. The main research contents include:(1) A series of damage indexes which are based on modal parameters of structure before and after damage are derived, and the relationship between damage indexes and the position of damage or the degree of damage are studied. Then the basis is provided for the input parameter selection of damage identification at different stage.(2) The sub-structure is introduced to make the damage identification of large-scale complex structure into possible. This method keeps clear of network no convergence when putting all signals into the neural network for one time. The damage sub-structure is subdivided into the different sub-structure on the basis of identifying the damage element; the scope of damage is narrowed.(3) An artificial neural networks method is applied for multiple steps damage identification, truss structure damage identification is divided into three steps. Firstly, the sub-structure of damaged member is determined by probabilistic neural network with inputting the change rate of standardized modal frequency and the change ratio of standardized modal frequency separately, and the two indicators outputs are compared. Then the member of damaged is also determined by radial basis function neural network with inputting the normalized damage signal index and the combination damage index separately, and analysis is done to compare the two indicators outputs. Finally, the damage degree is determined by radial basis function neural network with input of the change rate of squared modal frequency. Using three steps method to valuate the localization and quantification of truss damage can effectively reduced network data-in scale, also reduced the computation quantity, and reduced the network training time, and then the network recognition precision can be improved.(4) The method using partial modes to replace the integrity modes to identify the position of damage is introduced. It is confirmed validity by taking a complex three-dimensional truss as the example, which has significant instruction for the complex structure damage recognition.