Structural Damage Identification Based On Neural Network

The detection, location, and estimation problem of structural damage has been the subject of much current research in recent years. The damage identification methods have been widely used in aeronautical, civil, mechanical and nuclear structures. It is polytechnic method constructed on the damage theory, sensor technology, signal analysis, computer science and artificial networks. Relative to the traditional technology of structural damage identification, this paper mainly studies the method based on the combination of the static modal analysis and artificial networks.In this paper, it verified theoretically the combined-parameters that adapts to locate structural damage location and confirm damage degree. (The parameters consist of natural frequencies and mode shape data at a few selected points) based the theory, numerically simulated a three-story frame and a cantilever beam respectively. And by appropriate means form the input parameters of the improved BP algorithm neural network . Detect the structural damage by the trained networks.The paper includes the following contents:Firstly, through analysis of principle with neural network.the paper get that the combined-parameter can identify structural damage location and damage degree.Secondly, structural physics parameter (natural frequencies > modal parameters etc) is function of structural parameters(mass, stiffness ) . That is, the change of physical parameters induce the change of structural vibration. Therefore, through structural modal analysis with different damage degree using ANSYS, get natural frequencies and mode shape data as neural network input vector after unitary. By comparing the numerical results with theoretic result, to affirm if neural network can numerically identify structural damage.Lastly, through a three-layer frame and a cantilever beam testify that the method can very accurately identify structural damage location and degree. Yet with little damage the result is relatively low. To resolve the problem, augmenting node of hidden layer or number of hidden layer. At the same time, adding the dimension of input parameters.The result shows the effectiveness of this means.