Abstract :
In this study a method for identification of crack in variable cross section beam is presented. Theprocess of crack identification in the suggested method is consists of four steps. In first step, three natural frequencies of a variable cross section beam for different locations and depths of cracks were obtained using Finite Element Method (FEM). For applying FEM, ANSYS commercial Software was used. The FEM results were verified by data of one of available references. It was concluded that the FEM analysis has been done accurately. In second step, two Multi-Layer Feed Forward (MLFF) neural networks were created. In third step, Particle Swarm Optimization (PSO) method was coded in MATLAB commercial software. Then PSO was used to training the neural networks. In another word, the weights of MLFF neural networks were calculatedby minimizing the root mean square of differences between neural network outputs and targets. The inputs of neural networks were first three natural frequencies and the outputs of first and second neural networks were corresponding locations and depths of cracks, respectively. The all of data were applied to neural networks in normalized form. In forth step, some of natural frequencies of variable cross section beam with distinct crack conditions as inputs applied to trained neural networks. Obtained results displayed that cracks characteristics were in good agreements with actual data. Finally it was concluded that the proposed procedure in suitable for crack detection in variable cross section beams. Also it was found that the trained Artificial Neural Networks (ANNs) predicted crack location better than crack depth.
