Study On The Theory Of Structural Damage Identification Based On Neural Network And Data Fusion

Because of the incompletion of measured data and the existence of noise, the application of system identification method and damage index method is restricted in the field of structural damage identification. The problem can be solved by taking structural damage identification as a pattern recognition issue and using neural network as a classifier. But with the multi-sensor system widely used on the large and complex structures, how to use the various measured data effectively becomes a key problem. In order to solve the problem of neural network inputs selection, the use strategies of information and too many positions to be diagnosed in large and complex structures, the theories of data fusion, evidence, fuzzy integral, substructure and clustering were adopted in the thesis. The following studies have been done in this thesis:Firstly, the theories of data fusion and neural network have been deeply studied. Two decision level fusion methods and two neural network models have been compared respectively. The applicable conditions of each method or model have been gained.Secondly, a further study has been made on the selection of inputs for neural networks used in structural damage identification. Three rules have been put forward: the function of neural network, the class separability and the noise influence on the class separability. The numerical example proves the validity of the rules, and the rules are instructive to the inputs selection of neural networks used in structural damage identification.Thirdly, the two-steps identification method has been adopted to increase the recognition rate of some elements in the feature level data fusion. By using sub-nets, the influence of other mode types on the modification of weights has been decreased. The numerical example shows that the two-steps identification method can increase the recognition rate of effectively. Fourthly, the decision level data fusion of neural networks in structural damage identification has been studied. The results of D-S evidence theory and fuzzy integral have been compared. The numerical example shows that the effect of D-S evidence theory is "the minority subordinates to the majority", while the fuzzy integral theory can take the objective estimation of each evidence and the importance of each evidence into consideration at the same time.Fifthly, the computation method of fuzzy densities has been studied. Three kinds of computation method being used currently have been compared. The creditability of the neural network identification result was taken as the reasonable fuzzy density.Lastly, the general sub-structure method has been put forward to solve the large and complex structural damage identification problem. The general sub-structure method can overcome the shortcoming of subjectivity in sub-structure division. The numerical example shows the validity and advantage of the method. By combining the general sub-structure method and the multi-steps method, the multi-step general sub-structure method: is presented. A cable-stayed bridge example shows that the method can effectively decrease the difficulty of solving the problem and make it possible to identify the damage location in large and complex structures exactly.