Model Updating with Neural Networks and Genetic Optimization

In dynamic analysis of structures, the accuracy of the mathematical model plays a crucial role. However, because of several uncertainties like local nonlinearities, welding points, bolted joints, material properties and geometric tolerances, the mathematical model will contain differences compared to the manufactured product. Hence, it is essential to update mathematical models by using vibration test data taken from the structure. This paper presents a new approach to model updating via utilizing neural networks and genetic optimization algorithms. The key point in this new approach is that the model updating capability of neural networks is improved by a genetic optimization algorithm by guiding the optimization problem with results obtained from neural network identification. Employing the nominal mathematical model created for a particular structure, a data set of selected mode shapes and natural frequencies is created by a number of simulations performed by perturbing selected updating parameters randomly. A neural network is then created and trained with this data set. Upon training the network, it is used to update the initial model with the test data. The results are then improved further by using the “network updated mathematical model” as an initial model and updating it again by employing a genetic optimization algorithm. The most important advantage of the proposed approach is the possibility of using different number of degrees of freedom for each mode shape; as a result, additional flexibility is introduced to the approach, since the proposed method can be used with incomplete test data. The application and capabilities of the proposed approach is illustrated via real test data taken from a GARTEUR test bed, where it is seen that the proposed method updates mathematical models associated with such complex structures efficiently.