ESTIMATION OF STRESS INTENSITY FACTORS FOR CRACKED FINITE PLATES USING A HYBRID MODEL OF ARTIFICIAL NEURAL NETWORK AND FINITE ELEMENT METHOD (ANSYS)
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
The aim of this study was to develop an artificial neural network model in order to estimate stress intensity factor values of two different types of semi elliptical surface cracked plates which have no explicit stress intensity factor formula. One of these plates contains one semi elliptical surface crack at both sides, front side and back side. The other one contains two parallel semi elliptical surface cracks at the front and back side. First of all, data which were needed for neural network model training were generated in aid of finite element method (Ansys). In the first case (one semi elliptical surface crack at both sides), 179 Ansys simulations were done in total using different values of minor radius (a), the ratio of minor radius to major radius (a/c) and the ratio of minor radius to plate thickness (a/t). As a result of these simulations in case 1, 8234 data were generated (46 different parametric angles for each simulation) and 1061 of these data were used to train the artificial neural network model. Besides, 523 Ansys simulations were done using different values of a, a/c, a/t and h (vertical distance between two parallel cracks) for the second case (two parallel semi elliptical surface cracks at both sides). 24058 data were obtained after these simulations and 4248 of them were utilized for the network training process. Matlab neural network module (nntool) was used for artificial neural network modelling. Different types of network structures were used in the training process. The accuracy of the trained neural networks for the first and second case were tested using 760 and 1139 new data respectively, which were generated via Ansys so as to determine the best network model. Consequently, minimum deviation value (difference between Ansys and Matlab neural network result) of case 1 was obtained for the network model that has 2 hidden layers and 15 neurons for each hidden layer and minimum deviation was calculated as 0.32%. Similarly, minimum deviation value of the model for the second case was calculated as 0.49% and this model has 3 hidden layers and 14 neurons for each hidden layer. As a result of this study, for two different types of cases, two artificial neural network models which have the potential to estimate stress intensity factor values without doing any simulations, were obtained.