An investigation of the low-cycle fatigue parameters and estimation using artificial neural networks

Traditional and non-traditional prediction methods are used to estimate low-cycle fatigue (LCF) parameters and fatigue lives. Artificial neural networks (ANN) are the most widely used method for estimating the LCF parameters and fatigue lives. In this thesis, the LCF parameters and fatigue lives were estimated by two different ANN models and at the same time the effects of ANN structure on the estimation results were analyzed. According to the obtained results, LCF parameters of high-strength steels were estimated at over 99.99% and fatigue lives were estimated at over 98.7% estimation accuracy. For one hidden layer structure, the best activation function was found to be logistic sigmoid (logsig), epoch number 100, training function Levenberg-Marquardt (trainlm) and hidden neuron range 5-20. With Model 2, LCF parameters and transition fatigue lives of various steels were estimated using one, two, and three hidden layer ANN structures. The most optimal activation functions for the output layers were hyperbolic tangent sigmoid (tansig) and linear (purelin). On the other hand, logsig almost never gave good results in the output layer. In hidden layers, tansig gave better results than logsig and purelin. The best combination for three hidden layered structure was found to be tansig-tansig-tansig-tansig/purelin and 10-15 hidden neuron range for each hidden layer, tansig-logsig-purelin/tansig and 10-15 hidden neuron range for two hidden layers structure, and purelin-tansig and 1-5 hidden neuron range for one hidden layer structure. All estimation parameters were estimated at 94.4%/1.685% for one hidden layer structure, 93.4%/1.389% for two hidden layer structure, and 97.1%/0.873% for three hidden layer structure with estimation accuracy/mean absolute percentage error rates.