Prediction of fatigue–crack growth with neural network-based increment learning scheme

An increment learning scheme based on fully-connected neural network is proposed to predict the fatigue–crack growth in middle tension, M(T), specimens of 7B04 T6 aluminum and TA15 titanium alloy under constant amplitude stress. Usually, the measurement of fatigue–crack growth rates is labor-intensive and time-consuming, and the dataset of fatigue–crack growth is small. Neural networks with back-propagation algorithm are not good at training on small dataset. Here we design network inputs which employ multiple increment information to overcome this shortage. Given the first part data points of crack growth in a specimen, the trained network can predict the rest for both aluminum alloy and titanium alloy without any prior knowledge. The trained network learns the underlying rules in experimental data of crack growth. Our method shows superiority to conventional fitting formulas and common neural networks such as recurrent neural network and long short-term memory method. Our work demonstrates the capacity of neural network and provides an alternative method to predict fatigue–crack growth.