Dynamic neural modeling of fatigue crack growth process in ductile alloys
School of Engineering
In this paper, the dynamic neural modeling of fatigue crack growth process in ductile alloys is studied. It is shown that a fatigue crack growth process is treated as a virtual nonlinear dynamic system. A nonlinear model can then be developed with two dynamic neural networks (DNNs), designed to learn the dynamics of crack opening stress and crack length growth, respectively. The DNNs are constructed by adding the tapped-delay-line memories to both the input and the output layers of conventional single layered feed-forward neural networks (SLFNs). Since the delayed output feedback components are placed in parallel with the hidden nodes, a generalized hidden layer is formulated. The DNNs are then trained in the sense that the input weights of the DNNs are uniformly randomly selected in a range, and the generalized output weights are globally optimized with the batch learning type of least squares. The well-trained dynamic neural model is capable of capturing all dynamic characteristics of crack growth process. The excellent performance of the dynamic neural model of fatigue crack growth process is confirmed with the experimental data of 2025-T351 and 7075-T6 aluminum alloy specimens.