A data-driven approach based on long short-term memory and hidden Markov model for crack propagation prediction

Abstract

We present in this paper a combined technique of long short-term memory and hidden Markov model to prediction problems of crack propagation in engineering. The primary advantage of the hidden Markov model is that the ability to learn with less information, in other words, its future states do not depend on past ones, based only on the present state. We use long short-term memory to train data, and output consequences improved by adding predicted different changes that are computed by hidden Markov model. Applying this combined method to numerical examples of forecasting crack propagation of singled-edge-notched beam forced by 4-point shear, crack-height growth in Marcellus shale under the hydraulic fracturing and deformations of dam structures made from fiber reinforced concrete material is addressed. The tests were carried out with many different sizes of experimental data. It was found that a combined long short-term memory - hidden Markov model results in more accurate solution than only using long short-term memory, especially in the case of the dataset that is lack of information.