Spectral element simulation data-driven acoustic emission damage location in orthotropic steel decks

Data-driven acoustic emission (AE) damage location methods yield promising performances in large-scale complex structures like orthotropic steel decks (OSDs), but rely on abundant training data that are generally obtained by pencil lead break (PLB) tests. A new data-driven AE location method based on spectral element simulation and machine learning is proposed for more efficient practical applications. Especially, a CPU-GPU heterogeneous parallel computing framework is developed for three-dimensional time-domain spectral element method (SEM) simulation. It helps to generate high-quality numerical AE waves without excessive computational resources for training the artificial neural network (ANN)-based location model. Through the experiment on a full-scale OSD model, the method was proved to achieve an accuracy significantly higher than the widely-used time of arrival (TOA) method and comparable to the traditional data-driven method with experimental input data. The innovation lied in obviating burdensome PLB tests to collect training data for the AE location machine learning model.