Fatigue crack size evaluation using acoustic emission signals for wire and arc additive manufactured material

Abstract

As one of the promising additive remanufacturing techniques, the wire and arc additive manufacturing has shown great potential in automated repair and restoration. However, the layer-by-layer fabrication process may lead to an uneven heating and cooling rates, which has a significant effect on the microstructure and mechanical behavior, as compared to conventionally processed materials. In this study, the fatigue performance of the low-carbon steel components made by wire and arc additive manufacturing is studied systematically. The components made by conventional hot-rolled method is used for comparison purposes. By leveraging the advanced causal network and graph neural networks techniques, a novel method for online fatigue crack size evaluations is proposed using acoustic emission signals. Low dimensional latent variables are extracted from the acoustic emission signal through a one-dimensional convolutional autoencoder (1D-CAE), and the data-driven causal discovery algorithm is adopted to construct the causal network of these latent variables, which characterizes their inherent cause-effect relations. The graph attention network (GAT) is adopted to aggregate a node information with its parent nodes as the node embedding in a causal network. Fully connected layers are constructed between the node embeddings and fatigue crack size label. For online scenarios where limited data is available, the above GAT model previously trained is transferred with different strategies for crack size evaluation. Experimental data with three different scenarios: specimens with same direction, specimens with different direction and hot-rolled specimens are employed to validate the proposed online fatigue crack size evaluation method.