Deep learning-assisted structural health monitoring: acoustic emission analysis and domain adaptation with intelligent fiber optic signal processing

Abstract

Abstract The Acoustic Emission (AE) technique is a well-established method for detecting and locating cracks as they occur. One of the key challenges in fracture mechanics is determining the cracking levels during crack propagation. By linking the AE signature with crack geometry, we can determine the stages of crack growth. The classification of cracking levels is established by correlating micro and macro crack damage to the stable and unstable propagation stages. In this paper, a deep learning architecture based on Convolutional Neural Network is proposed to automatically extract features from the AE signal and classify cracking levels. The performance of this Deep Learning (DL) approach is evaluated using a simulated and experimental Hsu-Nielsen Pencil-led break test conducted in Finite Element Modeling and laboratory setup via COMSOL and a novel High-finesse Fiber Bragg Gratings Fabry-Perot Interferometer sensing system respectively. The results demonstrate the effectiveness of the DL approach for cracking level identification with high accuracy. The dataset was artificially augmented, and a comparison was made between balanced and imbalanced datasets for training purposes. In addition to gaining a deeper understanding of the fracture process, a finite element model was created to differentiate between active and passive AE sources by simulating the Pencil-led break test and crack propagation under cyclic loading conditions respectively. The analysis of the simulation model reveals that the energy associated with AE hits, as calculated from fracture mechanics, increases. This is the primary distinction between active and passive AE sources.