Concrete crack evolution analysis employing convolutional neural networks and acoustic emissions

Analyzing the distribution, types, and evolution of concrete cracks is vital for preventing structural failure. This study utilized Brazilian discs with central cracks (BDCC) specimens, which effectively provide data on tension, tension-shear, and shear cracks under mixed-mode loading. Crack activity in BDCC specimens was monitored using acoustic emission (AE) techniques. Crack sources were located through a three-dimensional (3D) time difference of arrival (TDoA) method. To distinguish between tension and shear cracks and accurately locate them, this paper integrates Gramian angular summation field (GASF) and convolutional neural network (CNN) algorithms into AE signal analysis, introducing the GASF-CNN crack classification model. Based on the load-time curves and AE parameter characteristics, the failure process of the BDCC specimens is categorized into damage accumulation and crack propagation stages. Proportion analysis reveals that the proportion of shear cracks in BDCC specimens gradually increases as the loading mode transitions from pure mode I to mode II; however, tensile cracks remain predominant. This is because the loading mode primarily affects stress distribution near the crack tip, while tensile stress governs regions farther from the crack tip. Crack distribution results show that tensile cracks display a random, disordered spatial distribution throughout the specimen, whereas shear cracks are mainly localized between the central crack tip and the loading ends. The proposed GASF–CNN crack classification model merges one-dimensional time-series signals with two-dimensional image data, enabling highly precise identification of crack types.