A data-driven approach for crack damage detection in ship plate structures utilizing stress field

Abstract

Cracks are prevalent and safety-critical defects in ship hull structures, rendering real-time and accurate detection essential for the assurance of structural integrity. However, traditional crack detection methods are constrained by inefficiency, high costs, and vulnerability to human error. This study aims to develop a crack detection method based on stress field data and convolutional neural networks (CNNs) to improve detection accuracy and efficiency. The extended finite element method (XFEM) was utilized to simulate crack propagation and generate a comprehensive stress field dataset, whereas actual stress measurements were acquired via sensors. A sliding window technique was employed to preprocess the stress sequences, segmenting continuous data into ordered subsets. Subsequently, a multi-layer CNN model was developed to automatically identify crack types, lengths, and locations. Experimental results indicate that the proposed method achieves high accuracy across various crack features, with classification accuracies of 100 % for crack length and position, and over 96 % for crack angle, thereby validating its effectiveness and robustness. The findings suggest that this data-driven detection approach is applicable to ship hull plates, offering an efficient and reliable tool for crack diagnosis in marine engineering.