Development of an intelligent CNN-LSTM-attention model for acoustic emission-based fracture detection and structural health monitoring in marine steel structures
Jialin Cui, Chunwang Lv, Xianqiang Qu, Jinbo Du, Hanxu Wang
Abstract
Structural Health Monitoring (SHM) of steel structures under cyclic loading is critical for ensuring structural integrity and safety. This study proposes an optimized CNN-LSTM-Attention deep learning framework for real-time and high-precision damage identification using Acoustic Emission (AE) signals. The model integrates CNN for spatial feature extraction, LSTM for temporal dependency modeling, and an Attention mechanism to enhance feature selection, effectively capturing the complex evolution of Crack Initiation, Crack Growth, and Crack Fracture. To improve model generalization and robustness, Gaussian noise-based data augmentation and optimized AE signal segmentation were employed. Experimental results demonstrate that the CNN-LSTM-Attention model achieves 98.5 % classification accuracy, significantly outperforming conventional CNN, RNN, and LSTM-based architectures. Furthermore, the study highlights the impact of signal segmentation length on classification performance, revealing that T = 8 provides the best balance between local feature extraction and global temporal modeling. The proposed framework offers an intelligent, scalable, and highly accurate solution for real-world structural health monitoring, reducing misclassification risks and enhancing early damage detection in steel structures.