SE-CNN-Based Damage Detection in Welded Joints of Spatial Grid Structures

Abstract

Owing to welding process-related factors and environmental influences, stress concentration is prone to occur in the welded regions of grid structure joints, rendering these joints the most damage-susceptible locations within the entire structural system. In this paper, an intelligent damage diagnosis framework that integrates deep learning techniques with joint stress field analysis is proposed to enable damage diagnosis of structural joints. A quasidistributed fiber Bragg grating (FBG) sensor network deployed at the joints facilitated the construction of a dual-dimensional damage index, characterized by high sensitivity to damage location and severity. The proposed index enables precise identification of joint damage by capturing the abrupt changes in the local stress field caused by joint failure. By leveraging the spatial distribution patterns of the damage index, a convolutional neural network (CNN) model incorporating the squeeze-and-excitation (SE) attention mechanism was developed, with Bayesian optimization algorithms employed for hyperparameter tuning to perform simultaneous localization and quantification of joint damage. Moreover, the model exhibits robust noise tolerance, retaining 93.8% identification accuracy at a signal-to-noise ratio (SNR) above 10 dB without requiring preprocessing denoising. To validate the feasibility of the proposed SE-CNN model, a spatial grid structure model with adjustable damage joints and controllable damage levels was designed and fabricated, and identification tests were conducted under various joint damage conditions. The damage diagnosis accuracy for joint damage in the spatial grid structure reached 97.2% using the proposed method with damage indices derived from stress data measured by quasidistributed FBG sensors. Therefore, the SE-CNN model established based on the characteristics of the damage index proves to be an effective method for joint damage diagnosis in spatial grid structures.