Intelligent Recognition and Life Prediction of Bridge Fatigue Cracks Based on Multimodal Visual Fusion and Spatiotemporal Convolutional Networks

Read full text Summarise with VIRAT
Authors: Jianzhen Wu, Qi Yang, Lifeng Song, Qu Wang
Cited by: 0
References: 19

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Bridges, as critical components of transportation infrastructure, require early identification of fatigue cracks and accurate life prediction to ensure structural safety.However, existing detection methods suffer from low efficiency, insufficient feature representation in singlemodal data, and inadequate modeling of spatiotemporal feature coupling.To address these limitations, this study proposes an intelligent bridge fatigue crack recognition model based on multimodal visual fusion and spatiotemporal convolutional networks.The model innovatively designs a feature transformation unit that embeds a Graph Convolutional Network (GCN) into a Long Short-Term Memory (LSTM) network, enabling the collaborative extraction of spatial topology and temporal dependencies from crack evolution sequences.A dual-stream architecture with independent 3D Convolutional Networks (C3D) is constructed to process spatiotemporal features from RGB and depth video modalities, respectively.During the feature fusion stage, a "cascade fusion-high-dimensional mapping" strategy is employed to enhance the abstract representation capability of multimodal features.Finally, classification results are produced through a Softmax layer, and supervised training is conducted using a multi-class cross-entropy loss function.To evaluate the performance of the proposed model, ablation studies, comparative experiments, and feature fusion analyses are conducted, using both crack recognition and life prediction as evaluation metrics.Experimental results demonstrate that the proposed model achieves a crack recognition accuracy of 96.2% and an F1-score of 95.6%, significantly outperforming conventional approaches.The Mean Absolute Error (MAE) of life prediction is only 28.3 days, representing a reduction of 34.4 days compared with the traditional Paris-law-based method.Moreover, the model exhibits strong robustness in micro-crack detection and under challenging conditions such as varying illumination and occlusion, validating the advantages of the integrated "recognition-prediction" framework.This research provides an efficient and accurate technical solution for bridge fatigue crack detection and promotes the intelligent development of structural health monitoring.

Fields of Study & Subjects

Infrastructure Maintenance and Monitoring Structural Health Monitoring Techniques Fatigue and fracture mechanics

Publication details

Journal / source
Traitement du signal
Volume
43
Issue
1
Pages
159–170
Publication date
28 February 2026
Language
EN