Temporal Risk Forecasting Of Infrastructural Health Using Deep Learning With Attention Mechanism
Abstract
This paper introduces a unified deep learning framework for automated damage assessment and temporal risk forecasting in civil infrastructure, addressing both crack and corrosion deterioration. A Spectral Mixed Attention (SMA) Network is developed for pixel-level semantic segmentation by combining convolutional feature extraction, mixed attention, and a global context bottleneck to identify fine crack patterns and multi-class corrosion states under diverse surface conditions. To improve training stability, the framework incorporates an singular value decomposition (SVD)-based reparameterized optimization strategy that adjusts gradient updates in the spectral domain using reflector-based orthogonal factors, helping maintain directional consistency and robustness under challenging optimization conditions. Following segmentation, structural descriptors are extracted from the predicted damage regions: crack dimensions are quantified using skeletonization and distance transforms, while corrosion severity is evaluated using class-wise area estimation aligned with inspection-based condition states. These descriptors are arranged into temporal sequences and used by an LSTM-based forecasting model to predict Time-to-Attention (TTA). Experimental results show that the proposed SMA segmentation network achieves 71.2% MIoU on DeepCrack, corresponding to an 8.37% relative improvement over the best baseline MIoU, while the proposed learning strategy improves U-Net Dice and IoU by 7.86% and 12.76%, respectively, compared with the baseline.
Recommended Citation
I. Y. Sheikh and S. Jagannathan, "Temporal Risk Forecasting Of Infrastructural Health Using Deep Learning With Attention Mechanism," IEEE Transactions on Artificial Intelligence, Institute of Electrical and Electronics Engineers, Jan 2026.
The definitive version is available at https://doi.org/10.1109/TAI.2026.3713437
Department(s)
Electrical and Computer Engineering
Second Department
Computer Science
Keywords and Phrases
Deep Learning; Infrastructural Health; LSTM; Mixed Attention; SVD optimizer; Time-to-Attention
International Standard Serial Number (ISSN)
2691-4581
Document Type
Article - Journal
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2026 Institute of Electrical and Electronics Engineers, All rights reserved.
Publication Date
01 Jan 2026
Comments
Intelligent Systems Center, Grant 69A3552348339