Date of Award

Spring 2026

Abstract

Bridges are critical components of transportation infrastructure that face continuous degradation due to aging, fatigue, environmental stresses, and natural hazards. Traditional inspection methods, primarily visual assessments, are labor-intensive, subjective, and often fail to detect internal or hidden damage. Deep learning-based Structural Health Monitoring offers a promising alternative for automated, continuous bridge condition assessment. However, a fundamental deployment challenge persists: real-world bridges rarely have labeled damaged data available for training, while models trained on numerical simulations suffer from severe performance degradation when applied to field conditions due to domain shift arising from modeling simplifications, operational variability, and measurement uncertainty. This dissertation addresses this challenge by developing and validating transfer-learning-based deep learning frameworks that enable reliable bridge damage detection using limited or entirely unavailable damaged field data. The research systematically progresses from supervised classification with full labeled data toward operationally realistic deployment scenarios requiring minimal or healthy-only field measurements. In this dissertation, the term Artificial Intelligence “AI” is used specifically to refer to data-driven deep learning models, primarily convolutional neural networks (CNNs), rather than general artificial intelligence systems.

The first study establishes the foundational deep learning methodology by introducing a novel two-dimensional Convolutional Neural Network architecture for damage classification using strain time-history data from a full-scale bridge deck mock-up. Validated under real commercial vehicle loads across four controlled damage states, including a visually undetectable damage level, the model achieved over 95 percent classification accuracy and demonstrated robustness across varying sensor configurations and elevated noise levels.

The second study investigates transfer learning strategies for generalizing damage detection across bridge structures with varying geometry. Using Convolutional Neural Network models trained on finite element simulation data from simply-supported beam structures, the study reveals the challenge of catastrophic forgetting in sequential feature extraction and demonstrates that joint training across multiple simulated domains produces geometry-invariant feature representations capable of maintaining high accuracy on unseen structural configurations. These findings establish the methodological foundation for the simulation-to-field transfer learning framework developed in subsequent chapters.

The third study presents the first known implementation of one-dimensional Convolutional Neural Networks for indirect, drive-by bridge health monitoring validated on a full-scale, in-service reinforced concrete bridge in Nebraska. Using acceleration data collected from a truck-mounted sensor array under four progressive cumulative damage states at multiple vehicle speeds, the model achieved 93 percent testing accuracy by processing raw time-series data directly, without handcrafted feature extraction or model-based calibration.

The fourth study, representing the primary technical contribution of this dissertation, addresses the critical simulation-to-field domain shift through an uncertainty-aware transfer learning framework. A one-dimensional Convolutional Neural Network was jointly pre-trained on synthetic strain data from three deliberately uncalibrated finite element models of a steel truss bridge and then adapted to field measurements from a full-scale bridge with physically induced damage states. Direct application of the simulation-trained model to field data yielded only 54.8 percent accuracy, which represents the conservative baseline and lower-bound performance metric before transfer learning improvements, quantifying the domain shift. Through supervised fine-tuning with limited labeled field data, classification accuracy improved to 95.8 percent, demonstrating that reliable three-class damage classification can be achieved without resource-intensive model calibration. A systematic data efficiency analysis across training percentages from 10 to 90 percent revealed that approximately 30 percent of field data suffices for acceptable performance, while the model exhibited a desirable fail-safe behavior at low data levels by conservatively predicting damage rather than missing structural deterioration.

The fifth study extends the framework to the operationally realistic scenario in which no labeled damaged field data is available. Building upon the simulation-to-field transfer learning framework developed in the previous study, two complementary D0-only adaptation strategies are developed and evaluated under identical cross-validation settings. The first, D0-only fine-tuning, adapts the fully connected classification head using only healthy field data and identifies damage through a reduction in healthy-class confidence. The second is a statistical detection framework that uses frozen CNN features, principal component analysis, and Mahalanobis distance, with particular emphasis on the fully unsupervised Mode 4. At the P90 operating point, Mode 4 attains 78.2% balanced accuracy (AUC = 0.837), while D0-only fine-tuning achieves 76.0% balanced accuracy (AUC = 0.798); both methods substantially outperform direct transfer without adaptation, which reaches only 54.8%. These strategies play complementary roles: Mode 4 offers a robust first-stage screening approach with no need for model adaptation, whereas D0-only fine-tuning provides a classifier-based alternative that connects naturally to the supervised framework presented in Chapter IV.

Collectively, this dissertation establishes a hierarchical framework for deploying deep learning in operational bridge health monitoring: from immediate anomaly screening using only healthy-state data, through progressively refined classification as labeled data becomes available, to detailed severity assessment with moderate labeled data. This continuum provides bridge owners and engineers with a practical, scalable pathway for AI-based structural health monitoring across diverse data availability scenarios.

Document Type

Dissertation

First Advisor

Yashar Eftekhar Azam

Second Advisor

Erin Bell

Third Advisor

Masoud Sanayei

Department or Program

Civil Engineering

Degree Name

Doctor of Philosophy

Download

Available for download on Wednesday, June 09, 2027

Share

COinS