Multi-Sensor Attention Networks for Automated Subsurface Delamination Detection in Concrete Bridge Decks

Subsurface delaminations in concrete bridge decks remain undetectable through conventional visual inspection, necessitating automated non-destructive evaluation methods. This work introduces a deep learning framework that integrates Ground Penetrating Radar (GPR) and Infrared Thermography (IRT) through hierarchical attention mechanisms. Our architecture employs temporal self-attention to process GPR electromagnetic signals, spatial attention to analyze thermal imagery, and cross-modal attention with learnable embeddings to model inter-sensor correspondences. We integrate Monte Carlo dropout-based uncertainty quantification, decomposing prediction confidence into model uncertainty and data-driven uncertainty components. Testing across five real-world bridge datasets from the SDNET2021 benchmark reveals that our approach delivers substantial performance gains over single-sensor and concatenation-based baselines when applied to balanced or moderately imbalanced data distributions. Comprehensive ablation analysis confirms that cross-modal attention mechanisms contribute meaningful improvements beyond unimodal attention alone. Critically, we identify and characterize specific failure modes: under extreme class imbalance, attention-based architectures demonstrate susceptibility to majority class bias, indicating scenarios where simpler architectural choices may prove more robust. Our findings equip practitioners with empirically-grounded criteria for selecting appropriate fusion strategies based on dataset characteristics, rather than promoting universal architectural superiority.

💡 Research Summary

This paper presents a novel deep‑learning framework for fully automated detection of subsurface delamination in concrete bridge decks by fusing Ground Penetrating Radar (GPR) and Infrared Thermography (IRT) data through hierarchical attention mechanisms. The authors argue that visual inspection cannot reveal hidden defects and that GPR and IRT offer complementary sensing capabilities: GPR penetrates deeper (10–30 cm) but struggles with shallow anomalies, while IRT excels at detecting shallow delaminations (0.5–10 cm) but is sensitive to environmental conditions. To exploit these complementary strengths, the proposed architecture processes each modality with a dedicated attention module and then integrates them via cross‑modal attention.

Modality‑specific attention

• Temporal self‑attention for GPR: Raw 1‑D A‑scan signals (512 samples) are first passed through a three‑layer Conv1D stack (1→32→64→128 channels) and then fed into an 8‑head self‑attention block. This captures long‑range dependencies in the time series, highlighting critical reflection patterns while being robust to noise.
• Spatial attention for IRT: Thermal images (112 × 112 × 3) are processed by a 2‑D Conv stack (3→32→64→128 channels) followed by batch‑norm and max‑pooling, yielding a 14 × 14 × 128 feature map. Channel‑SE and CBAM modules jointly learn channel‑wise and spatial importance, allowing the network to focus on temperature anomalies that correspond to delamination.

Cross‑modal fusion
Learnable modality embeddings (e_GPR, e_IRT) of size 128 are concatenated with the modality‑specific features and passed through an 8‑head multi‑head attention layer. This cross‑modal attention learns complex correspondences such as strong GPR reflections paired with weak thermal signatures for deep defects, and the opposite for shallow defects. Multi‑head design encourages specialization of heads, improving probabilistic calibration compared with a single‑head alternative.

Uncertainty quantification
Monte Carlo dropout (T = 25) is applied during inference to decompose predictive variance into epistemic (model) uncertainty σ²_e and aleatoric (data) uncertainty σ²_a. The authors use these estimates to route high‑uncertainty predictions to human review, achieving an overall accuracy of 93 % while providing calibrated confidence intervals.

Experimental evaluation
The framework is evaluated on the SDNET2021 benchmark, which contains five real‑world bridge decks with 663 k annotated GPR scans, 4.58 M IRT pixels, and three class labels (intact, shallow delamination, deep delamination). When class imbalance ratios are ≤ 8:1 (minority class ≥ 12 % of samples), the proposed model outperforms single‑sensor baselines (1‑D CNN for GPR, 2‑D CNN for IRT) and simple concatenation fusion by 10–25 % in accuracy and yields AUC improvements up to 0.998. The Park River Median deck achieves 96.6 % accuracy and 99.8 % AUC.

Failure analysis
Under extreme imbalance (> 9:1, minority class < 2 %), the attention‑based architecture exhibits a strong bias toward the majority class, and performance degrades relative to simpler models such as SVM or shallow CNNs. The authors attribute this to insufficient gradient signal for minority samples, causing attention weights to collapse onto dominant patterns.

Ablation studies

• Removing cross‑modal attention reduces accuracy by 6–12 % and lowers AUC, confirming its contribution.
• Replacing multi‑head with single‑head attention harms calibration (higher Expected Calibration Error).
• Disabling MC dropout eliminates uncertainty estimates and slightly reduces accuracy (2–3 %).

Contributions and implications

1. Tailored temporal and spatial attention modules that respect the physical nature of GPR and IRT data.
2. Cross‑modal attention that learns synergistic defect signatures across modalities.
3. Integrated epistemic/aleatoric uncertainty quantification enabling risk‑aware deployment.
4. Empirical guidelines on when attention‑based fusion is advantageous versus when simpler architectures are preferable, based on dataset class distribution.

Future directions suggested include: (a) employing cost‑sensitive learning or synthetic minority oversampling to mitigate extreme imbalance, (b) optimizing the model for edge‑device inference to enable real‑time bridge inspection, (c) extending the framework to incorporate additional NDE modalities (impact‑echo, ultrasonic), and (d) coupling uncertainty‑driven predictions with maintenance decision support systems.

In summary, the paper delivers a comprehensive, experimentally validated multi‑sensor attention network that advances automated, non‑destructive bridge deck inspection, while transparently addressing its limitations and offering practical guidance for field practitioners.