Enhancing time-frequency representations via gradient-based attention: The GIFT methodology for structural damage detection
Structures, cilt.92, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 92
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.istruc.2026.112859
- Dergi Adı: Structures
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
- Anahtar Kelimeler: Damage detection, Deep learning, Explainable artificial intelligence, Gradient-based attention, Structural health monitoring
- Eskişehir Osmangazi Üniversitesi Adresli: Evet
Özet
Converting vibration signals into time–frequency (TF) representations for deep learning (DL)-based damage detection has become a common approach in structural health monitoring (SHM). However, TF images often contain substantial background noise that obscures damage-sensitive features. Although explainable artificial intelligence (XAI) techniques are widely used for post-hoc interpretation in SHM, their use as active preprocessing mechanisms represents an emerging research direction. Building upon the emerging Explanation-Guided Learning (EGL) paradigm, this study introduces the Gradient-Informed Feature Transformation (GIFT) framework, which repurposes input-gradient-based saliency maps as an input-level feature transformation mechanism rather than using them solely for post-hoc interpretation. Using a guidance network, GIFT extracts spatial attention from raw TF images via input-derivative sensitivity analysis, suppressing noise while enhancing damage-related patterns prior to classification. The framework was evaluated across five deep learning architectures on two benchmark datasets (IASC-ASCE building and Z24 bridge) using both Continuous Wavelet Transform (CWT) and Short-Time Fourier Transform (STFT) representations. Under STFT representation, average accuracy increased from 87.59% to 98.33% on ASCE and from 76.03% to 95.61% on Z24. Statistical significance was confirmed across all evaluated experimental scenarios (p < 0.05). Overall results demonstrate that GIFT substantially improves classification performance, offering a favorable performance trade-off for practical SHM applications.