基於人工智慧於橋梁結構劣化辨識與量化評估
Other Title
Defect Recognition and Quantification of Bridge Structures Based on Artificial Intelligence
Journal
結構工程
Journal Volume
40
Journal Issue
3
Start Page
1
End Page
23
ISSN
1021-7878
Date Issued
2025-09
Abstract
橋梁為民生交通之重要連結，需定期檢測以確保安全，然而橋梁檢測工作面臨著多樣化和複雜的環境挑戰，特別是跨越區域的橋梁，其結構現況難以全面掌握。本研究提出一套可在全球導航衛星系統（global navigation satellite system, GNSS）受限環境運作的智慧化橋梁檢測流程：以無人飛行載具（unmanned aerial vehicles, UAV）取得高解析影像；結合超寬頻（ultra-wideband, UWB）雙向測距與即時動態定位進行混合定位，於橋下達到約0.1 m精度；透過深度學習模型Mask R-CNN（mask region-based convolutional neural network）自動辨識裂縫、剝落、鋼筋外露與滲水等劣化，模型經實橋影像訓練精確率達0.74、召回率0.83。檢測結果再透過電腦視覺量化並與橋梁檢測方法－DER&U評估連結，配合三維重建技術能將劣化構建重建為點雲模型，以及設定判斷標準，為驗證技術的可靠性和實用性，本研究選取兩座現役橋梁作為示範場域，結果顯示，與傳統的人工目視檢測相比，UAV檢測提供更全面的劣化資訊和更廣泛的檢查角度（如帽梁、主梁）。同時深度學習模型檢測到的劣化位置可以在橋梁評估中清晰記錄，提高檢測的客觀性和可追溯性。本研究提出一套完整的智慧化橋梁檢測流程，從資料獲取、定位技術到損壞評估，均進行深入研究，期能為橋梁檢測領域提供新的可應用方向。
Bridges are critical transportation links requiring regular inspections to ensure safety. However, inspections face challenges in complex environments, especially for large spans where global navigation satellite system (GNSS) signals are weak. With advances in intelligent technology, unmanned aerial vehicles (UAV) combined with deep learning are increasingly applied to bridge inspection, yet existing studies lack a complete, validated workflow for such conditions. This study proposes an integrated intelligent bridge inspection framework. UAV capture high-resolution images of key structural components, while the Mask R-CNN (region-based convolutional neural networks) deep learning model automatically detects and evaluates deterioration. To address weak GNSS signals, ultra-wideband (UWB) and real-time kinematic (RTK) positioning with two-way ranging (TWR) are combined, achieving sub-0.1 m accuracy under bridges. Mask R-CNN, trained on extensive bridge deterioration datasets, reached an accuracy of 0.74 and recall of 0.83, effectively identifying cracks, spalling, exposed rebar, and seepage. Detection results are integrated with the DER&U rating method and 3D reconstruction to generate point cloud models and objective assessment criteria, reducing subjectivity. Two operational bridges were inspected as demonstration sites. Compared with manual visual inspection, the proposed approach provided more comprehensive deterioration data and wider inspection coverage (e.g., cap beams, main beams). The deep learning results improved the clarity, objectivity, and traceability of evaluations. This framework offers a practical and scalable solution for advanced bridge inspection.
Subjects
無人機（UAV）
電腦視覺
深度學習
超寬頻（UWB）
三維重建
unmanned aerial vehicles (UAV)
computer vision
deep learning
ultra-wideband (UWB)
3D reconstruction
Type
journal article