- 學位論文
運用數位孿生與深度學習之端對端鋼筋查驗框架
End-to-End Rebar Inspection Framework using Digital Twins and Deep Learning
摘要
近年來，基於數位孿生解析複雜資料的方法，展現出被運用在營建管理上的潛力。然而，過去數位孿生的研究在鋼筋查驗上仍有諸多限制，例如低現地可行性及低自動化程度。因此，此研究提出一套端對端的鋼筋尺寸性品質管制 (dimensional quality control) 架構，輸入影像後便能直接得到缺失報表。 首先透過運動回復結構 (SfM)、多視角立體視覺 (MVS) 及基於基準點的尺度校正演算法，獲取鋼筋結構的三維模型。接著混和二維實例分割 (instance segmentation) 深度學習模型與三維分群 (clustering) 電腦視覺演算法，辨識出個別鋼筋。後續以隨機採樣一致法 (RANSAC) 預測三維空間中的線，及霍夫變換法 (Hough transform) 預測二維平面上的圓，達成直徑查驗；而分群演算法及分割結果則被用於間距查驗中。最後，建築資訊塑模 (BIM) 技術被整合進此研究。透過現地與模型 (Scan-vs-BIM) 的匹配與比對，尺寸性缺失能被辨識及追蹤。而透過調整現地與模型匹配演算法與開發進度偵測演算法，可一次性查驗多個結構構件。 在實驗室與現地進行的實驗驗證了此框架可被應用於複雜的鋼筋結構。分割模組在實驗柱上達到超過90%的精確率 (precision) 與超過97%的查全率 (recall)。此外，查驗模組在大號數與小號數鋼筋上各取得95.5%與70.8%的直徑分類準確率 (accuracy)；間距查驗方面則是取得0.98%的平均相對誤差 (MRE)。在缺失追蹤模組中，刻意安排的不合格間距被成功地辨識與追蹤。最終，進度偵測演算法的精確率與查全率皆為100%，具備將抽樣檢查擴展為全面查驗的能力。
並列摘要
Recently digital twin (DT)-based approaches that perform analytics on comprehensive data show potential in assisting construction management. However, in terms of reinforcing bar (rebar) inspection, previous research on DT reveals limitations such as impracticality on construction site and low level of automation. Hence, this study proposes an end-to-end framework for rebar dimensional quality control (DQC) that takes images as input and directly outputs issue reports. The combination of structure from motion (SfM), multi-view stereo (MVS), and fiducial-marker-based scale calibration algorithm were leveraged to reconstruct 3D scenes of rebar structures. The rebars within the 3D model were recognized by a hybrid model of deep learning (DL)-based 2D instance segmentation and computer vision (CV)-based 3D clustering algorithm. Then, 3D line random sample consensus (RANSAC) and 2D circle Hough transform were utilized to inspect the diameter; while clustering and segmentation results were utilized to inspect the spacing. Finally, building information modeling (BIM) was incorporated. The dimensional issues were identified and tracked by Scan-vs-BIM registration and comparison. To inspect multiple structural components at once, Scan-vs-BIM registration was modified and progress detection was explored. Experimental results on lab and on-site structures show that the framework is promising to be applied to complex rebar structures. The segmentation model reached a precision of over 90% and recall of over 97% on the lab column. Besides, the diameter inspection module achieved a classification accuracy of 95.5% and 70.8% for large-size and small-size rebars. A mean relative error (MRE) of 0.98% was reported during the spacing inspection. In issue tracking, the purposely defective spacing was identified and traced. Finally, both precision and recall reached 100% in the progress detection, which extends spot check to total inspection.
參考文獻
延伸閱讀
- 陳翊翔、莊仕杰、張鈞程、羅昱恆、黃琮煒、邱永全、林冠成、黃志民、周頌安、陳俊杉（2021）。以深度學習與數位孿生打造工地鋼筋查驗新法。土木水利，48(2)，15-21。https://doi.org/10.6653/MoCICHE.202104_48(2).0003
- 林湘霖、王貞淑（2015）。結合模糊層級分析法與關聯法則建構數位學習系統活動流程評估準則。管理與系統，22(1)，125-147。https://www.airitilibrary.com/Article/Detail?DocID=10239863-201501-201502100055-201502100055-125-147
- 陳楚云（2022）。基於數值分析與深度學習之感測器佈設最佳化與結構健康監測方法建立〔碩士論文，國立臺灣大學〕。華藝線上圖書館。https://doi.org/10.6342/NTU202203189
- Hu, C. Y. (2020). Applying Deep Neural Network for STR-based DNA Mixture Interpretation [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU202003142
- Yang, T. Y. (2019). Deep Feature Learning for Fast and Efficient Regression Application [doctoral dissertation, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU201902521