基於圖神經網路之鷹架結構完整性檢測

Abstract

鷹架為工地最常採用之一種施工設備，提供作業人員通行或在其上從事高處作業。每單元鷹架由立架、踏板和交叉拉桿所組成，且在搭建過程中需採取扶手先行工法以保證作業人員安全。在檢查鷹架完整性的過程中，需要確保元件被安裝在正確的位置上。然而，作業人員時常會拆除元件以方便作業，此行為有可能會造成鷹架結構不穩定或者產生容易墜落的空隙。在搭建鷹架的過程中，作業人員也會為了作業方便而忽略安裝扶手。目前工地的安全檢查多倚賴人工巡檢，耗時、費力之外，也難以即時檢查到鷹架單元元件細節或完整的安裝過程。本研究提出一種結合圖神經網路(Graph Neural Network)、關鍵點偵測(Keypoint Detection)和電腦視覺(Computer Vision)的集合方法，以檢查鷹架的完整性以及搭建順序，協助巡檢過程更自動化。具體來說，本研究首先利用圖像分割(Image Segmentation)來確認施工架立架、踏板和扶手之位置，接著建立一個多模態模型(Multimodal)，此模型結合深度學習的關鍵點偵測、傳統電腦視覺和深度預測的特徵來訓練，用於確認交叉拉桿之完整性。再來，利用圖神經網路分析圖像中元件的空間關係，確認鷹架搭建結構是否符合規範。最後，設計一個符合台灣鷹架搭建法規的邏輯判斷，以確認扶手先行之搭建過程和搭建結果是否有墜落的可能性。本研究經工地實際案例測試後，驗證提出之方法可成功辨識不完整的施工架單元，也能夠分析搭建鷹架過程是否符合規範。實驗結果顯示，二維的圖像分割精度在踏板、立架、扶手中，分別達到96.2%、90.6%、86.7%，且在交叉拉桿偵測模型中關鍵點精度達到97.8%。在圖神經網路之驗證達到85.2%。這些成果能夠證實所提出之方法可有效辨識鷹架之完整性以及搭建過程的安全性。

Scaffolding is a common construction equipment, which provides workers with access and a platform for high-altitude tasks. Each unit comprises uprights, footboards, and cross braces, with a handrail-first approach necessitating the installation of handrails during scaffolding erection process. During the process of inspecting the scaffolding's integrity, it is necessary to ensure that the components are installed in the correct positions. However, workers often remove components to facilitate their tasks, which can potentially cause the scaffolding structure to become unstable or create gaps that are prone to falling. Additionally, during the scaffolding construction process, workers may neglect to install handrails for the sake of convenience. Currently, safety inspections on construction sites rely heavily on manual checks, which are time-consuming and labor-intensive and fail to instantly capture the details of scaffold unit components or the completeness of the installation process. This study proposes an integrated approach combining Graph Neural Network (GNN), keypoint detection, and computer vision to inspect the integrity and assembly order of scaffolding, aiding in the automation of inspection processes. Initially, Image Segmentation identifies upright, handrail and footboard positions. A multimodal model then combines deep learning keypoint detection, computer vision and depth estimation features to verify cross brace integrity. GNN analyze component spatial relationships to ensure standard compliance. Finally, a logic assessment that aligns with scaffolding construction regulations in Taiwan is designed to confirm whether the handrails-first assembly process and results present any fall hazards. Tested on real sites, the method successfully identified incomplete units and checked assembly compliance. Image segmentation for footboards, uprights and handrails showed accuracies of 96.2%, 90.6% and 86.7%, respectively. Keypoint accuracy for cross brace detection reached 97.8%. GNN validation was 85.2%, proving the method's effectiveness in identifying scaffolding integrity and erection safety.