應用深度學習於建物RC牆和磚牆震後損傷程度分級

Abstract

台灣地震頻繁，建物受損風險高，震後快速評估對於安全判定至關重要。隨著顯卡算力提升與人工智慧影像辨識技術的進步，本研究旨在開發影像辨識模型以利用電腦視覺和深度學習技術自動檢測牆體損傷，評估建物安全性，判斷是否需補強或拆除，由此減少工程師的時間與人力成本。本研究透過收集世界各地包含台灣等多個著名地震資料，其中有Datacenterhub網站上來自DEEDS的數個地震資料、來自國家地震工程研究中心資料庫中的2016年美濃地震和2018年花蓮地震資料以及2024年0403花蓮地震資料等，篩選出地震後建物的承重牆的牆體照片。接著，依次將牆體照片按照RC牆和磚牆分類，再將RC依照內政部國土管理署（舊稱營建署）的五級損傷標準分級，將磚牆按照國家地震中心五級損傷標準分級。本研究將採用YOLO(You Only Look Once)一個熱門的電腦視覺模型作為本研究的模型架構，所採用的YOLO版本分別為YOLOv8和YOLO11的分類模型。將種類和級別都分好牆體圖像放入YOLOv8和YOLO11模型中分別進行訓練。以此開發出一個能夠根據RC牆的震後損傷特徵進行損傷等級分級的模型，以及另一個能夠根據磚牆的震後損傷特徵進行損傷等級分類的模型。除此之外，本研究也會使用交叉驗證的方式，將資料樣本切割成複數較小的資料子集，通過將不同的資料子集當初訓練集，而其他資料子集當成驗證集進行交叉驗證，重複進行，便可以交叉驗證。交叉驗證的目的是避免模型出現過擬合和選擇偏差等問題，同時可以反映出模型在各個不同資料集下的性能，從而得出模型的平均性能和最佳性能。本研究也採用Grad-CAM++的技術去對模型的識別能力進行視覺化檢驗。

Taiwan experiences frequent earthquakes, putting buildings at high risk of damage. Post-earthquake rapid assessments are necessary for determining safety. With advancements in GPU computing power and Artificial Intelligence (AI) image recognition technology, this study aims to develop an image recognition model using computer vision and deep learning techniques to automatically detect wall damage, assess building safety, and determine whether reinforcement or demolition is required, in order to reduce the time and labor costs for engineers. This study collects earthquake data from various well-known seismic events worldwide, including those in Taiwan, such as multiple datasets from the DEEDS collection available on the Datacenterhub website, as well as data from the National Center for Research on Earthquake Engineering (NCREE) in Taiwan, specifically the 2016 Meinong earthquake, the 2018 Hualien earthquake, and the 2024 April 3 Hualien earthquake, to systematically filter and analyze wall structure images of load-bearing walls in buildings affected by these earthquakes. Subsequently, the wall images are categorized into RC walls and brick walls, with RC walls classified based on the 5-level damage assessment standard established by the National Land Management Agency (formerly the Construction and Planning Agency) and brick walls classified according to the 5-level damage assessment standard of the NCREE. This study adopts YOLO (You Only Look Once), a popular computer vision model, as the core model architecture by employing the classification models YOLOv8 and YOLOv11. Labeled wall images with predefined categories and damage levels are input into the YOLOv8 and YOLOv11 models for separate training. Through this process, one model is developed to classify post-earthquake damage levels based on the labeled features of RC walls, while another model is designed to classify post-earthquake damage levels according to the labeled features of brick walls. Moreover, this study employs cross-validation by partitioning the dataset into multiple smaller subsets. In each iteration, different subsets are assigned as the training set while the remaining subsets serve as the validation set. By repeating this process, cross-validation is performed systematically. The purpose of employing cross-validation is to reduce the risk of overfitting and selection bias, while also enabling a comprehensive evaluation of the model performance across varying data splits. This approach facilitates the estimation of both the average performance and optimal performance of the model. This study also employs the Grad-CAM++ technique to examine the recognition capability of the model visually.