Bridges, as essential structures in infrastructure of any country, require timely and efficient maintenance practices to ensure their longevity, structural durability, and safe functionality. Current bridge monitoring and inspection procedures are labor-intensive, and time-consuming in managing a large inventory of the bridges. Bridge owners and local infrastructure authorities often rely on traditional bridge management system (BMS) to maintain bridges. However, traditional BMSs do not provide up-to-date information on the structural condition of the bridges, often resulting in high costs for the bridge owners as bridge maintenance practices are then characterized by reactive rather than predictive procedures. Therefore, emphasizing on BIM technology to effectively maintain and monitor bridges is often recommended. In the domain of BIM, a digital twin (DT) of a bridge is defined as a virtual representation of an existing bridge which continuously provides updated and useful information. While utilizing DT in maintenance of bridges provides numerous advantages to bridge owners, the DTs are not available for many of the existing bridges that were built several decades ago. On the other hand, most of the bridge owners have bridge legacy data in the form of 2D technical drawings. Creating DTs of bridges is a manual and time-consuming process. Therefore, 2D technical drawings of existing bridges could play a role in generating DTs of the bridges. Therefore, in this thesis, the task of utilizing 2D technical drawings in extracting useful information has been tackled with a particular focus on object detection of bridge elements in technical drawings by deep learning and parametric modeling techniques. YOLO, a deep learning model based on CNN, has been trained on technical drawings of bridges which contained 1142 instances for a total of 14 different classes. The model was trained several times to arrive at optimum combination of hyperparameters. Once the well-performing hyperparameters were selected, the model was trained for even a large number of iterations until the mean average precision (mAP) and average loss values converge. It was concluded that YOLO object detector has provided reasonable performance in predicting detections on validation and test dataset with an mAP of 89.15%.     «

Bridges, as essential structures in infrastructure of any country, require timely and efficient maintenance practices to ensure their longevity, structural durability, and safe functionality. Current bridge monitoring and inspection procedures are labor-intensive, and time-consuming in managing a large inventory of the bridges. Bridge owners and local infrastructure authorities often rely on traditional bridge management system (BMS) to maintain bridges. However, traditional BMSs do not provide...     »