Sonar combines deep learning and building information modeling for underwater crack detection of concrete structures

Timely detection and positioning of cracks is an important task for concrete structure operation management. This paper presents a rapid underwater crack detection-localization system combining deep learning (DL) and building information modeling (BIM) based on forward-looking multibeam sonar and remotely operated vehicle (ROV). Due to the non-orthogonal characteristics of the sonar, the crack clarity is closely related to the sonar attitude. To improve the accuracy of crack shooting, a sonar image simulation platform is built. The objective function of sonar attitude-detection accuracy is established, and the sparrow search algorithm is introduced to preset the optimal sonar attitude. Because the cracks have complex feature in sonar images, a multi-scale feature extraction module and an attention mechanism are introduced to improve YOLOV9c to increase the detection accuracy. The results show that the improved algorithm outperforms other algorithms and has high robustness. Subsequently, to achieve the localization of the detected damage and map to the corresponding location, the coordinate transformation relationship between the real-world and the sonar images is derived. Based on Dynamo's secondary development for Revit, the crack coordinates were mapped to the BIM model. Indoor tests and field tests show that the method in this paper can be used for underwater crack detection and location recording, which intuitively reflects the concrete damage.