Abstract: To identify the operational risk of dam construction quickly and accurately, we develop an intelligent risk identification method (YOLO-CDSRI) for the safety risks of cross operation on a concrete dam surface, based on the YOOv8 network and considering the characteristics of complex scenes of such operation. First, a backbone network is constructed using a Cross Stage Partial Network (CSPNet) module and a Spatial Pyramid Pooling-Fast (SPPF) module to enhance the model's situational awareness of safety risks shown in the construction site images. Then, to address the issues of misidentification and missed identification of small target safety risks, this method adopts the Bidirectional Feature Pyramid Network (BiFPN). And using bidirectional cross scale connections and weighted feature fusion, it strengthens information coupling between the risk features and enhances the model's attention to small target safety risks. Finally, the method evaluates the quality of the anchor box via an "outlier" to avoid the excessive influence of geometric factors of the label box on the model, by using Wise-IoU as the boundary box regression loss function and combining with the dynamic non-monotonic focusing mechanism. Results show that after 500 iterations of training, the comprehensive performance of YOLO-CDSRI is superior to YOLOv5s, SSD, and Faster-RCNN models, thus promoting intelligent identification of the safety risks in cross operation on concrete dam surfaces.

Key words: concrete dams, cross operations, complex scenarios, safety risks, intelligent identification