Deep learning-based detection and condition classification of bridge elastomeric bearings

Abstract

The condition of the bearings is crucial for the operational functionality of the bridge. However, detecting damage in elastomeric bearings, which are often deeply positioned in bridges and surrounded by complex structures, remains a challenging task when using computer vision. To address this issue, a comprehensive dataset of bearing images relevant to engineering was first established. Subsequently, a lightweight network that incorporates a multi-level information fusion attention mechanism called LSK-YOLOv8n was introduced. The optimal embedding location for Large Selective Kernel Attention was determined to improve the lightweight nature of the LSK-YOLOv8n model, ensuring optimal performance in detecting bearing damage. Furthermore, the Focal-EIoU loss function was utilized to address the issue of positive and negative sample inhomogeneity more effectively. Experimental results demonstrate that LSK-YOLOv8n outperforms the benchmark model by 5.7% in recognizing bearing diseases, as measured by mAP0.5%. Moreover, LSK-YOLOv8n surpassed several state-of-the-art networks in both the number of integrated parameters and mAP0.5%, demonstrating its effectiveness and robustness.