基于改进YOLOv11算法的桥墩损伤点智能检测方法

Intelligent detection method for bridge pier damage points based on improved YOLOv11 algorithm

摘要:

为了提升桥墩损伤点检测准确性，设计了基于改进YOLOv11算法的桥墩损伤点智能检测方法。利用振型变化识别结构损伤，提取模态阵型变化列向量的最大元素，确定桥墩结构模态阵型损伤点程度变量。以C3k2模块替换原有C2F模块，并在C3k2模块中并行引入C2PSA空间注意力机制，对特征图进行自适应的权重重标定，使网络能够自主聚焦于裂缝尖端、剥落边界和锈蚀中心等关键损伤区域。通过空间索引结构查找邻域点并拟合桥墩轴心中线，建立移动最小二乘拟合平面，拟合全局点云曲面，计算损伤点到曲面的法向距离实现损伤点定位。实验结果表明，Precision、Recall、F1-Score、mAP@0.5、mAP@0.5:0.95等指标均高于90%，在桥墩损伤点智能检测任务中具有高精度与强鲁棒性，对于桥墩的维修与加固决策具有重要支撑作用。

Abstract:

To enhance the accuracy of detecting damage points on bridge piers, an intelligent detection method based on an improved YOLOv11 algorithm was designed. Structural damage was identified using changes in vibration modes, and the maximum element of the modal shape change column vector was extracted to determine the degree variable of modal shape damage points on the bridge pier structure. The original C2F module was replaced with a C3k2 module, and a C2PSA spatial attention mechanism was introduced in parallel within the C3k2 module to perform adaptive weight recalibration on the feature map, enabling the network to autonomously focus on key damage areas such as crack tips, spalling boundaries, and corrosion centers. Neighboring points were searched through a spatial index structure, and the centerline of the pier axis was fitted. A moving least squares fitting plane was established to fit the global point cloud surface, and the normal distance from the damage point to the surface was calculated to locate the damage point. Experimental results showed that Precision, Recall, F1-Score, mAP@0.5, and mAP@0.5:0.95 were all above 90%. This method demonstrated high accuracy and strong robustness in the intelligent detection task of bridge pier damage points, providing important support for bridge pier maintenance and reinforcement decisions.In the detection of bridge pier damage points, the damage point variable cannot describe the damage degree. There is a local optimal situation in the point cloud under the C2F module, resulting in indicators such as Precision, Recall, F1-Score, mAP@0.5, and mAP@0.5:0.95 being lower than actual demand, which affects the use quality of bridge piers. Therefore, an intelligent detection method for bridge pier damage points based on the improved YOLOv11 algorithm was designed. From the perspective of structural dynamics, structural damage is identified by using the change of vibration shape. By establishing the free vibration differential equation of the bridge pier, the curvature function of any section under the damage state is solved, and the maximum element of the modal formation change column vector is extracted to determine the modal formation of the bridge pier structure. The degree of damage point degree variable combines the amplitude information of local change of modal vibration shape with the degree of stiffness reduction caused by damage. The original C2F module is replaced with a C3k2 module, which provides a non-attenuating gradient channel for backpropagation through a cross-stage local connection mechanism, and a C2PSA spatial attention mechanism is introduced in parallel into the C3k2 module to perform self-adaptive re-calibration of the feature map in spatial dimensions allows the network to autonomously focus on key damage areas such as crack tips, spalling boundaries, and corrosion centers, while suppressing irrelevant responses in background areas. The damage point degree variable is encoded as prior information into the BMFPN feature fusion path, realizing cross-modal joint driving of mechanical degradation parameters and visual representations. Neighboring points are found through the spatial index structure and fitted to the center line of the pier axis. Using this as a geometric basis, a moving least squares fitting plane is established, and then the global point cloud surface of the intelligent detection model for pier damage points is fitted. The normal distance from the damage points to the surface achieves accurate positioning of the damage points. Experimental results show that indicators such as Precision, Recall, F1-Score, mAP@0.5, and mAP@0.5:0.95 approach 100% infinitely. The high accuracy and strong robustness in the intelligent detection task of bridge pier damage points play an important supporting role in the maintenance and reinforcement decisions of bridge piers.