基于改进YOLO与预测光流结合的机翼振动测量方法

Wing Vibration Measurement Method Combining Improved YOLO with Predictive Optical Flow

摘要:

针对机翼模态试验中传统接触式传感器附加质量效应明显、测点有限及多点布设复杂等问题，提出了基于改进YOLO与预测光流结合的机翼振动测量方法YK-WingVib；提出QCT-YOLO振动靶标检测模型，通过EdgeRepConv边缘增强卷积、CSP-EDLAN动态聚合模块及EUCB高保真融合模块，提高四象限十字靶标在复杂工况下的检测精度；针对传统KLT光流在大位移场景易失配、长时程测试易漂移等问题，提出PRA-KLT预测性光流跟踪方法，实现测量点稳定连续跟踪；经消融实验、多算法对比实验及机翼缩比模型试验验证，QCT-YOLO模型mAP@0.5:0.95达到0.904，PRA-KLT算法多频段RMSE稳定在0.67px以内，真实环境下机翼缩比模型试验中测量点振动位移平均RMSE为0.35mm，与光纤传感器测量结果相关系数平均为0.989，一阶固有频率平均相对误差为0.126%，满足机翼地面振动试验与结构健康监测中的高精度非接触测量需求。

Abstract:

To address the mass-loading effect, limited measurement points, and complicated multi-point deployment of conventional contact sensors in wing modal tests, a wing vibration measurement method combining improved You Only Look Once (YOLO) with predictive optical flow, termed YK-WingVib, was proposed. A quadrant-cross-target detection model, QCT-YOLO, was developed by introducing edge-enhanced re-parameterized convolution (EdgeRepConv), a cross-stage partial edge dual-level aggregation network (CSP-EDLAN), and an efficient upsampling convolution block (EUCB) to improve detection accuracy under complex conditions. To reduce mismatching under large displacement and tracking drift during long-duration tests, a predictive robust adaptive Kanade–Lucas–Tomasi method, PRA-KLT, was proposed for stable and continuous measurement-point tracking. Ablation experiments, algorithm comparisons, and scaled-wing model tests showed that QCT-YOLO achieved a mean average precision (mAP)@0.5:0.95 of 0.904, while the root mean square error (RMSE) of PRA-KLT remained within 0.67 px across multiple frequency bands. In scaled-wing model tests under real experimental conditions, the average displacement RMSE was 0.35 mm, and the mean correlation coefficient with fiber-optic sensor measurements was 0.989. The mean relative error of the first natural frequency was 0.126%. The method meets the requirements for high-precision non-contact measurement in wing ground vibration tests and structural health monitoring.