基于卷积神经网络的荷载大小与位置同步识别

Synchronized identification of dynamic load magnitude and location based on convolutional neural network

结构健康监测和状态评估中现有大多数研究均需要精确的荷载作用位置或详细的荷载时程,为了同时获得荷载大小和位置,构建并训练了同时具备分类和回归能力的两分支卷积神经网络,建立了融合分类问题和回归问题的损失函数,提取结构响应与荷载大小、结构响应与荷载位置间的映射关系.通过数值简支梁算例和三层试验框架验证了该方法识别结构荷载大小和位置的精度.结果表明:噪声条件下数值模型的荷载识别误差在8%以内,荷载位置识别准确率在95%以上;实际结构的荷载识别误差在18%以内,荷载位置识别准确率为100%.两分支卷积神经网络可以很好地同时识别荷载大小和位置.

Most of the existing research in structural health monitoring and status assessment requires accurate load action locations or detailed dynamic load schedules.To simultaneously obtain the size and location of the dynamic load,a two-branch convolutional neural network with both classification and regression capabilities is constructed and trained.A loss function that integrates classification and regression problems is established to capture the mapping relationship between structural response and load magnitude,as well as between structural response and load location.The identification accuracy of load magnitude and location is demonstrated through numerical cantilever beam examples and a three-layer experimental framework.Results show that under noisy conditions,the error in load magnitude identification of the numerical model is within 8%,and the accuracy of load location identification is above 95%.For real structures,the error in load magnitude identification is within 18%,and the accuracy of load location identification is 100%.The two-branch convolutional neural network can effectively identify both the magnitude and location of dynamic loads.