基于Kalman-BP的高边坡表面变形预测研究与应用

Research and application of Kalman-BP based prediction of surface deformation to high slopes

施工期中,泵站高边坡内部受力复杂,造成的表面变形可能导致滑坡坍塌等危及现场的安全作业事故,因此表面变形趋势预测具有实际工程意义。笔者选取4个现场监测的特征量——温度、锚杆应力、孔隙水压力和土压力作为模型的输入层,经过Kalman滤波对现场监测特征量的白噪音进行剔除,再进入BP神经网络进行训练,融合模型既能克服Kalman滤波的离散性现象,又提高了BP模型的泛化能力和收敛速度。将模型应用到重庆某泵站高边坡监测项目进行分析验证,结果表明:与传统BP神经网络相比,Kalman-BP神经网络针对H、Y、X三个方向累计变形量的训练迭代步数分别从1 086步减少至1 047步、1 090步减少至1 050步及1 080步减少至1 044步,预测H、Y、X三个方向累计变形量的均方根误差分别从0.974减小至0.684、1.037减少至0.564、0.982减少至0.526,模型的预测能力得到了提高,为高边坡安全作业提供了有效保障。

During construction,the internal forces on the high slopes of pumping stations are complex and the resulting surface deformation may lead to accidents such as landslides and collapses that would endanger the safe operation of the project.Therefore,the trend prediction of the surface deformation is of practical significance.In this paper,four field-monitored characteristic quantities-temperature,anchor stress,pore water pressure and soil pressure-are selected as the model input layer and the white noise of the field-monitored characteristic quantities is removed by Kalman filtering.What followed is the BP neural network for training.The combined model can overcome the discrete phenomenon of Kalman fil-tering and improve the generalization ability and convergence speed of the BP model.The model is ap-plied to the monitoring of high slopes of a pump station in Chongqing,and the results show that com-pared with the traditional BP neural network,the number of training iterations of Kalman-BP neural net-work for the cumulative deformation in the H,Y and X direction is reduced from 1086 to 1047,from 1090 to 1050 and from 1080 to 1044 respectively.The root mean square error of the predicted cumu-lative deformation in the H,Y and X direction is reduced from 0.974 to 0.684,from 1.037 to 0.564 and from 0.982 to 0.526 respectively,which has improved the prediction capability of the model and provid-ed an effective guarantee for safe operation of high slopes.