基于带约束BP神经网络的黄土湿陷系数预测模型研究

通过现场试验和室内试验得到某黄土场地Q3黄土的湿陷系数及相关物理力学参数，分析了黄土湿陷系数与其主要物理力学影响因素之间相关性。针对室内试验数据具有离散性较强的特点，基于传统BP神经网络模型进行改进，建立了带约束的BP神经网络模型并对黄土的湿陷系数进行预测。结果表明:湿陷系数与含水率、天然密度、饱和度成负相关，与压缩系数和孔隙比成正相关；带约束的BP神经网络模型可以定量保证预测结果的准确性，其预测误差在10%以内。

Study on Prediction Model of Loess Collapsibility Coefficient Based on BP Neural Network with Constraint

Both the collapsibility coefficient and the related physical and mechanical parameters of loess in a Q3 loess field were obtained through field and laboratory tests to analyze the correlation between the coefficient and the main factors influencing its physical and mechanical properties. Considering the laboratory test data is of strong discreteness, a constrained BP neural network model was established on the basis of the improved traditional model to predict the collapsibility coefficient of loess. The result shows that the collapsibility coefficient relates negatively to water content, natural density and saturation, and positively to compressibility factor and void ratio; the constrained BP neural network model can quantitatively guarantee the error of prediction results within 10%.