在地铁隧道建设过程中，渣土改良效果是影响盾构掘进速度的关键因素。为确保盾构顺利掘进，以坍落度试验、渗透试验和电阻率测试的结果作为数据样本集，根据深度学习领域常用的数据划分方法将训练集、验证集和预测集按照6∶2∶2进行划分，基于BP神经网络建立渣土改良效果的预测模型，对南昌地区富水砂层进行渣土改良效果预测，并将预测值和实际值进行对比分析。研究结果表明： 坍落度、渗透系数、内摩擦角在模型学习过程中的平均预测值分别为1728 mm、3355×10-6 cm/s、216°，相对误差平均值分别为176%、453%和360%； 预测集的输出结果与实测值的部分数据重合，三者的平均误差均在5%以内，可决系数R2值分别为088、090和085，表明该神经网络结构属于高精度模型。预测结果的误差均在现场渣土改良的允许误差范围内，可见BP神经网络模型能够对渣土改良的效果进行精准预测。

The ground conditioning effect is the key to shield metro tunneling speed. As a result, the testing results of the slump, penetration and resistivity are taken as the data sample set. And then the training set, verification set and prediction set are divided into 6∶2∶2 according to the data partitioning method commonly used in the field of deep learning. Finally, a prediction model for ground conditioning effect is established based on the BP neural network, and the model is applied to the ground conditioning effect prediction of waterrich sandy stratum in Nanchang. The study results show that: (1) The average predicted values of slump, permeability coefficient and internal friction angle during model learning are 1728 mm, 3355×10-6 cm/s, and 216°, respectively, and the average relative errors are 176%, 453%, and 360%, respectively. (2) The output of the prediction set coincides with part of the measured data; the average errors of the slump, permeability coefficient and internal friction angle are all within 5%; and the determinable coefficients R2 are 088, 090, and 085, respectively, which indicates that the neural network structure is a highprecision model. The errors of the prediction results are within the allowable error range of the ground conditioning, which shows that the BP neural network model can accurately predict the effect of ground conditioning.