基于偏最小二乘的小波网络膨胀土边坡稳定性预测

Wavelet neural network prediction based on partial least square for expansive soil slope stability

为准确预测膨胀土边坡稳定性,避免人员和财产损失,提出一种基于偏最小二乘（PLS）和WNN的组合预测方法。首先利用PLS提取样本数据的主元特征;然后利用获取的主元特征建立WNN边坡稳定性预测模型;最后通过所建立的模型实现对边坡稳定性的预测。结果表明,该预测方法仅经过200余次迭代即可完成模型训练,预测均方误差达到0. 012 5,比单一WNN有了很大改善,从而验证了其可行性和有效性。

The given paper intends to propose a wavelet neural network （WNN） prediction method based on the Partial Least Squares （PLS） feature extraction for the expansive soil slope stability. To achieve the purpose, the author has first of all to solve the problem of the constrained optimization in the PLS process, so as to make the extracted principle components from the independ- ent variables and dependent variables endowed with （contain） the changing information in the independent and dependent variables as much as possible, in hoping to maximize the correlation between the extracted principal components from the independent variables and the dependent ones. This can help to execute the preprocessing steps such as dimension reduction, denoising and multiple-correlation elimination for the sampling data so as to obtain the principal componential features of the slope stability. And, then, it would be possible to establish a slope stability prediction model based on the principal componential features governing the nonlinear modeling ability of WNN. And, for the said demand, it is also necessary to adopt BP algorithm with momentum and adaptive learning rate to avoid falling into the local minimum to regulate the adaptive learning rate according to the local error surface so as to make the algorithm more stable and the learning steps big enough in training WNN, in case the momentum has to be mainly adopted. Finally, the slope stability can be predicted by building a model. The above mentioned experiments show that, through the PLS preprocessing of the sampling data of 9 factors influencing the stability of the expansive soil slope, it has become possible for the WNN model gained with merely one principal component feature to increase its comparatively satisfactory accuracy. All this implies that only through more than 200 iterations for training, the predicted mean square error of the model can be made to reach 0. 012 5, which is supposed to be much more advantageous than that of the single WNN： This result also shows that, the complexity of the WNN modeling can be simplified via the PLS feature extraction, which can not only heighten the accuracy of the slope stability of the prediction model, but can also improve the convergence. Therefore, the method the paper has proposed turns out to be feasible and effective in predicting the stability of the expansive soil slope, it can also be used as a new approach to solving the prediction instability to gain the expansive soil slope stability.