Research on Settlement Prediction and Evaluation Methods for Soft Soil Layers Based on Deep Learning

The aim of this study is to investigate an effective prediction and evaluation method of soft soil layer settlement using deep learning method, which will solve the challenge of traditional methods being unable to achieve high-precision predictions under a complex soil environment. Therefore, in this paper, we design a deep learning model based on Long Short-Term Memory Network (LSTM), and also make innovative improvement on predictions regarding the input of soft soil layers settlement prediction. Therefore, in the specific methods, the first is to improve the LSTM model, by combining adaptive activation functions (such as Swish) to optimize the nonlinear expression ability of the network, and introducing a multi-scale feature extraction mechanism for time series data, so that the model can better deal with the settlement changes in different time periods, thus improving the accuracy of the model prediction results. The model, however, with this multi-scale feature extraction approach, demonstrated better performance to reflect the dynamic variation and temporal deviation of the soft soil settlement with time. Experimental results demonstrate that the proposed improved LSTM model outperforms the traditional method for predicting the settlement of various actual datasets, which affirms its effectiveness and superiority in targeting the assessment of soft soil layers.