Abstract:

This study aims to meet the engineering demand for high-precision prediction of surface deformation in open-pit mine slopes. Firstly, a hybrid prediction model of CNN-LSTM was constructed. With a certain open-pit mine taken as a case, time-series deformation data were acquired using S-SAR. Wavelet denoising was applied to preprocess the raw data and enhance data quality. An autoregressive integrated moving average (ARIMA) model was then introduced as the baseline for traditional statistical methods to evaluate the necessity of employing deep learning-based prediction approaches. Finally, the hybrid prediction model of CNN-LSTM was compared with single-structure LSTM and CNN models to verify the scientific soundness and effectiveness of the proposed models. The results show that the LSTM, CNN, and CNN-LSTM models all outperform the ARIMA model in terms of root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and mean absolute percentage error (MAPE), demonstrating the pronounced nonlinear characteristics of slope-deformation sequences. Among the deep learning architectures, the CNN-LSTM model exhibits the best overall performance. Relative to those of the LSTM model, its RMSE and MAE are reduced by 55.54% and 59.24%, respectively. Relative to those of the CNN model, they are reduced by 48.63% and 52.15%, respectively. In addition, the hybrid model's R² and MAPE are improved by 5.29% and 59.80% compared with those of the LSTM model, and by 3.27% and 52.51% compared with those of the CNN model.

Key words: open-pit mine slope, deformation prediction, convolutional neural network-long short-term memory network (CNN-LSTM), slope-synthetic aperture radar (S-SAR), wavelet denoising