Stability prediction of multi-material complex slopes based on self-attention convolutional neural networks

This study proposes an integrated slope stability prediction model for various complex slope scenarios, including soil, rock, and rock-soil mixed situations. First, a small number of numerical slopes are constructed using the digital twin (DT) technique, and then these slope parameters are sorted and fine-tuned to build a database containing 19,666 soil, single/multiple sets of inclined joints, and rock-soil mixed slope scenarios. Second, the self-attention (SA) mechanism that can analyze the correlation of data features is connected to a classical convolutional neural network (CNN), forming a trained CNN-based SA model (CNN-SA) with 80% of the samples from the built database. The remaining 20% of the database and the stability of six actual slopes are then used for prediction. The performance of the CNN-SA is compared and evaluated. The results indicate that the DT technique is a reliable tool for providing the data to train the AI models, especially when the sample data is limited. As the complexity of the slopes increases, the prediction error of the models increases, and the CNN-based SA mechanism can effectively reduce these prediction errors compared to a classical CNN and other attention mechanisms.