Slope Displacement Prediction Based on Cross Distillation for Small Samples

Slope displacement monitoring is one of the important means for geological disaster early warning. Since slope displacement is affected by many factors and the slope deformation time is long, there is few effective data that can be used for slope warning. For few effective samples, a lightweight prediction model based on knowledge distillation is established by introducing model compression to predict the slope displacement while improving the timeliness of the prediction. The difficulty in designing a compression prediction model for small samples is how to avoid the deterioration of network output due to the accumulation of estimation errors during the layer-by-layer propagation in the training process. To solve this problem, cross distillation technology is introduced. By interweaving the hidden layers of the teacher network and the student network, and pruning the student network, the accumulated estimation errors and training parameters are reduced. To verify the validity of the proposed prediction model, an experiment was carried out with the slope displacement data collected by Jinyu Zenith Cement plant. Experimental results show that when there are only a few samples, the proposed model can reduce estimation errors and improve the accuracy of slope displacement prediction, which can provide certain reference value for early warning of slope instability in actual mines.