A lightweight and explainable model to predict the dynamic normal stress on silo wall assisted with knowledge distillation

Abstract

It is challenging to develop a widely accepted model for predicting the dynamic normal stress (DNS) on silo wall. A lightweight and explainable model was proposed to predict the dynamic normal stress (DNS) on silo wall, assisted by knowledge distillation (KD). Feature distribution and correlation analyses were performed to establish a novel DNS database by eliminating redundant features. Various machine learning models were compared for accuracy and efficiency. Results indicate that the optimized model shows significant performance improvement. Adding residual modules and increasing the number of network layers effectively enhance the performance of the ResGRU model. Feature fusion and derivative modules drives derivation distillation (DD) model, reducing the number of parameters and storage space. Finally, SHAP method was used to help the proposed model to explain the importance ranking of the features, and it is recommended first to consider granular material type, aspect ratio, and hopper angle when designing silo structures.