Optimising maize threshing by integrating DEM simulation and interpretive enhanced predictive modelling

Maize threshing is a complex and dynamic process, and optimisation of operating parameters is essential to improve threshing quality and efficiency. In this study, machine learning was combined with interpretability analysis to investigate the dynamic effects of operating parameters on maize threshing quality and to optimise the threshing process. The maize cob model used to simulate threshing was validated by stacking angle and tensile test. Real-time drum operating parameters and threshing quality data obtained through Discrete Element Method (DEM) threshing simulation were used to train a threshing quality prediction network. The prediction accuracy was improved by incorporating an attention mechanism into the Long Short-Term Memory (LSTM) model with an optimised Root Mean Square Error (RMSE) of 0.0041. The global feature importance and dynamic Shapley Additive Explanations (SHAP) value analyses demonstrated that rotational speed is a key determinant of unthreshed and damaged rates and that its effect varies significantly at different stages of the threshing process. Guided by these analyses, a staged speed adjustment experiment was conducted. Specifically, an increase in rotational speed during the initial threshing phase markedly lowered the initial unthreshed rate for medium and high-speed groups to 6.63% and 2.73%, respectively, a significant improvement over the 67.70% observed in the low-speed group. The final damage rate in the high-speed group decreased by 9.79% relative to the low-speed group. This dynamic analysis approach provides a novel paradigm for optimising complex agricultural processes under varying conditions, offering interpretable insights for precise process control and improvement.