SHAP-LightGBM stacked learning prediction model for NdFeB production process optimisation

NdFeB permanent magnetic materials exhibit a very high magnetic energy product and coercive force, making them widely used in modern industry and electronic technology due to their excellent magnetic properties. The pressing process plays a crucial role in the production of NdFeB. Currently, most process parameters are determined empirically, which often leads to process instability and adversely affects product quality. To optimize the process parameters in the NdFeB powder compacting stage, this paper proposes a stacking model based on LightGBM with SHAP for feature selection. Furthermore, to enhance the prediction accuracy of the model, hyperparameter optimization is performed using the NSGA-II multi-objective optimization algorithm. Finally, model optimization is achieved by combining Optuna with the NSGA-II optimization algorithm. The predicted process parameters include demagnetization current, upper cylinder position during magnetization, and lower cylinder position during magnetization. The experimental results indicate that the model can effectively optimize the process parameters, providing valuable reference data for actual production, thereby improving production efficiency, reducing costs, and enhancing product quality.