Dynamics modeling method for fully-mechanized mining equipment system based on improved neural relational inference

The fully mechanized coal mining equipment system serves as the core foundation for safe and efficient production in modern coal mines. Its complex dynamic behaviors and multi-equipment coupling relationships pose severe challenges to system dynamics modeling. Traditional methods often rely on expert experience to predefine system dynamics relationships, making it difficult to adaptively uncover implicit time-varying correlations within the system. Moreover, their generalizability and interpretability are limited in multi-equipment, multi-modal coupling scenarios. To address these issues, the neural relational inference model is pioneered into the dynamics modeling of coalmine fully mechanized mining equipment systems, and an improved neural relational inference method based on an attention mechanism (ANRI) is proposed. This approach achieves the synergistic optimization of system coupling structure inference and dynamic behavior modeling. First, the method accomplishes the construction of a node system for the fully mechanized mining equipment system based on multi-dimensional evidence support. It systematically integrates three categories of evidence — “standards and research basis,” “inherent equipment attributes,” and “expert knowledge” — to build a node system that combines theoretical rigor, physical essence, and engineering practicality. Second, the method learns both equipment node dynamics and internal system coupling relationships in an unsupervised manner, enabling explicit inference of various complex interaction relationships, both between equipment (e.g., “shearer traction unit–scraper conveyor drive unit”) and within equipment (e.g., “shearer ranging arm–shearer drum”). To address the limitation of the original Neural Relational Inference model in distinguishing the importance of neighbor node messages during message aggregation, an attention mechanism is introduced to improve the message-passing process. This allows the model to adaptively weight key coupling relationships, significantly enhancing the accuracy and interpretability of dynamics modeling. Finally, a joint learning framework for dynamics modeling and relational reasoning tailored to the fully mechanized mining equipment system is constructed. Through end-to-end training, the tasks of coupling relationship inference and state prediction are simultaneously optimized, achieving accurate characterization of system dynamic behavior. Experiments on fully mechanized mining equipment state data show that the prediction errors of the ANRI model for key parameters (such as shearer traction motor speed) are significantly lower than those of traditional time-series models. Specifically, the mean squared error is reduced by approximately 92% and the mean absolute error is optimized by about 74% compared to the Recurrent Neural Network model. Furthermore, the visualization results of the coupling structure align with the system’s physical mechanisms, verifying the rationality and interpretability of the inferred relationships. This method provides a new approach for the dynamics modeling of fully mechanized mining equipment systems.