Ball Mill Load Classification Method Based on Multi-Scale Feature Collaborative Perception

Against the backdrop of intelligent manufacturing, the ball mill, as a key energy-consuming piece of equipment, requires an accurate perception of its load state, which is crucial for optimizing production efficiency and ensuring operational safety. However, its vibration signals exhibit typical nonlinear and non-stationary characteristics, intertwined with complex noise, posing significant challenges to high-precision identification. A core contradiction exists in existing diagnostic methods: convolution network-based methods excel at capturing local features but overlook global trends, while Transformer-type models, although capable of capturing long-range dependencies, tend to “average out” critical local transient information during modeling. To address this dilemma, this paper proposes a new paradigm for multi-scale feature collaborative perception. This paradigm is implemented through an innovative deep learning architecture—the Residual Block-Swin Transformer Network (RB-SwinT). This architecture subtly achieves hierarchical and in-depth integration of the powerful global context modeling capability of Swin Transformer and the excellent local detail refinement capability of the residual module (ResBlock), enabling synchronous and efficient representation of both the macro trends and micro mutations of signals. On the experimental dataset covering nine types of fine operating conditions, the overall recognition accuracy of the proposed method reaches as high as 96.20%, which is significantly superior to a variety of mainstream models. To further verify the model’s generalization ability, this study was tested on the CWRU public bearing fault dataset, achieving a recognition accuracy of 99.36%, which outperforms various comparative methods such as SAVMD-CNN. This study not only provides a reliable new technical approach for ball mill load identification but also demonstrates its practical application value in indicating critical operating conditions and optimizing production operations through an in-depth analysis of the physical connotations of each load level. More importantly, its “global-local” collaborative modeling concept opens up a promising technical path for processing a broader range of complex industrial time-series data.