Application of a multi-residual unit fusion lightweight network model in ball mill load recognition

To address complex architectures, parameter redundancy, high computational costs, weak generalization, and limited accuracy of traditional deep learning in ball mill load recognition, this paper proposes a multi-residual unit fusion lightweight network. The model constructs a multi-scale feature fusion architecture using dual residual modules (Resblock-I/II) with Dropout to extract hierarchical features. Gaussian error linear unit is introduced to improve gradient propagation continuity, and L1 norm-based structured pruning reduces model complexity. Results show that, under the chronological split, multi-residual unit fusion achieves 98.81% accuracy on the Case Western Reserve University data set, outperforming SAVMD-CNN and the other comparative methods, and achieves 97.35% accuracy in ball mill load recognition. Pruning and structural optimization reduce model complexity and time overhead to a certain extent while retaining effective recognition performance. Overall, the method balances recognition, complexity, and time overhead, providing a reference for lightweight ball mill load recognition under constrained computational conditions.