To address the challenges of human-experience dependency and low anomaly detection efficiency in traditional discrete assembly systems, this study proposes a digital twin-driven anomaly response and hierarchical early-warning framework. Focusing on a planetary gear reducer assembly line, we establish a virtual-physical mapping monitoring system through integrated data sanitization, reduced-order modeling (ROM), and machine learning. An IPOA-optimized K-Means +  + clustering algorithm enables preliminary anomaly identification, while ROM techniques reduce mechanical simulation complexity. An enhanced backpropagation neural network achieves hierarchical early-warning with 95.7% prediction accuracy. Experimental results demonstrate a 70% improvement in computational efficiency for the reduced-order model. The developed digital twin platform validates the framework’s efficacy, offering theoretical support for intelligent transformation in discrete manufacturing.