Digital twin-driven key events and product quality prediction for time-varying manufacturing system

Real-time monitoring and precise quality prediction in modern manufacturing face severe challenges due to the dynamic time-varying characteristics of production systems. To address the limitations of traditional methods in adapting to these characteristics and the disjointedness between event identification and quality prediction, a collaborative prediction framework based on a Time-Varying Digital Twin System (TV-DTS) is proposed. First, a key event identification method that fuses temporal reconstruction deviations with multi-dimensional dynamic statistical features is designed to simultaneously capture global dynamic pattern deviations and local fluctuations during system operation. Second, by combining rule-based reasoning with mutual information analysis, an event-scenario matching mechanism is established to achieve an effective association between key events and production scenario semantics. Finally, event information and scenario semantics were incorporated into a quality prediction model to form an event-scenario fusion-driven quality prediction approach. A case study on the tobacco loosening and conditioning process demonstrated that the proposed framework effectively distinguished complex anomaly patterns, improved the accuracy of key event identification, and significantly reduced quality prediction errors, indicating its potential applicability for continuous manufacturing processes with similar time-varying characteristics.