Power-Aware State Recognition for Digital Twins: Non-Intrusive Industrial Monitoring of Solder Paste Printers

The construction of high-fidelity virtual factories relies heavily on the accurate reconstruction of historical production timelines. While traditional Manufacturing Execution Systems (MESs) provide idealized, static schedules, they inherently struggle to capture the “stochastic variability” and “undefined events” caused by machine-specific behaviors on the industrial shop floor. To bridge the gap between top-down scheduling and bottom-up physical reality, this study proposes the Power-Aware State Segmentation for Solder Paste Printers (PAS-SPP) algorithm. Utilizing non-intrusive, high-frequency continuous power features captured via an Industrial Internet of Things (IIoT) architecture with PA310 meters, the algorithm employs a synergistic combination of amplitude thresholding (θhigh) and temporal constraints (τblank, τmin) to actively filter transient electrical noise and accurately bound macroscopic operational blocks. This robust filtering thereby avoids the accuracy degradation commonly caused by noise interference in the analysis processes of traditional machine learning models. Consequently, the mechanism effectively decouples operational states into a virtual Solder Paste Printer (vSPP) behavioral meta-model integrated with a Finite State Machine (FSM). Empirical validation across distinct production cases demonstrates that the proposed model not only accurately extracts standard 33–35 s cycle times but also reveals critical hidden characteristics, such as 68 s automated cleanings and dynamically adjusted “3-to-1” print-to-clean ratios. Furthermore, a comprehensive comparative analysis was conducted against static MES logs, Naive Power Thresholding (NPT), and a Gaussian Hidden Markov Model (GMM-HMM). Evaluated under identical manufacturing-process conditions, the results reveal that PAS-SPP effectively mitigates the cascading misalignments in static schedules and avoids the severe over-segmentation limitations inherent in point-by-point probabilistic decoding, thereby achieving highly accurate state decoupling. Finally, this study systematically defines the method’s applicability boundaries across diverse DT domains, confirming its indispensable role as a non-intrusive, broadly applicable event-triggering foundation for the broader smart manufacturing ecosystem.