Temporal Modeling for Domain-Invariant Fault Diagnosis in Robotics Digital Twin Systems

Fault diagnosis in robotics frequently encounters obstacles due to the scarcity of labeled real-world data and the disparities between simulated and physical systems. Digital twin technology mitigates data limitations by producing simulated datasets; however, the sim-to-real gap remains a barrier to effective model generalization. This study introduces the TemporalTwinNet (TTN) framework, a novel approach designed for digital twin-supported fault diagnosis in robotics. By embedding bidirectional Long Short-Term Memory (LSTM) layers into the feature extraction process, the TTN framework adeptly captures temporal dependencies within time-series data, enhancing the alignment of simulated and real-world trajectories. Tested on an open-source robotics dataset comprising 3,600 simulated and 90 real samples, the proposed TTN achieves a real-world test precision of 86.67\%, significantly narrowing the sim-to-real gap to 9.44\%. Notably, the model elevates F1 scores for difficult categories, such as the healthy state (improving from 0.06 to 0.63), while sustaining strong simulation performance with a precision of 96.11\%. Additionally, the framework integrates severity prediction, boosting its practical utility. These outcomes underscore the efficacy of temporal modeling in overcoming the sim-to-real divide, presenting a resilient solution for predictive maintenance in robotics.