Abstract: To address the challenges of indistinct features in freezing-type slow-varying faults and the limited availability of measurable parameters in the Rudder-Loop-System, this study introduces digital twin technology into Rudder-Loop-System health management and proposes a digital-twin-based fault prediction method. First, a general digital-twin-driven framework is designed for freezing fault prediction, comprising three core modules: fault feature construction, feature enhancement, and fault prediction. Second, based on a digital twin virtual model, a multi-parameter integrated health baseline for the Rudder-Loop-System is established. By integrating measurable data with extended digital twin data, a comprehensive feature set for slow-varying fault characterization is constructed. Third, a dynamic prior distribution and a differential Kullback-Leibler (KL) divergence constraint are introduced to improve the LSTM-VAE model, thereby enhancing the representation of fault features. Finally, a fault prediction model combining LSTM-HMM and dynamic adaptive thresholds is constructed. In this model, the HMM uses prior data to predict future trends of KL divergence-based fault features. Experimental results demonstrate that, under the proposed digital-twin-driven framework, the developed fault prediction module achieves lower prediction error, higher precision, and greater accuracy, enabling effective extraction of fault evolution characteristics and early prediction of freezing faults.

Key words: digital twin, rudder loop, health baseline, fault prediction, health management