Abstract: To address the engineering challenges of weak early failure characteristics and insufficient prediction accuracy of remaining useful life（RUL）for harmonic reducer，a condition monitoring and fault prediction method for harmonic reducer based on digital twin is proposed.Firstly，a 5-dimensional digital twin framework including physical entities，virtual models，data twins，service data，and connection interactions is constructed to achieve real-time mapping and collaborative evolution of multi-physical field sensing data and high-fidelity digital twin models.Secondly，by combining multiple sources of data，the attention mechanism is introduced into the long short-term memory network（LSTM），and an LSTM-Attention fault prediction model is constructed to improve the accuracy of degradation trends and early fault predictions.Finally，the effectiveness of the method is verified through accelerated life tests.The results show that compared with the standard LSTM，the root mean square error（RMSE）of the method decreases by 18.7%，the mean absolute error（MAE）decreases by 22.3%，and the fatigue cracks of the flexible wheel are identified 150 working cycles earlier，providing an effective technical path for predictive maintenance of harmonic reducer.