Continuous estimation of elbow joint angles is vital for motion intention recognition and real-time control in rehabilitation robotics. This paper proposes a novel time series prediction framework based on the Mamba model, which integrates surface electromyography (sEMG) and inertial measurement unit (IMU) signals to achieve high-precision elbow joint angle estimation. The proposed method involves synchronized feature extraction and fusion from multi-channel sEMG and IMU signals within sliding time windows, followed by advanced sequence modeling using the Mamba architecture. Experimental validation on data collected from eight healthy subjects demonstrates that the Mamba model achieves lower root mean square error (RMSE) and higher R2 values compared to long short-term memory (LSTM) and gated recurrent unit (GRU) baselines in both single-angle and multi-angle prediction modes. The Mamba-based framework also shows robust performance under varying window lengths and advance prediction times. These results confirm the effectiveness and generalization capability of the proposed method for dynamic joint angle estimation, providing valuable support for real-time prosthetic control and multimodal human-machine interaction applications.