Dual-Branch Fusion Network: Precise Decoding of Lower Limb Multi-Joint Torque

OBJECTIVE: To address the critical challenge of providing accurate, real-time lower-limb joint torque estimation across diverse locomotion conditions for adaptive human-exoskeleton interaction. METHODS: We developed a novel dual-branch architecture that synergizes temporal convolutional networks (TCN) and transformers to process surface electromyography and kinematic data. The TCN captures local temporal dynamics, while the transformer extracts global dependencies. A joint-specific task-aware residual fusion mechanism was introduced to dynamically synthesize these features, employing residual enhancement to adapt precisely to the distinct biomechanics of individual joints. RESULTS: Validated across twelve diverse locomotion patterns, the framework achieved root mean square errors (Nm/kg) and Pearson correlation coefficients of 0.1655/0.9904 (ankle), 0.1405/0.9588 (knee), and 0.1975/0.9698 (hip). It maintained a 4.2912 ms latency and showed strong adaptability on public datasets. CONCLUSION: The proposed method effectively balances high estimation accuracy with the strict computational efficiency needed for real-time applications, successfully addressing previous issues in adapting to dynamic environments. SIGNIFICANCE: This work advances biomedical engineering by providing a fast, reliable solution for adaptive exoskeleton torque control, significantly enhancing seamless and natural human-robot interaction in assistive exoskeleton technologies.