TimesNet-Based Human-Robot Interaction Intent Recognition for the Orthopedic Surgical Robot

Robot-assisted Total Knee Arthroplasty (TKA) osteotomy demands highly compliant and safe physical humanrobot interaction (pHRI). Heavy surgical instruments and high-frequency non-stationary vibrations prevent conventional time-series models from robustly extracting surgeons' genuine operational intents from 6D force signals. This paper proposes a TimesNet-based intent recognition and control framework for orthopedic robots. Real-time data preprocessing first performs gravity/eccentric moment compensation and adaptive Butterworth filtering. Subsequently, TimesNet transforms 1D temporal force signals into 2D tensors, effectively decoupling local high-frequency vibrations from long-term intentional trends across multiple period scales. Recognized intents are fed into a smooth anisotropic admittance controller for compliant tracking. Offline evaluations show TimesNet outperforms baseline models like LSTM and Transformer. In simulated osteotomy experiments, the system exhibits robust end-to-end responsiveness, effectively bounds the impact of occasional misclassifications, and significantly enhances pHRI safety and compliance.