The ability to continuously recognize locomotion modes and accurately predict transition intentions is essential for intelligent prosthetic knees. In this study, an innovative framework for locomotion recognition and transition prediction was introduced based on fusing mechanical (inertial measurement unit (IMU)) and biomechanical (force myography (FMG)) signals. This framework integrated an FMG-IMU dual-modal sensing system implemented on a prosthetic knee, enabling simultaneous acquisition of FMG-IMU fusion signals from transfemoral amputees during dynamic walking. A novel feature-driven CNN-BiLSTM model was developed and trained as the classifier, enhancing the accuracy and efficiency of locomotion mode prediction. The RelifF-MI algorithm was employed to optimize FMG-IMU features, ensuring efficient data processing by effectively eliminating feature redundancy. The framework was evaluated using data collected from eight transfemoral amputees. The results demonstrated that the fusion of FMG-IMU dual-modal gait data with the feature-driven classifier significantly improved classification performance, achieving an overall average recognition accuracy of 98.51% and an average prediction time of 274 ms (21.82% of the gait cycle) across five locomotion modes—level walking (LW), stair ascent/descent (SA/SD), and ramp ascent/descent (RA/RD)—and eight transitions between these modes. These promising results highlighted the considerable potential of the proposed method for application in prosthetic knee control.

To enhance the adaptability and human-machine coordination of intelligent prosthetic knees, this study proposes a motion intention mapping direct myoelectric control method based on an LSTM network and a human-machine coupling model. Multichannel surface electromyography (sEMG) and knee joint angle data were collected during level-ground walking. Time-domain features were extracted to construct an LSTM prediction model, enabling temporal mapping between muscle activity and joint kinematics. Experimental results show that the LSTM model outperforms traditional neural networks in terms of prediction accuracy and temporal consistency. Furthermore, by integrating the human-machine coupling dynamics model with a hydraulic actuation system, a direct myoelectric control framework for a variable-damping prosthetic knee was established, achieving continuous damping adjustment and smooth gait transition. The results verify the feasibility and effectiveness of the proposed method in human-machine coordinated control.

To address the limitations of current prosthetic knees that lack personalized adaptability to users’ gait characteristics and walking speeds, this study proposes a personalized gait parameter prediction–based speed-adaptive control method for a hybrid active–passive intelligent prosthetic knee (HAPK). The proposed system integrates a perceptron-based model to predict individualized gait parameters by mapping anthropometric data and walking speed to key points of the knee trajectory. A fuzzy logic–based damping control for the swing phase and a position–torque control for the stance extension phase are developed to achieve real-time adaptation to different walking speeds and user-specific biomechanics. The hybrid actuation system combines hydraulic damping and motor torque assistance to ensure both compliance and power delivery across gait phases. Experimental results from variable-speed walking tests demonstrate that the proposed control method improves gait symmetry indices—reducing stance and swing asymmetries by approximately 30–38%—and achieves smoother, more natural gait transitions compared to traditional fixed-gait control strategies. These findings validate the effectiveness of the proposed approach in achieving continuous, personalized, and speed-consistent gait control for intelligent prosthetic knees.