ObjectiveTo propose a prediction method combining transition-aware soft labels and dynamic weighting to mitigate transition-zone prediction jitter of hard labels in locomotion intent prediction for intelligent lower-limb prostheses. MethodsSeven healthy subjects were recruited from Tongling University from January to July, 2025, wearing simulated lower-limb-prosthesis braces. Six inertial measurement units were used to acquire acceleration and angular-velocity under 21 locomotion modes. Cosine-based soft labels were introduced to reconstruct the smooth evolutionary gradient of locomotion-mode transitions. Probability-distribution features derived from soft labels were used to build a transition-aware weighting mechanism, which drived the model to focus dynamically on transition boundaries during training. A temporal convolutional network served as the prediction backbone. ResultsThe proposed method achieved an overall prediction accuracy of 94.03% across all locomotion modes, which substantially improved prediction performance during mode transitions, and suppressed probability jitter near decision boundaries and yielded smooth continuous probability trajectories of intent evolution. ConclusionThe proposed transition-aware soft-labeling and dynamic-weighting approach alleviates transition-region prediction oscillation caused by traditional discrete labels. It can provide smooth, robust and anticipatory intent-decision signals for intelligent lower-limb prostheses.

Objective To address the issue of motion misclassification caused by individual gait differences and fixed time window data extraction in motion intention recognition for intelligent lower limb prostheses,this study proposes a motion intention recognition method based on adaptive temporal alignment.Methods In lower limb motion analysis,for continuous gait cycle data,inter-class variability across different steady-state modes was utilized to detect gait pattern consistency through inter-cycle frame differencing.For samples identified as single steady-state modes,the dynamic time warping algorithm was introduced to align adjacent mo-tion sequences,thereby reducing individual variability.Haar wavelet 4-level decomposition was applied to ex-tract low-frequency coefficients for feature vector construction,and classification was performed using a support vector machine.The experimental protocol was designed as follows:three inertial measurement units were used to collect lower limb acceleration and angular velocity data from subjects performing thirteen locomotion modes.The test subjects included ten healthy participants and one transtibial amputee.The locomotion modes consisted of five steady-state modes(level walking,stair ascent,stair descent,ramp ascent,and ramp descent)and eight transition modes(mutual transitions between level walking and stair ascent/descent,as well as ramp ascent/de-scent).Results Simulation tests on ten healthy individuals and one amputee showed recognition accuracies of 99.24%and 100%for five steady-state modes,and 98.51%and 89.11%for all thirteen motion modes,respectively.Conclusion This study proposes an adaptive temporal alignment-based motion intention recognition method.The pro-posed approach effectively reduces the interference of individual gait variability on feature representation,en-hances the consistency and discriminability of gait features,and ultimately improves recognition performance.

Objective To address the issue of motion misclassification caused by individual gait differences and fixed time window data extraction in motion intention recognition for intelligent lower limb prostheses,this study proposes a motion intention recognition method based on adaptive temporal alignment.Methods In lower limb motion analysis,for continuous gait cycle data,inter-class variability across different steady-state modes was utilized to detect gait pattern consistency through inter-cycle frame differencing.For samples identified as single steady-state modes,the dynamic time warping algorithm was introduced to align adjacent mo-tion sequences,thereby reducing individual variability.Haar wavelet 4-level decomposition was applied to ex-tract low-frequency coefficients for feature vector construction,and classification was performed using a support vector machine.The experimental protocol was designed as follows:three inertial measurement units were used to collect lower limb acceleration and angular velocity data from subjects performing thirteen locomotion modes.The test subjects included ten healthy participants and one transtibial amputee.The locomotion modes consisted of five steady-state modes(level walking,stair ascent,stair descent,ramp ascent,and ramp descent)and eight transition modes(mutual transitions between level walking and stair ascent/descent,as well as ramp ascent/de-scent).Results Simulation tests on ten healthy individuals and one amputee showed recognition accuracies of 99.24%and 100%for five steady-state modes,and 98.51%and 89.11%for all thirteen motion modes,respectively.Conclusion This study proposes an adaptive temporal alignment-based motion intention recognition method.The pro-posed approach effectively reduces the interference of individual gait variability on feature representation,en-hances the consistency and discriminability of gait features,and ultimately improves recognition performance.