An Automatic Labeling Strategy for Locomotion Mode Recognition with Robotic Transtibial Prosthesis

Doffing and Donning the prosthetic socket seriously influenced the performances of the locomotion mode recognition. To make the recognition algorithm adaptive to the disturbances, labeling the new coming data without human intervene was a key step. In this study, we proposed an automatic labeling method with on-prosthesis mechanical signals. The strategy was designed based on the dynamic time warping (DTW) to measure the similarities between the data of a whole completed gait cycle and the pre-defined templates. The automatic labeling algorithm was validated on 6 unilateral transtibial subjects wearing the robotic prosthesis. 2 experimental sessions with 5 locomotion mode and 8 locomotion transition tasks were investigated. Between the sessions, donning and doffing the prosthetic socket were done by the subjects without expertise manual configuration. We evaluated two template generation methods, i.e. the fixed template and the sliding template. The average accuracy of the automatic labeling after re-wearing the socket achieved 96.2% across the subjects, which was comparable to existing sEMG-based studies. The strategy provided a promising tool to accumulate training with inertial signals for locomotion mode recognition.