Developing a Classification Algorithm for Plantarflexor Actuation Timing of a Powered Ankle–Foot Orthosis1

摘要

The development of powered orthoses and exoskeletons for robotic gait assistance has led to new issues related to their control. Limited work has been published regarding when to provide plantarflexor torque to the ankle [1–3]. If actuation is given too early, four times more energy for walking may be necessary [1]. Some groups have followed a functional approach and used minimization of metabolic cost as the optimization criteria for detecting appropriate plantarflexor actuation timing (e.g., see Ref. [2]). Optimizing timing using metabolic data can be time-consuming (10 s of minutes of walking). In addition to functional energetics, the effect of these devices on joint biomechanics should also be considered.The portable powered ankle foot orthosis (PPAFO) is capable of providing both plantarflexor and dorsiflexor torques by using a pneumatic rotary actuator at the ankle [4]. The PPAFO has three sensors: force-resistive sensors under the heel and ball of foot to detect foot contact, and a Hall effect sensor to record ankle angle.Joint biomechanics have been observed to vary considerably when the plantarflexor torque actuation timing was varied when using the PPAFO (Fig. 1). Sudden jumps toward the plantarflexion direction are observed, when the plantarflexor torque was earlier than necessary (30% and 40% of the gait cycle (GC), dotted lines). When actuation was given late, the maximum amplitude and timing in the plantarflexion direction were inconsistent (55% and 60% GC, thin solid lines). Plantarflexion maximum range of motion was diminished, and a discontinuity in the motion was observed due to delayed assistance. The plantarflexor actuation timing that created the smoothest ankle angle profile was found to be 49% GC (thick solid line). At this timing, these early and late artifacts were minimized.By estimation of a walker's instantaneous state of the limb during a single stride, as represented by a specific % GC, it is possible to detect and control for various gait events. Properly timed control of powered assistance during walking is a crucial task to prevent tripping or fall risk. Early state estimation-controlled PPAFO studies used dorsiflexor and plantarflexor actuation timings based on the normative event timings for healthy able-bodied adult gait [4]. The plantarflexor actuation timings were often further fine-tuned for a given subject by using a trial and error method.In this study, we took a biomechanics approach and used replication of ankle angle kinematics as the optimization criteria. We proposed a multistep, supervised learning classification algorithm to identify the plantarflexor actuation as early or late using the ankle angle collected during walking.Five healthy adult males (age 26.40 ± 5.0 yr, height 1.79 ± 4.7 m, and weight 80.70 ± 5.8 kg), without any neurological, gait, or postural disorders, participated in the study. All the subjects gave informed consent, and this study was approved by the university's Institutional Review Board.Each subject participated in two sessions of experiments in the same day with a break of up to 10 min. There were two test conditions per session (shoes only and PPAFO on right leg). In session 1, data were collected to create training data for the classifier. In session 2, the trained classifier was combined with a bisection search technique to quickly identify the appropriate actuation timing for a specific user. To assess the accuracy of the classifier, cross-correlation analysis was used to compare the shoe-only and PPAFO data. The shoe-only data were considered to be the true reference data, being generated during normal walking conditions; thus indicating normal plantarflexion timing. Training of the classification algorithm was done in between sessions 1 and 2. More detailed explanations of session 1, training, and session 2 are described below.First, to record normative ankle angle using motion capture (Vicon), each subject walked on a treadmill for 30 s while wear