Development of Hybrid Ankle-Foot Prosthesis

Abstract

Limb loss of humans produces a permanent disability that impacts the amputee’s self-confidence, self-care, and limb movement. Specifically, lower limb loss results in slow and less stable locomotion. Below-knee (transtibial) amputation is one of the most frequent types of amputation around the world. The ankle-foot prosthesis is one of the solutions meant to provide support and assistance to the transtibial amputee. The majority of available ankle-foot prostheses are passive and not able to provide required push-off and seamless movement during locomotion. However, active ankle-foot prostheses can provide desired characteristics and seamless operation to achieve by embedding subject locomotion intention in the control sequence of the prosthesis. EMG signal is a well-reorganized way of capturing neural information of human limb movements to identify the locomotion intention of the subject. The prime objective of the present research work was to develop a hybrid ankle-foot prosthesis prototype. Here, the EMG signal of six lower limb below-knee muscles for non-weight bearing and weight-bearing dorsiflexion/plantarflexion ankle-foot movements had been analyzed. Tibialis Anterior and Gastrocnemius muscles had been found suitable for repetitive ankle-foot movements. Further, to incorporate subject locomotion intention in the control sequence of the prosthesis, hybrid information (EMG+knee joint angle) based activity and locomotion prediction module had been designed. It had a two-level classification approach, the first level of classification predicts the activity (standing or walking) of the subject, whereas the second level predicts the locomotion (level walking, stair ascent, stair descent, ramp ascent, and ramp descent) intention of the subject only if at level-1 walking activity had been identified. EMG signal of four lower limb muscles (Tibialis Anterior, Gastrocnemius, Vastus Lateralis, and Biceps Femoris) with knee joint angle signal were used in both levels of classification. A continuous windowing technique with window size 256 msec and window shift 32 msec had been applied at level-1 classification to monitor subject activity continuously. Whereas for level-2 classification, four subwindows, each of size 200 msec, had been extracted over the complete gait cycle, and an individual classifier was trained for each subwindow. The performance of three classifiers (LDA, SVM, and NN) had been compared for both levels of classification. The performance of the LDA classifier had been found better as compared to other classifiers for both levels of classification. iv After that, an ankle joint angle estimation module had been developed using the Non-negative Autoregressive Exogenous model as a training algorithm. The angle estimation module served as a low-level controller for the prosthesis and estimate ankle-foot angular position continuously using the EMG signal of two lower limb below-knee muscles (Tibialis Anterior, and Gastrocnemius muscles) and knee joint angle signal. An individual model had been designed for each locomotion. The performance of the estimation model had been found satisfactory for each locomotion, which justifies its applicability for ankle-foot prosthesis control. Moreover, the knee joint angle signal information had been found very useful, and its incorporation with EMG signal resulted in significant improvement in model performance. Finally, an electric actuator operated an ankle-foot prosthesis prototype, with one-DoF and dorsiflexion/plantarflexion movement had been developed. The estimated ankle joint angle was used to generate a control signal to operate the developed prototype in offline mode.