Machine Learning for Active Gait Support with a Powered Ankle Prosthesis

In recent years active prosthetic devices emerged as research area. In comparison to passive devices, active ones mimic the motion of our intact limbs more perfectly. Hence, they overcome several drawbacks a user of a passive devices is confronted with. Examples for benefits caused by switching from a passive to an active prosthesis are a more normal gait and reduced metabolic costs of transport. Even though active devices can improve amputee’s every day life, there are still challenges to be solved. Prosthesis control is besides design one of those challenges. The controller needs to decide when and how to adapt active support. Regarding, for instance, the push off phase, the active push off must be aligned with the timing and strength intended by the user. This work introduces recent supervised machine learning methods such as Gaussian process regression and support vector machines for control of active prosthetic devices. The machine learning methods are used to form a supervisory controller that infers the user’s intent. As input to the supervisory controller, we use the data obtained by an inertial measurement unit mounted at the shank of the considered active ankle prosthesis. The output or rather the users intent is given by gait, speed and gait percent predictions. If the intent is known, the desired nut position can be determined by a lookup. To enforce the desired trajectory a slave controller, here a PD controller, is applied. The supervisory controller is designed, implemented and tested based on walking and running data recorded on a treadmill. At first, noise free motion capturing data is used to demonstrate the applicability of supervised machine learning methods in context of active ankle control. Afterwards, sensor data obtained with the prosthesis’s intertial-measuremt unit is used to proof real-world applicability of the introduced supervisory controller. The obtained supervisory controller is fast and moreover accurate. For gait percent prediction the error is bounded to ±5%. In case of speed and gait prediction, accuracies close to 100% and 95% are achieved, respectively.