Multiclass real-time intent recognition of a powered lower limb prosthesis

Abstract

This paper describes a control architecture and intent recognition approach for the real-time supervisory control of a powered lower limb prosthesis. The approach infers user intent to stand, sit, or walk, by recognizing patterns in prosthesis sensor data in real time, without the need for instrumentation of the sound-side leg. Specifically, the intent recognizer utilizes time-based features extracted from frames of prosthesis signals, which are subsequently reduced to a lower dimensionality (for computational efficiency). These data are initially used to train intent models, which classify the patterns as standing, sitting, or walking. The trained models are subsequently used to infer the user's intent in real time. In addition to describing the generalized control approach, this paper describes the implementation of this approach on a single unilateral transfemoral amputee subject and demonstrates via experiments the effectiveness of the approach. In the real-time supervisory control experiments, the intent recognizer identified all 90 activity-mode transitions, switching the underlying middle-level controllers without any perceivable delay by the user. The intent recognizer also identified six activity-mode transitions, which were not intended by the user. Due to the intentional overlapping functionality of the middle-level controllers, the incorrect classifications neither caused problems in functionality, nor were perceived by the user.

Fig. 1

Powered prosthesis control structure.

Fig. 2

The state chart depicting the phase transitions for standing, walking and sitting modes.

Fig. 3

Self-contained powered knee and ankle transfemoral prosthesis.

Fig. 4

Block diagram of the activity mode intent recognizer.

Fig. 5

Demonstration of the controller and real mode discrepancy during mode transitions.

Fig. 6

Controller mode switching logic for gait mode intent recognition.

Fig. 7

PCA (left) and LDA(right) dimension reduced features extracted from 200 sample-long frames.

Fig. 8

Classification performance of different model order GMM’s with PCA and LDA dimension reduction to one (a), two (b) and three (c) dimensions for frame length. Note the y-axis scaling for each dimension.

Fig. 9

Gaussian Mixture Models surface plots for standing, walking and sitting showing the portions of the feature space, where probability density function is greater than 0.05, for three dimensional LDA reduced data.

Fig. 10

Real-time activity mode switching (top), knee angle (middle) and ground reaction force (bottom) for a 190 seconds long trial. A video corresponding to this trial is included in the supplemental material.

Fig. 11

Knee angle (top) and real-time activity mode switching (bottom) for a 90 seconds sit to stand and stand to sit trial.