Toward design of an environment-aware adaptive locomotion-mode-recognition system

Abstract

In this study, we aimed to improve the performance of a locomotion-mode-recognition system based on neuromuscular-mechanical fusion by introducing additional information about the walking environment. Linear-discriminant-analysis-based classifiers were first designed to identify a lower limb prosthesis user's locomotion mode based on electromyographic signals recorded from residual leg muscles and ground reaction forces measured from the prosthetic pylon. Nine transfemoral amputees who wore a passive hydraulic knee or powered prosthetic knee participated in this study. Information about the walking terrain was simulated and modeled as prior probability based on the principle of maximum entropy and integrated into the discriminant functions of the classifier. When the correct prior knowledge of walking terrain was simulated, the classification accuracy for each locomotion mode significantly increased and no task transitions were missed. In addition, simulated incorrect prior knowledge did not significantly reduce system performance, indicating that our design is robust against noisy and imperfect prior information. Furthermore, these observations were independent of the type of prosthesis applied. The promising results in this study may assist the further development of an environment-aware adaptive system for locomotion-mode recognition for powered lower limb prostheses or orthoses.

Fig. 1

(a) Mechanical design of a powered knee. (b) Experimental setup of a TF amputee wearing the prototype.

Fig. 2

Definitions of evaluation and optimization parameters for the locomotion-mode-recognition system. An example of results of continuous gait-phase detection and locomotion-mode recognition is shown. Red vertical line denotes the critical time T_c . T_{p r e} indicates the prediction time for the transition from level ground walking to stair ascent. Definitions of two optimization parameters in transitions, i.e., error pulses (#1-#2) and duration of error pulse, are also demonstrated.

Fig. 3

Influence of the objective function on the optimized prior probability models and their performance for locomotion-mode recognition in static states. (a) Effect of the optimization threshold c in the objective function (13) on the a value in the prior probability model. (b) Comparison of the accuracy for classifying level-ground walking without using prior knowledge, with correctly applied optimized prior knowledge model, and with incorrectly applied prior models when the value of c varies. (c) Influence of c on the P-value in a one-way ANOVA of classification accuracy. The P-values in gray lines were derived from comparison of the accuracy without using prior knowledge (uniform prior probability) to the accuracy with incorrectly applied prior knowledge. The P-values in black lines were derived from comparison of the accuracy without using prior knowledge to the accuracy with correctly applied prior knowledge. The statistical analysis was conducted on the classification accuracy of each class obtained from all the subjects. P < 0.05 means a statistically significant difference between studied conditions. P = 0.05 is highlighted in the plot.

Fig. 4

Influence of parameters b and d in the prior probability model for transitions on (a) the number of error pulses and (b) the longest duration of error pulses. The presented number of error pulses and the longest duration of error pulses were averaged over subjects.

Fig. 5

Classification accuracy for each locomotion mode in static states averaged over ten experiments. The average classification accuracy was derived from the LDA classifier without prior knowledge of walking environment (black bars), with correct prior knowledge (gray bars), or with incorrect prior knowledge (white bars). * indicates a statistically significant difference (one-way ANOVA, P < 0.05). “Red crosses” denote the classification accuracy derived from TF09 when donning the powered device.