IEEE Access (Jan 2019)

Motion Estimation From Surface Electromyogram Using Adaboost Regression and Average Feature Values

  • Feiyun Xiao,
  • Yong Wang,
  • Liangguo He,
  • Hu Wang,
  • Weihan Li,
  • Zhengshi Liu

DOI
https://doi.org/10.1109/ACCESS.2019.2892780
Journal volume & issue
Vol. 7
pp. 13121 – 13134

Abstract

Read online

The method to estimate joint motion quickly and accurately from surface electromyogram (sEMG) has been explored by many researchers. However, the effect of different grabbing loads is ignored by most of them, which limits its clinical and daily applicability. In order to eliminate this effect in the course of motion estimation, an Adaboost-AF method based on Adaboost regression was proposed to identify the load information and estimate motion intention. This method is composed of three parts: the sEMG preprocessing part, the load identification part, and the motion intention estimation part. The average value of sEMG feature signal, which is obtained during sEMG preprocessing part, was used to identify the load information. Five features, root mean square (rms), waveform length, difference absolute standard deviation value, the integral signal of sEMG (IEMG), and low-pass filtered signal of sEMG (LPFEMG) similar to sEMG envelope were explored whether the average value of features signal can be used to identify the load under different speeds. After the load was identified, the Adaboost regression framework with the decision tree as the weak learner was applied to be the motion intention estimation part. Experimental results showed that the average rms difference between an actual angle and estimated angle by using the Adaboost-AF method was 0.1193 ± 0.0148, and the average execution time for a section of data with 17.58 s was 0.3932 ± 0.0610s. (Scikit-learn packet was adopted, and the version number was V0.21. The running software was Spyder, Python3.6.) This paper shows that the method based on the average value of sEMG feature signals can eliminate the influence of different loads without adding new sensors; IEMG and LPFEMG from the biceps brachii muscle show the optimal loads identification performance; the proposed Adaboost-AF method takes the influence of different loads into account and ensures the rapid and accurate estimation performance of motion intention.

Keywords