Movements Classification of Multi-Channel sEMG Based on CNN and Stacking Ensemble Learning

Shu Shen,Ming Yang,Xin-Rong Chen,Kang Gu,Ru-Chuan Wang

doi:10.1109/access.2019.2941977

Abstract

In recent years, the analysis of surface electromyography (sEMG) signals by feature engineering and machine learning has developed rapidly. However, when feature engineering is applied to feature extraction of sEMG signals, important feature information in the signals will inevitably be omitted, which will reduce the performance of signal analysis and recognition. Therefore, this paper proposes a method to complete classification of sEMG hand movements based on convolutional neural network (CNN) and stacking ensemble learning. In this method, a primary classifier based on CNN is designed to extract sEMG data features, which avoid omission of important feature information. A secondary classifier based on the stacking method is designed to integrate three primary classifiers trained with time domain, frequency domain and time-frequency domain data of the sEMG signal respectively. Then, several experiments on NinaPro DB5 dataset is performed to evaluate the proposed models. When the window length is 200ms, primary classifier is trained and tested with the sEMG signal data divided by the 80ms, 100ms, and 125ms sliding length. The best accuracy can reach 71%. The primary classifier and the secondary classifier trained and tested with sEMG signal data divided by window lengths of 200ms and 300ms in the case of a sliding length of 100ms. When the window length is 200ms, the best primary classifier accuracy and the best secondary classifier accuracy can be 70.92% and 72.09%, respectively. On the window length of 300ms, the best primary classifier accuracy and the best secondary classifier accuracy can reach 75.02% and 76.02%, respectively. Finally, the model designed is compared with Linear Discriminant Analysis (LDA), Long Short Term Memory-CNN (LCNN), Support Vector Machine (SVM), and Random Forests. Under the same conditions, the average accuracy of the secondary classifier is 11.5%, 13.6%. and 10.1% higher than LDA, SVM, and LCNN, respectively. Also, the average accuracy rate is 3.05% higher than SVM and Random Forests.

Highlights

Electromyography (EMG) is a superposition of bioelectrical signals generated by the muscles of the human body. surface electromyography (sEMG) is a way to detect muscle activity from the surface of human skin [1]
When performing Discrete Fourier Transform (DFT), assuming that the sequence length of the current sEMG signal is N, we extend the length of the sequence to 2N by adding 0 at the end.The extended sequence is subjected to DFT, and the first N amplitude data of the transformed sequence is taken as a frequency domain representation of the signal
The combination of convolutional neural network (CNN) and Stacking ensemble learning is used for sEMG hand movements classification, which can overcome the limitation of the feature engineering requiring better feature quality

Summary

INTRODUCTION

Electromyography (EMG) is a superposition of bioelectrical signals generated by the muscles of the human body. sEMG is a way to detect muscle activity from the surface of human skin [1]. The results showed that the accuracy of calibrated classifier would be 10.18% (intact, 50 movement types) and 2.99% (amputee, 10 movement types) higher than uncalibrated classifier He et al [11] combined long-short-term memory networks and multi-layer perceptrons to classify sEMG signal in NinaPro DB1 dataset. Many works have shown that the accuracy of classifying sEMG signals using DNN is generally higher At this stage, the sEMG signal recognition based on deep learning model is hopeful to be improved in terms of the accuracy and the feature engineering complexity. We propose a multi-channel sEMG signal recognition model based on convolutional network (CNN) and Stacking ensemble learning.

METHODS

LOSS AND OPTIMIZATION

Findings

CONCLUSION