Underwater sEMG-based recognition of hand gestures using tensor decomposition

Highlights

• Acquire sEMG signals underwater based on waterproof the electrode.

• Generate a four-dimensional tensor by wavelet transform.

• Extract features of underwater signals using tensor decomposition.

Abstract

Amputees have limited ability to complete specific movements because of the loss of hands. Prosthetic hands can help amputees as an effective human-computer interaction system in their daily lives, and some amputees need to use the prosthetic hands for underwater operations. Therefore, it is necessary to solve the problem of using prosthetic hands underwater. There are two main problems in underwater surface Electromyogram (sEMG) signal recognition. The underwater sEMG signals are disturbed by noise, and the traditional sEMG features are easily affected by noise, decreasing the recognition accuracy of underwater sEMG signals. It is difficult for subjects to acquire quantity training data underwater, and satisfactory sEMG recognition accuracy needs to be obtained based on small datasets. Tensor decomposition has the advantage of finding potential features of signals, and it is widely used in many fields. Tucker tensor decomposition was used for feature extraction and recognition of underwater sEMG signals. Seven subjects were selected to complete four hand gestures underwater and two-channel sEMG signals were collected. Wavelet transform was applied to generate a three-dimensional tensor and extracted signal features by tensor decomposition. The recognition accuracy based on K-Nearest Neighbor reaches 96.43%. The results show that the proposed sEMG feature extraction method based on tensor decomposition helps improve the recognition accuracy of underwater sEMG signals, which provides a basis for applying prosthetic hands in a water environment.

Introduction

The hand plays an essential role in human interaction with the external environment. The hand can sensitively perceive changes in environmental information, and it can convey human emotions through different gestures. More importantly, the hand can perform many delicate operations, such as grasping, pinching, and pushing. Amputees cannot undertake some daily living activities due to the loss of the hand, which seriously affects their quality of life and has harmful impacts on their lives and work. Amputees can improve their self-care ability with the help of prosthetic hands [1].

The surface Electromyogram (sEMG) signal is the bio-electrical signal that accompanies muscle contraction, and it reflects the level of muscle activity [2]. The sEMG signal can be obtained by attaching non-invasive electrodes to the skin. The sEMG signal is widely used in the clinical and laboratory. The sEMG signal can be used to monitor human activity for the care of elderly persons [3]. It is also important for athletes to improve their performance based on the sEMG. The sEMG signal can be used to evaluate the rehabilitation level of patients [4].

The sEMG signal of the forearm can be used to recognize hand gestures, which is widely used in human-computer interaction. For amputees, the sEMG signal of the forearm can also reflect their hand movement intention. The sEMG signal is ideal control signals for artificial prosthetic hands. Amputees can control prosthetic hands based on the sEMG signal according to their movement intention, and they can use prosthetic hands to replace part of hand functions in daily life [5].

Amputees need to use the prosthetic hands for a long time in their daily life, which requires the prosthetic hands could be used underwater. However, there are currently two problems in the application of prosthetic hands underwater. The traditional feature extraction methods are unsuitable for sEMG signal collected underwater and have adverse effects on recognizing underwater sEMG signal. It is difficult for amputees to collect quantity sEMG signals, which leads to the limitation of the size of the dataset and has an impact on the recognition of hand gestures. It is necessary to propose a novel feature extract method.

Currently, the Ag/AgCl electrode is widely used in laboratory and clinical, and it is considered the gold standard in collecting sEMG signal [6]. The Ag/AgCl electrode is a kind of wet electrode. It has a hydrogel layer at the electrode-skin interface, which reduces the skin-electrode impedance to improve the signal quality [7]. The flexible fabric electrode is also used to collect sEMG signal. For instance, a silver-plated knitted electrode is proposed for the user’s comfort in daily life [8]. Traditional sEMG electrodes are not suitable for underwater sEMG signal acquisition. Our existing research proposed a flexible waterproof electrode based on conductive silicone to collect sEMG signals underwater [9]. Our existing research has verified the underwater sEMG signal acquisition performance of the proposed electrode [10].

It is meaningful to obtain helpful information of raw sEMG through feature extraction [11]. The sEMG features mainly include time-domain, frequency-domain, and time-frequency features [12]. Krasoulis et al. used the time-domain features, which were the Mean Absolute Value (MAV), Waveform Length, 4th-order Auto-Regressive coefficients, and Log-Variance [13]. Zhang obtained four features to characterize the sEMG signals, and the features were MAV, Standard Deviation (SD), SD of the frequency-domain, and Wavelet Transform coefficient [14]. Pizzolato et al. extracted five features for signal recognition: Root Mean Square (RMS), Histogram features, and other features [15].

There are many studies on sEMG signal recognition. Samuel et al. proposed three novel time-domain features. The recognition accuracy of upper-limb motions based on Linear Discriminant Analysis (LDA) reaches 92.00% $\pm$ 3.11% [16]. Hu et al. introduced a novel attention-based hybrid Convolutional Neural Network and Recurrent Neural Network model. The sEMG recognition accuracy based on the proposed method is higher than that based on the state-of-the-art method [17]. Shi et al. used K-Nearest Neighbors (KNN) for hand gesture online recognition [18]. Qi et al. used the General Regression Neural Network for nine hand gesture recognition [19].

Tensor decomposition can extract potential features of signals, and it is an important tool for multi-dimensional data analysis. It has been widely used in data mining, signal processing, image recognition, and other fields [20]. Currently, commonly used tensor decomposition methods include CANDECOMP/PARAFAC (CP) decomposition [21]and Tucker decomposition [22].

To improve the recognition accuracy of underwater sEMG signals, this paper proposes a feature extraction method of sEMG signals through Tucker decomposition. The contributions of this paper mainly include the following three aspects:

• Propose the feature extraction method based on tensor decomposition and a method for determining the size of the core tensor;

• Compare the recognition accuracy of hand gestures based on tensor decomposition and that based on the traditional sEMG features in the normal environment;

• Prove that the features based on Tucker decomposition can effectively improve the recognition accuracy of underwater hand gestures.

The structure of this paper is as follows. Section 2 introduces the process of the sEMG signal acquisition. Section 3 illustrates the method of feature extraction and recognition. Section 4 presents the experimental results. Section 5 discusses the experimental results. Finally, Section 6 summarizes the conclusions and future work.