An Experimental Study of EMG Signal Features for Motion Discriminations Using Support Vector Machine

This paper treats a discrimination problem of wrist/hand motion patterns from EMG signal. We examined which of the following signal features was appropriate: raw signal, integrated signal (IEMG), the max frequency component, power spectrum or rising voltage level. For the discrimination algorism, a Support Vector Machine (SVM) was introduced. As a result, around 80% discrimination rate was accomplished from integrated signal, power spectrum and rising voltage level. The IEMG signal scored the highest 83.3% discrimination rate.


Introduction
The electromyographic (EMG) signal is considered to be applicable for the control of prosthetic arms and legs [1] [2].It has many advantages: Surface-EMGs are non-invasively measurable by attaching electrodes to a person's skin.Moreover, exerted force information can be estimated from EMG signals.This makes the EMG signal a potential candidate for the control of prosthetics.A surface-EMG consists of a set of complex signals from many muscle fibers around the electrode and is affected by many different muscle fibers [2].Thus the action of an individual muscle from the surface-EMG is difficult to discriminate.Furthermore, the voltage level of an EMG signal is very low and so easy to be effected from noise.To solve these problems, EMG signals and their application have been studied from many aspects [3][4].This study aims at establishing a new human interface using some forearm EMG signals so as to control mainly robotic devices, especially a hand rehabilitation support system we are developing [7].This paper investigates which of the following signal features is more effective to discriminate finger and wrist motions: raw signal, integrated signal, the max frequency component, power spectrum, or rising voltage level, As an algorithm to discriminate the hand motion, we adopt the support vector machine because of its high classification performance.

Purpose
EMG signals would contain efficient information for the human motion, because it affects directly to human motion actuators, i.e., muscles.However, the raw signal contains noises as well as different components from the target muscle, and its direct, i.e., non-processed, usage to the discrimination requires much computational resources due to its high-dimensionality.Therefore, some pre-processing should be introduced before being utilized for motion discrimination.The hand rehabilitation system we have been developing [7] requires the detection of the following eight motions to assist the hand motion: pronation/supination and dorsal/volar flexion of the wrist, thumb extension/flexion and four fingers extension/flexion.These motions are generated mainly from the following eight muscles: pronator teres, spinatior, extensor/flexor carpi ulnaris, flexor/abductor pollicis longus, flexor digitorum superficialis and extensor digitorum.Thus we tried to discriminate motions by attaching the electrodes around these muscles.
The purpose of this paper is to investigate what kind of EMG signal features, obtained from the electrodes targeting to the above eight muscles and then be pre-processed, is effective to discriminate the above eight motions.As the feature of the signals, we choose the following five, integrated signal, the max frequency component, power spectrum and rising voltage level: how these features are obtained is briefly explained below.The sampling rate of our EMG measurement system was 3kHz.Ⅰ) Raw data (4800[dimension]) We detect the time max t when the total sum of all channel voltages is maximized.Then, the EMG signals 0.1sec (300[sample]) before and after max t are extracted, and normalized by the absolute value of its maximum voltage in each subject.Data dimension is 600[samples]*8[ch]=4800 [dimensions].This raw data is examined for comparison of other signals.

Ⅱ) IEMG (4800[dimension])
IEMG is an integration of raw data that are four times amplified after low-pass filtered with 4.8Hz cutoff frequency.They are automatically obtained from our EMG measurement systems.Afterwards, the same operation as I) is applied.Data dimension is 600 The 256 samples before and after max t are extracted from each channel.The frequency component in this time range is used as a signal feature, i.e., their power spectrum is calculated and used as the data for discrimination.Data dimension is 512 The largest component of power spectrum in each 8 ch is used for a signal feature.Data dimension is 8 The difference of average voltage between the rest state (1 second before motion: 3000[samples]) and activated state (1 second after motion: 3000[sample]) is calculated from each IEMG signal, and it divided by standard deviation.Number of dimension is 8 [dimension].

Methods
EMG signals were obtained from three male 22-24-year-old subjects total three times in the separate day.Eight electrodes are attached to the subject in the following muscles: pronator teres, spinatior, extensor/flexor carpi ulnaris, flexor/abductor pollicis longus, flexor digitorum superficialis and extensor digitorum.Fig. 1 shows a diagram of the measurement system.During EMG measurement, signal was recorded with a 3kHz-sampling rate.A 60Hz ham filter, 10Hz Low-pass filter and 100Hz High-pass filter were used to remove the noise.In each measurement, subjects were asked to perform 8 wrist/hand motions in 15 times respectively.The first 10 data were utilized for teaching data, while the last 5 data were for test data.For the computation of the SVM algorithm, SVM-perf and SVM-multiclass was utilized, which were developed by Cornell university computer science subject [8].The SVM-multiclass can classify the data into multiple categories, while the SVM-perf did only into two.So, we combined three SVM-perfs to discriminate 8 motions.Fig. 1 Measurement system.