Artif Life Robotics (2009) 13:483–487 DOI 10.1007/s10015-008-0607-4
© ISAROB 2009
ORIGINAL ARTICLE Shuji Kawano · Dai Okumura · Hiroki Tamura Hisasi Tanaka · Koichi Tanno
Online learning method using support vector machine for surface-electromyogram recognition
Received and accepted: July 10, 2008
Abstract Research surface electromyogram (s-EMG) signal recognition using neural networks is a method which identifies the relation between s-EMG patterns. However, it is not sufficiently satisfying for the user because s-EMG signals change according to muscle wasting or to changes in the electrode position, etc. A support vector machine (SVM) is one of the most powerful tools for solving classification problems, but it does not have an online learning technique. In this article, we propose an online learning method using SVM with a pairwise coupling technique for s-EMG recognition. We compared its performance with the original SVM and a neural network. Simulation results showed that our proposed method is better than the original SVM. Key words Surface electromyogram · Support vector machine · Neural network · Pattern classification problem
1 Introduction Surface electromyogram (s-EMG) signals are detected over the skin surface and are generated by the electrical activity of the muscle fibers during contraction.1 A muscle movement can be presumed by analyzing a s-EMG. Therefore, s-EMGs are used to control an artificial leg, etc. s-EMG recognition using a conventional neural network is a method which identifies the relation between s-EMG patterns which are reproduced using a neural network. However, it not sufficiently satisfying for the user because s-EMG signals change with muscle wasting or with changes in the electrode position, etc. S. Kawano (*) · D. Okumura · H. Tamura · H. Tanaka · K. Tanno Faculty of Engineering, University of Miyazaki, 1-1 Gakuken kibanadai-nishi, Miyazaki 889-2192, Japan e-mail:
[email protected] This work was presented in part at the 13th International Symposium on Artificial Life and Robotics, Oita, Japan, January 31–February 2, 2008
The support vector machine (SVM) proposed by Cortes and Vapnik2 is one of the most influential and powerful tools for solving classification and regression problems.3,4 but it does not have an online learning technique. Online learning techniques for the SVM have been proposed by many researchers (e.g., Ogura and Watanabe5), but these techniques needed longer computing times. Here, we propose a simple online learning method using SVM for sEMG recognition. We compare its performance with the original SVM and neural network, and also test our propose method of online s-EMG recognition.
2 Review of SVM The SVM is a mechanical learning system that uses a hypothetical space of linear functions in a high-dimensional feature space. The simplest model is called linear SVM, and this works for data that are linearly separable in the original feature space only. In the early 1990s, nonlinear classification with the same procedure as linear SVM became possible by introducing nonlinear functions called Kernel functions without being conscious of the actual mapping space. This extended techniques of nonlinear feature spaces called nonlinear SVM. Assume the training sample S = ((xi,yi), . . . ,(xi,yi)) consisting of vectors xi ∈ R with i = 1, . . . , N, and each vector xi belonging to either of the two classes. Therefore it is given a label yi ∈ {−1,1}. The pair of (w,b) defines a separating hyperplane of the equation as follows: ( w, x ) + b = 0
(1)
However, it is possible that Eq. 1 can separate any part of the feature space, and therefore one needs to establish an optimal separating hyperplane (OSH) that divides S without affecting anything else. Points of the same class are accumulated on the same side while maximizing the margin, which is the distance of the closest point from S. The closest vector xi is called the support vector, and the OSH (w′,b′) can be determined by solving an optimization problem. The
484
solution of this optimization problem is given by the saddle point of the Lagrangian. 1 ( w, w ) 2 yi ((w ⋅ x i ) + b) ≥ 1
Maximize margin
To solve the case of nonlinear decision surfaces, the OSH is found by nonlinearly transforming a set of original feature vectors xi into a high-dimensional feature space by mapping Φ : xi → zi and then performing the linear separation. However, it requires an enormous computation of inner products (F(x) · F(xi)) in the high-dimensional feature space. Therefore, using a Kernel function which satisfies Mercer’s theorem given in Eq. 2 significantly reduces the calculations to solve the nonlinear problems. Here we used the Gaussian kernel given in Eq. 3 as the kernel function, the SVM decision function g(x), and the output given in Eqs. 4 and 5. (F(x) ⋅ F(x)) = K (x, x i )
(2)
⎛ − x − xj ⎞ K (x, x j ) = exp ⎜ 2 ⎟ ⎝ 2σ ⎠ 2
(3)
N
g(x) = ∑ wi K (x i, x) + b
(4)
O = sign( g(x))
(5)
i=0
3 Proposed method In this section, we propose a method of decreasing the deviation of SVM by an online learning method for the s-EMG recognition problem. It is widely recognized that SVM is a binary patternrecognitionmachine. However, s-EMGrecognition problems need a multiclass pattern recognition machine. Therefore we extended the standard SVM based on a pairwise coupling method to make it into a multiclass pattern recognition problem for s-EMG recognition. The number of SVMs necessary for an n-pattern classification, N, is defined by Eq. 6. N=
n(n − 1) 2
We then conducted two experiments to confirm the effectiveness of the proposed method. The construction of the proposed s-EMG pattern recognition system is shown in Fig. 2. The system consists of an input part, a feature extraction part, and a learning discrimination part. The s-EMG of each movement pattern is measured with electrode sensors, and the feature quantity is extracted from the s-EMG. The feature quantity is given to the recognition machine as an input, and each movement pattern that generates an s-EMG is recorded. The feature quantity uses both the maximum and minimum values of the s-EMG signal.6,7 It was shown that the techniques in these articles are easier and superior
(6)
The SVM produces a large margin in a high-dimensional feature space. However, many experimental results have shown that the boundary line created by the SVM has deviFig. 2. The structure of the s-EMG recognition system. The s-EMG of each movement pattern is measured with electrode sensors, and the feature quantity is extracted from the sEMG. The feature quantity is then given to the recognition machine (SVM)
4 Experimental method
Fig. 1. Online learning method. Here t(x), which is supervised information, and g(x), which is of the class (+1 or −1) to which input xi belongs, are the outputs of the SVM to input xi. η is an update parameter
Feature Extraction part
Input part
Pattern Classification and SVM Online Learning part
Maximum Motion
s-EMG Signal Amplifier
Volt
Subject
ations. Therefore, adjustment of the SVM is required. In our proposed online learning algorithm, supervised input data are considered for updating the threshold b. Threshold b is calculated by solving the Lagrange multiplier equation. We used a method of adjusting threshold b of the SVM (Eq. 4) by additional online learning. Threshold b of the SVM is adjusted using the rule in Fig. 1 when the SVM misclassifies the supervised input xi. Here t(x), the supervised information, and g(x), of the class (+1 or −1) to which input xi belongs, are the outputs of the SVM to input xi. The proposed method adjusts the threshold b of the SVM so that it can be set to the desired output of the SVM about the supervised input xi. The equation in Fig. 1 has the action of making threshold b small if the desired output of SVM becomes −1. Moreover, the equation in Fig. 1 enlarges threshold b if the desired output of SVM becomes +1. Parameter η is an update parameter. Here, we used parameter η = 0.1.
Time Minimum
b
485
4.1 Experiments on muscle wasting We then tested the effectiveness of the proposed method of s-EMG recognition by considering the problem that the feature quantity changes in muscle wasting. We compared the performance of the proposed method with the original SVM and a neural network. The neural network used was the multilayer perceptron (MLP), which is composed of three layers. Moreover, a back-propagation algorithm was used for learning. The experimental subjects were two healthy men (SK and DO) in their twenties whose physiques looked alike. The subjects sat on a chair. The recognition experiments for 6 movements in the pattern of the fingers were conducted using the results from s-EMGs obtained from two sensors strapped to the wrist of the right hand (Fig. 3). The experimental method first acquires the training data from an s-EMG monitoring the movement of each finger. Next, the SVM and MLP learn the relation between the s-EMG and the movement from the training data. Each movement is identified 20 times with these recognition machines. The subjects train for 10 min while seeing the recognition results on the display. Afterward, additional supervised learning is done nine times for each motion. The experiment repeats the measurement nine times (recognition rate calculation). As a result, muscle wasting is caused artificially and the feature quantity is changed. The total experimental time was about 3 h. Figures 4 and 5 show the recognition results. These results showed that our proposed method is better than the original SVM, and that it gave the best results for both SK and DO.
100 95 90 85 80 75 70 65 60 55 50 1
SVM proposed method MLP 2
3
4
5
6
7
8
9 10 Trial
Fig. 4. Experimental results for subject DO. The vertical axis is the recognition rate. The horizontal axis is the number of times we compared the recognition rates of the original SVM (diamonds), the proposed method (triangles), and the MLP (circles)
Recognition result(%)
to fast Fourier transform (FFT) processing. The sampling frequency of the measurement data was 1 KHz, and the range was from 0 KHz to 500 KHz.
Recognition result(%)
Fig. 3. Images of finger movements. Six patterns of movements of the fingers were recorded using the s-EMGs obtained from two sensors strapped to the wrist of the right hand
100 95 90 85 80 75 70 65 60 55 50
SVM proposed method MLP 1
2
3
4
5
6
7
8
9
10
Trial 4.2 Experiments on the electrode position gap We also investigated the effectiveness of the method on the problem posed by the s-EMG recognition that the features
Fig. 5. Experimental results for subject SK. The vertical axis is the recognition rate. The horizontal axis is the number of times we compared the recognition rates of the original SVM (diamonds), the proposed method (triangles), and the MLP (circles)
486
100 95 90 85 80 75 70 65 60 55 50
MLP
83.4
SVM
Proposed Method
85.7 85.2 86.5
85.9 80.8
80.0
79.7 71.0
First day
Second day
Third day
Fig. 7. Averages of the recognition results. The vertical axis is the recognition rate. The horizontal axis is the number of days. We compared the recognition rates of the original SVM (hatched), the proposed method (black), and the MLP (gray)
quantity changes according to the gap in the electrode position. We compared the performance of the proposed method with that of the original SVM and MLP. The experimental subject was a healthy man (SK). A recognition experiment using the six movement patterns was conducted by using the results from s-EMGs obtained from four sensors attached to the right arm (Fig. 6). The experiments were conducted for 3 days. The purpose was to move the position of the sensor artificially and to change the quantity of the feature. The subject trained for a few minutes while seeing the recognition results on the display. Afterward, additional supervised learning was carried out nine times for each motion. The experiment repeated the measurement nine times (recognition rate calculation). When the experiment was finished, subject SK removed the sensor. The SVM constructed using the training data from the first day was used in the experiments on the second and third days. From the experimental results shown in Fig. 7, it can be seen that our proposed method is better than that of the original SVM, and that the proposed method showed the best performance on SK. Figure 8 shows the recognition results for the second day.
Recognition result(%)
Recognition result(%)
Fig. 6. Images of forearm movements. Six patterns of movement of the forearm were recorded using the s-EMGs obtained from four sensors strapped to the right arm
100 95 90 85 80 75 70 65 60 55 50
SVM proposed method MLP 1
2
3
4
5
6
7
8
9 10 Trial
Fig. 8. Experiment results of the gap of the electrode position (second day). The vertical axis is the recognition rate. The horizontal axis is the number of times we compared recognition rates of the original SVM (diamonds), the proposed method (triangles), and the MLP (circles)
5 Results The average results for subject DO in the muscle wasting experiment were 75.9% for the SVM, 77.3% for the MLP, and 83.6% for the proposed method. The average results for subject SK in the muscle wasting experiment were 75.5% for the MLP, 82.8% for the SVM, and 87.1% for the proposed method. We compared these results with those found using the nearest neighbor method. The result using the nearest neighbor method for subject DO was 71.5%, and that for subject SK was 65.2%. The average results of the electrode position gap experiment were 82.9% for the MLP, 79.0% for the SVM, and 84.1% for the proposed method. These results show that the recognition rate of the SVM improved after a period of online learning.
487
technology described in this article. The robot control experiments could be performed while the robot moved in four different directions from an s-EMG of a human face. In future work, we hope to (1) propose a technique of unsupervised learning, (2) miniaturize our system using hardware developments, and (3) apply the results in medical fields. Acknowledgment This research was supported by the Tateishi Science and Technology Foundation, (No.1081019).
References
Fig. 9. Robot controls. We have succeeded in performing robot control using the technology described in this article
6 Conclusion We have proposed an online learning method using a SVM for s-EMG recognition problems. The experimental results showed that the proposed method was effective in overcoming the s-EMG recognition problem. In this application, we succeeded in performing robot control (Fig. 9) using the
1. Tomohiro K, Tadashi M, Tohru K, et al (2006) Practical usage of surface electromyogram. Biomechanism Library 2. Cortes C, Vapnik VN (1995) Support vector networks. Mach Learn 20:273–297 3. Burges CJC (1998) A tutorial on support vector machine for pattern recognition. Data Mining Knowledge Discovery 2(2):121–167 4. Cristianini N, Shawe-Taylor J (2000) An introduction to support vector machines and other kernel-based learning methods. Cambridge University Press, Cambridge 5. Ogura N, Watanabe S (1999) Online learning for support vector machine using necessity estimation (in Japanese). Technical Report IEICE PRMU, vol 99, No, 182, pp 45–52 6. Tamura H, Okumura D, Tanno K (2007) A study of motion recognition without FFT from surface-EMG (in Japanese). J IEICE D J90-D(9):2652–2655 7. Okumura D, Tamura H, Tanno K (2007) Proposal of the motion recognition system corresponding surface-EMG changes (in Japanese). FIT2007 (CD-ROM)