New EMG Pattern Recognition based on Soft Computing Techniques and Its Application to Control of a Rehabilitation Robotic Arm Jeong-Su Han*, Won-Kyung Song*, Jong-Sung Kim**, Won-Chul Bang*, Heyoung Lee***, and Zeungnam Bien* * Dept. of Electrical Engineering and Computer Science, KAIST 373-1 Gusong-Dong, Yusong-Gu, Taejon 305-701, KOREA ** Internet Appliance Department, ETRI, 161 Kajong-Dong, Yusong-Gu, Taejon 305-350, KOREA *** Dept. of Electrical Engineering, Chonnam National Univ. 300 Yongbong-Dong, Puk-Gu, Kwangju, 500-757, KOREA Tel.: +82-42-869-3419, Fax: +82-42-869-8410, e-mail:
[email protected] Key Words: EMG, Rehabilitation Robot, Pattern Recognition, Feature Extraction, Soft Computing, Rough Set, and Fuzzy Min-Max Neural Network Abstract – A new EMG pattern classification method based on soft computing techniques is proposed to help the disabled and the elderly handle rehabilitation robotic arm systems. First, it is shown that EMG is more useful than existing input devices, such as voice, a laser pointer, and a keypad in view of naturality, extensibility, and applicability. Next, through soft computing techniques, such as the fuzzy logic and rough set theory, a new procedure is proposed to select an essential feature set of EMG signals that is independent of users. In order to classify pre-defined motions, a fuzzy pattern classification and fuzzy min-max neural networks (FMMNN) are adopted to handle the selected minimal feature set in systematical ways. As results, motions are recognized with success rates of 83 percent and 90 percent for fuzzy pattern classification and FMMNN, respectively.
I. Introduction The number of the elderly in the world has been growing fast due to advanced medical technology. Most of the elderly, however, have several chronic diseases [1], and they have difficulties in daily living without others’ assistance. In addition, posteriori physically disabled people increase by industrial or traffic accidents. Thus, developing a system that assists humans for their incomplete activities and lost senses is strongly desirable, and several prototype systems have been developed [2], [3], [4]. In case of Korea, it has been developing a wheelchair-mounted rehabilitation service robot, called KARES (KAIST Rehabilitation Engineering Service system) [19]. We use an EMG (electromyogram) signal as a user’s input to a rehabilitation robotic arm. An EMG signal is more useful than existing input devices such as voice [2], a
laser pointer [3], a keypad [4], and a 3D input device [5] in the following views. First, it is more natural to use one’s arm to control a rehabilitation robotic arm than to use a voice and a head movement with a laser pointer. It gives the user a natural feeling of control as if his/her original limb is live. Second, learning complex system commands, such as syntax in voice commands, is not required. For example, MANUS evaluation users reported that one of difficult thing using rehabilitation robots stems from too many functions to keep in mind at the beginning [16]. Third, users who have difficulty in using a keypad, a joystick, and a 3D input device could command their intentions by EMG signals. Finally, EMG signals provide the rehabilitation robot with useful additional information, e.g., the moving speed [6] or fatigue [7] of the arm. This paper is organized as follows: a problem of EMG pattern recognition is described in Section II. A new EMG pattern classification method based on soft computing techniques is proposed in Section III. A minimal feature set is extracted to classify pre-defined motions. In Section IV, experiment results and discussion are presented. Finally, concluding remarks are given in Section V.
II. Problem Description The EMG signal is a form of electric manifestation of neuromuscular activation associated with a contracting muscle [17]. EMG signals have following characteristics [8], [9], [10]. First, an EMG signal is nonstationary. Second, the number of muscle fibers consisting of one single motor unit is different from person to person. Skins and other muscle tissues modify EMG signals waveforms.
The thickness of these tissues is not equal in all people. Therefore, people have different waveforms, and thus EMG signals are complex to analyze. Third, the amplitude of EMG signals is too small to be affected by noise in the environment. Several approaches [11], [12], [13], [14] have been proposed to solve the motion command identification problem using EMG signals for assistive devices such as powered prostheses. Although previous works have brought some sort of theoretical and practical achievements for powered prostheses, these have some problems. Major defect of previous works is the dependency among users. It uses different threshold to classify the motions for each user [12], and learning process is needed before classification. On-line learning method was proposed to overcome the user dependency [10], but it did not mention which features should be trained to minimize the effect caused by different users. In this paper, a new EMG pattern classification method based on soft computing techniques is proposed. A method to select minimal feature set for minimizing the user dependency is proposed. We have also designed the fuzzy pattern classifier and Fuzzy Min-Max Neural Networks (FMMNN) for classifying eight primitive motions as shown in Fig. 1. These motions are basically necessary for the disabled, especially, people with spinal cord injury (C-5, C-6, and C-7 lesion), and the elderly for enjoying daily life.
Fig. 1: Primitive motions of a user. From left to right and top to bottom, wrist pronation (motion 1), wrist supination (motion 2), elbow flexion (motion 3), elbow extension (motion 4), humeral moving up (motion 5), humeral moving down (motion 6), humeral moving in (motion 7), humeral moving out (motion 8).
III. EMG Pattern Classification Based on Soft Computing Techniques 1. The Location of Channels for Collecting EMG Signals If many channels for measuring EMG signals are used, they give more information to classify motion commands but make the user uncomfortable. Therefore, it is very important to set the position of channel with minimum number of channels for classifying the predefined motions.
Primitive motions are divided into four classes according to rotation axes in the arm as shown in Fig. 1: class 1 is wrist pronation (motion 1) and wrist supination (motion 2), class2 is elbow flexion (motion 3) and elbow extension (motion 4), class3 is humeral moving up (motion 5) and humeral moving down (motion 6), and class4 is humeral moving in (motion 7) and humeral moving out (motion 8). Each motion in each class is separated easily using the fact that skeletal muscle in the body composed of an antagonistic muscle. Therefore, the classification problem of eight primitive motions changed as a classification problem of four classes and four channels are needed for classifying eight primitive motions. Fig. 2 shows the location of channels that are selected by experiments.
Fig. 2: The locations of channels to measure EMG. Channel 1 is pronator teres muscle (From 1a to 1d are candidates for electrode position of channel 1), channel 2 is brachialis muscle (From 2a to 2d are candidates for electrode position of channel 2), channel 3 is deltoid muscle, and channel 4 is levator scapulae muscle (From 3a to 3d are candidates for electrode position of channel 3 and 4).
2. Features for Pattern Recognition It is very important to define or select appropriate features of EMG signals in pattern classification problems [15]. Although there are various features for EMG pattern classification problems [6], [8], [12], choosing suitable features is not easy with respect to tasks or motions. According to analyses of previous works [6], [8], integral absolute value (IAV), average slope of IAV (ASIAV), variance (VAR), and zero crossing (ZC) are elected to abstract the EMG pattern. Frequency ratio (FR) is proposed as a new feature in this paper to distinguish the difference between contraction and relaxation of a muscle in frequency domain. By applying FFT to EMG signals in time domain, we calculate the ratio
between the low frequency components and the high frequency components of EMG signals. FR j =
F (⋅) j low freq.
F (⋅) j high freq.
=
F (⋅) j 30Hz- 250Hz F (⋅) j 250Hz- 1kHz
3. Proposed Minimal Feature Set Selection Algorithm Step #1. Data acquisition: EMG signals of the primitive motions are measured from the four predetermined muscles (channels).
(1)
where, F (⋅) is the FFT of EMG signal in channel j , the j
Step #2. Preprocessing: A second order high pass butterworth filter with 30 Hz as a cut-off frequency is used to eliminate low frequency noise such as motion artifacts.
threshold for dividing a low frequency band and a high frequency band is decided through experiments. Fig. 3 and Fig. 4 show EMG signals of motion3 (elbow flexion) and motion4 (elbow extension) in time and frequency domain, respectively. In time domain, the amplitude of EMG signals in channel 1 and 2 are bigger than those of the other channels. Because EMG signals are generated in proportion to the degree of contraction of the muscle, it was found that muscles in channel 1 and 2 contract more than others. In frequency domain, the amplitude of low frequency parts in channel 1 and 2 is big. Therefore, the high frequency ratio in Eq. (1) means that the degree of contraction of the muscle is big. Thus we can identify the degree of contraction of the muscle by means of the frequency ratio.
Step #3. Feature extraction: IAV, VAR, ZC and FR are extracted from preprocessed EMG signals. ASIAV is excluded from extracting minimal feature set for classifying four classes because ASIAV only need for classifying each motion in each class. Step #4. Feature clustering: This step gives the user dependent feature sets and consists of following three sub-steps: (a) Normalization: Feature values for the primitive motions are normalized by channel using Eq. (2). X k’ =
Xk N
∑X
(2) i
i =1
where, X i is a feature value in each channel, and N is the number of channels. (b) Linear Mapping: Normalized feature values are mapping linearly using Eq. (3). All normalized feature values are in [0 1] after linear mapping; therefore it is easy to design fuzzy pattern classifier. X k’ − ( X min )k ’
Yk =
Fig. 3: EMG signals of elbow flexion/extension in the time domain.
( X max )’k − ( X min )’k
(3)
where, X k’ is a normalized feature value in channel k , ( X min )’k and ( X max )’k are the minimum and maximum value
Fig. 4: EMG signals of elbow flexion/extension in the frequency domain.
in channel k before linear mapping, respectively. (c) Clustering of Features for Each Subject: Mean values of features are clustered into two, three, and four groups through fuzzy c-means algorithm. Because each motion in each class (class1 (motion1 and motion2), class2 (motion3 and motion4), class3 (motion5 and motion6), class4 (motion7 and motion8)) is separated by ASIAV, the classification problem of eight primitive motions is reduced to a classification problem of four classes. Therefore, we set limits to the number of clustering is four. Clusters are found systematically by fuzzy c-means algorithm. A justification to use the fuzzy c-means algorithm is to reduce the heuristic factor of grouping feature values. This step is repeated to the features from
each subject and each channel. The clusters only classify for a specific subject, and thus common cluster sets are desired. Step #5. Common cluster set selection: This step is required to make the feature sets have independency among subjects. A common cluster set is the set of clusters that are in the same membership function for four subjects. The overall number of tables via clustering results is 16 (= 4 channels * 4 features) including Table 1. The common cluster sets are made from the clustering results by assigning attribute value (Table 2). The procedure for Table 2 is following. The method to assign condition attributes is to assign the same value in the same cluster. FR in channel 1 for all subjects was clustered two groups such as {motion1, motion2, motion3, motion4} and {motion5, motion6, motion7, motion8}. Therefore, we assign the attribute value, Big, to {motion1, motion2, motion3, motion4} and another attribute value, Small, to {motion5, motion6, motion7, motion8} to consider the distribution in FR in channel 1. It means that FR in channel 1 classifies the primitive motions into the two groups: {motion1, motion2, motion3, motion4} and {motion5, motion6, motion7, motion8}. {{motion1, motion2}, {motion5, motion6}} or {{motion1, motion2}, {motion7, motion8}} are the only common cluster set in the 3 group clustering results of FR in channel 1. Because subject #3 has {motion1, motion2, motion3} rather others have {motion1, motion2} as one cluster, so motion3 could not be in a {motion1, motion2} cluster. Neither could motion4. Therefore, motion3 and motion4 are treated as don’t care terms. It means motion3 and motion4 are indiscernible terms in the 3 group clustering result of FR in channel 1. This process is repeated for four features from each channel. It is not possible to classify eight motions with one common cluster set. Thus, it is necessary to find a combination of common cluster sets to discern eight motions. Step #6. Minimal feature set selection: This step makes it possible to select the minimal combination based on the rough set theory [20]. The rough set theory gives us a systematic method of sufficient and necessary sets of inputs to classify outputs. By applying the rough set theory to Table 2, the reduct is found which classified four classes such as {motion1, motion2}, {motion3, motion4}, {motion5, motion6}, and {motion7, motion8}. The reduct is the minimal combination of common cluster sets and terminology in the rough set theory, makes decision attributes discernible through condition attributes. It may exist more than one reduct for a given decision table. To select the optimal reduct, the standard deviations of each feature value in reduct are used. In the view of separability, classification problem is easily solved when
the distance of between-class is big and the distance of within-class is small. The Bhattacharyya distance [15] µ (1 / 2) (Eq. (4)) is used as an important measure of the separability between distributions −1
1 Σ + Σ2 µ (1 / 2) = ( M 2 − M 1 ) T 1 (M 2 − M 1 ) 8 2 +
1 ln 2
Σ1 + Σ 2 2
(4)
Σ1 Σ 2
where M 1 and M 2 are the mean matrices of class 1 and 2, respectively, and Σ1 and Σ 2 are the covariance matrices of class 1 and 2, respectively. ⋅ means the determinant. Table 3 shows the two reducts with the highest separability in order.
Fig. 5: Flowchart for classifying the primitive motions.
4. Classification of Eight Primitive Motions Measured EMG signals are classified into eight primitive motions by minimal feature set and ASIAV. Minimal feature set is used as input-output pairs for learning in FMMNN that is suitable for classifying EMG signals, because it has an on-line adaptation function. Therefore it copes with the variation of EMG signals in
time caused by a subject’s sweat, fatigue of muscles, and so on. Learned FMMNN divide users’ motions into four classes using minimal feature set that is extracted from measured EMG signals, and then eight motions are classified according to ASIAV. Fig. 5 shows flowchart for classifying the primitive motions.
5. Safety and Classification Results One of the most essential factors is to enhance reliability of techniques in rehabilitation systems. This reliability is strongly related that the disabled and the elderly can use rehabilitation systems in safe and confidential ways. We must avoid actuating a robotic arm with high uncertainty of EMG recognition results. Therefore, the certainty of pattern classification results should be calculated. FMMNN gives a soft decision that provides a value [0, 1] that describes the degree to which a pattern fits within a class systematically. The certainty is calculated in form of Eq. (5) when the output of classification is k class. Pk =
m (c k ) p
∑ m(c )
× 100
specific constraint in experimental environment, and subjects are normal and untrained. To measure EMG signals, Ag/AgCl electrode and shielded cable are used to reduce the effect of the noise in environment. 10Hz high-pass filter and 1kHz low-pass filter are used to measure EMG signals, and 60Hz notch filter is adopted to eliminate hum noise, and measured EMG signals are amplified by 5000 times using Grass Instruments Model 15 amplifier. Filtered EMG signals are converted by A/D converter, AD976 of Analog Device, with 2kHz sampling frequency and 16 bit resolutions. Fig. 6 shows overall system configuration.
2. Data Acquisition Subjects split into two groups: group A and group B. Twenty trials for each of eight motions in group A (four subjects) are collected for extracting a minimal feature set. Each ten trials were repeated in two-day period. In group B (another four subjects), each ten trials used as test data sets to examine the user independency of the proposed method.
3. Pattern Classification Experiment
(5)
j
j =1
Where, m(c j ) and m(ck ) are an output membership value of class
(1) Experiments for Classifying Each Motion in Each Class using ASIAV Once the class of the motion to be classified is known, it can be easily classified according to ASIAV. For example, if a motion is known to be in class 1, it will be easily classified into motion 1 or 2 by ASIAV (Fig. 7 and Fig. 8).
j and class k in FMMNN, respectively.
(2) Experiments for Classifying Primitive Motions of New Subject Group To show the user independence of proposed method, fuzzy pattern classifier [21] and FMMNN pattern classifier composed of the minimal feature set in group A is used to classify the primitive motions in group B.
Fig. 6: Overall system configuration to measure EMG and to control a rehabilitation robotic arm.
Table 4. Comparison of EMG recognition rates in group B by without learning (fuzzy classifier) and with learning (FMMNN) composed of the feature sets of group A Fuzzy pattern Subject FMMNN classifier New subject 1 83 % 90 % New subject 2 83 % 89 % New subject 3 80 % 88 % New subject 4 85 % 91 % Average success rate 83 % 90 %
IV. Experimental Results 1. Experimental Environment In this paper, EMG signals are obtained in not a chamber but the normal environment, because the disabled and the elderly do not live in a chamber. There is no
Table 4 shows that the motions are recognized with success rates of 83 % and 90 % using fuzzy pattern classification and FMMNN, respectively. It means that the minimal feature set selected by the proposed method is user independent one.
(3) Experiments for Classifying Primitive Motions through Feature Sets There is not a general method to choose the feature sets to minimize the effects of other users as learning pattern in previous results. Because EMG signals are complex and has different waveforms from each person, it is difficult to confirm that specific feature sets will work well to other users.
Table 5. Comparison of EMG recognition rates according to feature sets All Heuristically Features via Subject features selected proposed (A) features (B) method (C) Subject 1 90 % 96 % 94 % Subject 2 91 % 83 % 93 % Subject 3 93 % 96 % 98 % Subject 4 85 % 89 % 94 % Average success rate 90 % 91 % 95 %
(4) Actuating Robotic Arm Experiments The EMG recognition results are used to control a wheelchair-based robotic arm. The primitive motions of a user correspond to the motions of 4 DOF robot arm via one-to-one mapping, therefore the results of classification lead to the motions of a robotic arm.
V.
Fig. 7: ASIAV of motion 1. X-axis and Y-axis represent experiment trial number and ASIAV, respectively.
Concluding Remarks
The new EMG pattern recognition procedure based on soft computing techniques is proposed to select the minimal feature set of EMG signals. A fuzzy pattern classification and FMMNN has been designed to classify EMG signals of the primitive motions of an arm. The feature sets selected by the proposed method are independent of a user. As results, the motions are recognized with success rates of 83 % and 90 % using fuzzy pattern classification and FMMNN, respectively. As a further work, we are studying segmentation of EMG signals.
Acknowledgments This work is supported by the Ministry of Science and Technology of Korea as a part of Critical Technology 21 Program.
References
Fig. 8: ASIAV of motion 2. X-axis and Y-axis represent experiment trial number and ASIAV, respectively.
Table 5 shows the success rates according to different feature sets. All features (A), feature sets selected heuristically that are not changed to each user (B), and minimal feature set selected by the proposed method (C) are used to classify the primitive motions. It is revealed that method (B) depends on subjects while method (C) gives even success rates for different subjects.
[1] Korea Institute for Health and Social Affairs, “Analysis and Polices for the Life Style of the Elderly”, Korea Institute for Health and Social Affairs, 1995. [2] Kazuhiko Kawamura, Sugato Bagchi, Moenes Iskarous, and Magued Bishay, “Intelligent Robotic Systems in Service of the Disabled”, IEEE Tr. on Rehab. Eng., vol. 3, no. 1, pp. 14- 21, March, 1995. [3] Z. Kazi, M. Salganicoff, Beitler, S. Chen, D. Chester and R. Foulds, “Direct Manipulation of 3-D Objects through Multimodal Control: Towards a Robotic Assistant for People with Physical Disabilities”, ASEL technical Report #ROB9509, AI duPont Institute, Univ. of Delaware. [4] Carol A. Stanger, Carolyn Anglin, Wiliam S. Harwin, and Douglas P. Romilly, “Devices for Assisting Manipulation: A Summary of User Task Priorities”, IEEE Tr. on Rehab. Eng., vol. 2, no. 4, pp. 256-265, Dec., 1994.
[5] Jin-Woo Jung, Won-Kyung Song, Heyoung Lee, Jong-Sung Kim, and Zeungnam Bien, “A Study on the Enhancement of Manipulation Performance of Wheelchair-mounted Rehabilitation Service Robot”, Proc. of International Conference on Rehabilitation Robotics (ICORR), pp. 4249, Stanford, California, U.S.A., July 1-2, 1999. [6] Brenda Bigland and O. C. J. Lippold, “The Relation between Force, Velocity and Integrated Electrical Activity in Human Muscles”, J. Physiology. vol. 123, pp. 214-224, 1954. [7] S. H. Kim et al, “A Digital Signal Processing System for Analysis of Skeletal Muscle EMG Signal”, Journal of KOSOMBE, vol. 17, no. 2, pp. 155-164, 1996. [8] G. N. Saridis and T. Gootee, “EMG Pattern Analysis and Classification for a Prosthetic Arm”, IEEE Tr. Biomed. Eng., vol. BME-29, pp. 403-409, June 1982. [9] S.-H. Hong et al, “A Study on the Pattern Classificatin of the EMG Signals Using Neural Network and Probabilistic Model”, Journal of IEEK, vol. 28, no. 10, pp. 85-95, 1991. [10] Daisuke Nishikawa, Wenwei Yu, Hiroshi Yokoi and Yukinori Kakazu, “EMG Prosthetic Hand Controller Discriminating Ten Motions using Real-Time Learning Method”, IROS '99, vol. 3, pp. 1592-1597, Kyungju, Korea, 1999. [11] D. S. Dorcas, V. A. Dunfield and R. M. Scott, “Improved Myoelectric Control System”, Med. Biol. Eng., vol. 8, pp. 333-341, 1970. [12] Daniel Graupe and William K. Cline, “Functional Separation of EMG Signals via ARMA Identification Methods for Prosthesis Control Purposes”, IEEE Tr. on Systems, Man, and Cybernetics, vol. SMC-5, no. 2, pp. 252-259, March 1975. Frequency Ratio Subject No. of Cluster 2
[13] Osamu Fukuda, Toshio Tsuji, Akira Otsuka and Makoto Kaneko, “An EMG-based Rehabilitation Aid for Prosthetic Control”, IEEE Int'l Workshop on Robot and Human Communication, vol. 1, pp. 214-219, Takamutsu, Kagawa, Japan, Sep.30- Oct.2, 1998. [14] Yong-Sheng Yang, F. K. Lam, Francis H. Y. Chan, Yuan-Ting Zhang, Philip. A. Parker, “A New Fuzzy Approach for Pattern Recognition with Application to EMG Classification”, The 1996 IEEE Int'l Conf. on Neural Networks, IEEE Part vol. 2, pp. 1107- 1114, New York, USA, 1996. [15] Keinosuke Fukunaga, “Introduction to Statistical Pattern Recognition”, Academic Press, first edition, 1972. [16] H. Eftring and K. Boschian, “Technical results from MANUS user trials”, In Proc. Sixth Int. Conf. on Rehabilitation Robotics, pp. 136-141, 1999. [17] John V. Basmajian and Carlo J. De Luca, “Muscle Alive”, Williams & Wilkins, fifth edition, 1985. [18] P. Simpson, “Fuzzy Min-Max Neural Networks- Part 1: Classification”, IEEE Tr. on Neural Networks, vol. 3, pp. 776- 786, Sep., 1992.` [19] Won-Kyung Song, Heyoung Lee, and Zeungnam Bien, “KARES: Intelligent wheelchair-mounted Robotic Arm System using Vision and Force Sensor”, Robotics and Autonomous Systems, vol. 28, no. 1, pp. 83-94, 1999. [20] Z. Pawlak, “Rough Sets: Theoretical Aspects and Reasoning about Data”, Kluwer Academic Publishers, 1991.
Table 1. Clustering results of frequency ratio for each subject in channel 1 Channel 1 Subject 1
Subject 2
Subject 3
Subject 4
{m1,m2,m3,m4} {m5,m6,m7,m8}
{m1,m2,m3,m4} {m5,m6,m7,m8}
{m1,m2,m3,m4} {m5,m6,m7,m8}
{m1,m2,m3,m4} {m5,m6,m7,m8}
{m1,m2}{m3,m4} {m1,m2}{m3,m4} {m1,m2,m3}{m4} {m1,m2,m3,m4} {m5,m6,m7,m8} {m5,m6,m7,m8} {m5,m6,m7,m8} {m5,m6}{m7,m8} {m1,m2}{m3,m4} {m1,m2}{m3,m4} {m1,m2}{m3}{m4} {m1,m2,m3,m4} 4 {m5,m6,m7}{m8} {m5,m6,m8}{m7} {m5,m6,m7,m8} {m5}{m6,m7}{m8} (note) ‘m1’ means wrist pronation (motion1), ‘m2’ means wrist supination (motion2), ‘m3’ means elbow flexion (motion3), ‘m4’ means elbow extension (motion4), ‘m5’ means humeral moving up (motion5), ‘m6’ means humeral moving down (motion6), ‘m7’ means humeral moving in (motion7), and ‘m8’ means humeral moving out (motion8). 3
Table 2. Decision table Conditions Frequency Ratio (channel 1)
Feature Attribute num.
a0
a1
a2
a3
Decision IAV (channel 4) a80
a81
Motion Cluster 2 Cluster 3 Cluster 4 Cluster 4 motion1 Big Big Big Big Small class1 motion2 Big Big Big Big Small class1 motion3 Big class2 × × × × motion4 Big class2 × × × × motion5 Small Small Big class3 × × motion6 Small Small class3 × × × motion7 Small Medium class4 Small × × motion8 Small Small Small class4 × × (note) Each attribute is defined on the basis of clustering results for each channel and each feature. Attributes consist of ‘Very big’, ‘Big’, ‘Medium’, ‘Small’, and ‘ × ’, and the attributes are gaussian membership functions in the fuzzy logic. The ‘ × ’ means don’t care condition.
Table 3. The two reducts which has the highest separability Class1 Class2 Attribute Number
Attribute
m1
M2
m3
m4
m5
Class3
Class4 m6
m7
m8
Frequency Ratio, Channel3, a9 Small Small Medium Big Big × × × Cluster 3 a65 IAV, Channel 2, Cluster 3 Medium Medium Big Big Small Small Small Small a73 IAV, channel 3, Cluster 4 Small Small Medium Big Big × × × Frequency Ratio, Channel3, Small Small Medium Big Big a9 × × × Cluster 3 Variance, Channel4, Small Small Big Medium Medium a38 × × × Cluster 3 a65 IAV, Channel2, Cluster 3 Medium Medium Big Big Small Small Small Small (note) Overall number of the extracted reducts through the proposed method is 21. One reduct is the combination of ‘a9’, ‘a65’, and ‘a73’. The other reduct is the combination of ‘a9’, ‘a38’, and ‘a65’. These two reducts have the highest separability among 21 reducts.