Two sets of electrodes placed on the Biceps Brachii, marked as CH1, and the Triceps Brachii, marked CH2, respectively, shown in Figure 3-1, are used to govern the motion of RV-2A.
First things first, the setting of CVu and CVl has to be determined for the invited subjects. In the first set of experiments, we used the empirical method to determine CVu and CVl for the three invited male subjects, with their physical data listed in Table 3-5 and derived CVu and CVl in Table 3-6. Their determination is customized for each individual subject through an extensive trial-and-error procedure according to the effectiveness on classification. The subject was asked to contract/extend his upper limb to move robot arm, J2 axis, from 0 to 90 degrees, 90 to 0 degrees, 0 to 45 degrees, 45 to 90 degrees, 90 to 45 degrees, and 45 to 0 degrees. The moving speed of the robot arm is 6 deg./sec. We define the successful discrimination rate (SDR) as the times that the robot arm successfully follows the motion of the subject out of the total number of classification:
% tion 100
classifica of
number Total
following motion
successful of
Number
SDR (3-1)
Table 3-5 Physical data of the three male subjects
Subject Height (cm) Weight (kg)
Muscle for electrode
A 174 70 Ordinary
B 166 60 Slender C 164 82 Fat
Table 3-6 Critical values via the empirical method
Biceps Brachi Triceps Brachi
Subject
CVu CVl CVu CVl
A 4.5 2.2 4 2.5 B 3.2 2.2 4.5 3 C 5 2.8 5 3.3
The experimental results are shown in Figures 3-9 to 3-11. Figure 3-9 shows the EMG signals after band-pass filtering for subjects A-C, in which larger amplitudes indicate larger forces during the movements of the flexor/extensor. Figure 3-10 shows the variations of MAV features corresponding to the filtered EMG signals in Figure 3-9. Based on these features, the classifier determines the corresponding upper limb movements, while more evident feature
variations lead to better discrimination. Figure 3-11 shows the outputs from the classifier,
where numbers 0-3 in the vertical coordinate denote STOP, UP, DOWN, and ERROR respectively, and I-VI the stages of 0-90, 90-0, 0-45, 45-90, 90-45, and 45-0.
Subject C reported that he felt a little bit fatigued. It might be due to higher CVu and CVl
demanded him to make more effort for movement. The SDR for the subjects is 95.5%, 97%, and 95.5%, respectively, indicating quite successful motion following.
0 200 400 600 800 1000 1200 1400 1600 1800 2000 -20
-10 0 10 20
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200 1400 1600 1800 2000 -20
-10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(a) Subject A
Figure 3-9 Filtered EMG signals for subjects A-C.
0 200 400 600 800 1000 1200
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200
-20 -10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(b) Subject B
0 200 400 600 800 1000 1200 1400 1600 1800
-20 -10 0 10 20
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200 1400 1600 1800
-20 -10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(c) Subject C
Figure 3-9 (Cont.) Filtered EMG signals for subjects A-C.
0 20 40 60 80 100 120 140 160 180 200
Figure 3-10 EMG feature variations for subjects A-C.
0 20 40 60 80 100 120 140 160 180 0
2 4 6 8
Biceps Brachii MAV
Time (S)
Voltage (V)
0 20 40 60 80 100 120 140 160 180
0 2 4 6 8
Triceps Brachii MAV
Time (S)
Voltage (V)
(c) Subject C
Figure 3-10 (Cont.) EMG feature variations for subjects A-C.
(a) Subject A
(b) Subject B
I II III IV V VI
I II III IV V VI
I II III IV V VI
(c) Subject C
Figure 3-11 Outputs from the classifier for subjects A-C.
Due to the individual fuzziness, the tuning of the system parameters for the individual user is not that straightforward. Thus the concept of the fuzzy system [35, 51, 52] is employed for CVu and CVl determination, so that the tedious process encountered in the trial-and-error method can be avoided. MAV (signal power) and BZC (zero crossing) were chose as the input variables of the fuzzifier, as each of them provides the time- and frequency-domain estimation, respectively. Figure 3-12 shows the fuzzy sets used for MAV, BZC, and CVu (CVl),
where W, M, and S stand for weak, middle, and strong, L, M, and H for low, middle, and high,
, , and () the strength of MAV, BZC, and CVu (CVl), respectively, and A, B, and C (D)
the membership function for MAV, BZC and CVu (CVl). With them, the fuzzifier transforms
the extracted features into the linguistic values. The fuzzy rules, listed in Table 3-7, are obtained from the empirical knowledge acquired via extensive experiments. The values of and are empirically set to be 1/3 of the MAV and BZC, respectively, and and are 2 and 1.4 times of . The fuzzy inference engine determines the jth firing strength j of the jth fuzzy
rule via Eq.(3-2):
And, the defuzzifier utilizes the center of gravity (COG) method to map the inferred fuzzy action into a nonfuzzy value of CVu (CVl):
where n is the number of fuzzy rule and zj the strength of CVu (CVl) at the jth fuzzy rule.
Table 3-7 Fuzzy rule base BZC
In the second set of experiments, the fuzzy system is used to determine CVs for each individual user, listed in Table 3-8, and asked the subjects to perform the same movements as
those they did in the first set of experiments. The experimental results are shown in Figures 3-13 to 3-15. Figure 3-13 shows the filtered EMG signals, Figure 3-14 variations of the MAV features, and Figure 3-15 outputs from the classifier. The SDR for the subjects is 95.5%, 97%, and 97%, respectively, also indicating quite successful motion following. From Tables 3-6 and 3-8, different set of CVs were derived by the empirical method and fuzzy system, while both of them led to successful motion governing. Meanwhile, the fuzzy system may be with better potential when dealing with more complex movement, as it possesses the ability of automatic parameter tuning.
Table 3-8 Critical values via the fuzzy system Biceps Brachi Triceps Brachi Subject
CVu CVl CVu CVl
A 4.09 2.86 4.25 2.98 B 4.99 3.49 4.34 3.04 C 4.36 1.96 4.95 2.23
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200 1400 1600 1800 2000 -20
-10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(a) Subject A
0 200 400 600 800 1000 1200 1400 1600 1800
-20 -10 0 10 20
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200 1400 1600 1800
-20 -10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(b) Subject B
Figure 3-13 Filtered EMG signals for subjects A-C.
0 200 400 600 800 1000 1200 1400 1600 1800
Filtered Biceps Brachii EMG
Sample
Voltage (V)
0 200 400 600 800 1000 1200 1400 1600 1800 -20
-10 0 10 20
Filtered Triceps Brachii EMG
Sample
Voltage (V)
(c) Subject C
Figure 3-13 (Cont.) Filtered EMG signals for subjects A-C.
0 20 40 60 80 100 120 140 160 180 200
Figure 3-14 EMG feature variations for subjects A-C.
0 20 40 60 80 100 120 140 160 180
Figure 3-14 (Cont.) EMG feature variations for subjects A-C.
(a) Subject A
(b) Subject B
I II III IV V VI
I II III IV V VI
I II III IV V VI
(c) Subject C
Figure 3-15 Outputs from the classifier for subjects A-C.
Chapter 4
Multi-DOF Robot Arm Movement Control
While the proposed system is shown to be effective for 1-DOF robot motion governing, it is not appropriate to serve as a classifier for more than 1-DOF upper limb motion as it has larger muscle mutual interference. To tackle this, the EMD method is applied in feature extraction design. To reduce the computational load in EMD, a sixth-order band-pass Butterworth filter and a window with 20 samples per second are employed. The muscle state was determined by the root mean square (RMS) of the 2nd IMF, c2(t), expressed as:
Fk = RMS(c2(t)), 0 t 20 (4-1)
Meanwhile, for multi-DOF upper limb motion, the fuzzy system adopted for 1-DOF motion is not efficient enough for the tuning of the critical valuesfor each individual user. For its excellence on adaptation, the ANFIS is employed to realize the fuzzy system.