Initial Point Detection

Because the movements involved in the proposed system are not complicated, an initial point detection method is proposed to deduce the motion intention from the EMG signal. The

reason for the naming is because this method determines the onset of the upper limb motion via detecting the instant when the magnitude of the extracted EMG feature reaches the critical values. Due to its simplicity, real-time motion governing can be achieved. In the proposed classifier design, we start with the single critical value detection, in which the state of the muscle MS is determined by checking if the initial value for the feature exceeds a predefined critical value:



 , 0

, MS 1 if

otherwise CV F_k 

(2-5)

where F_k stands for the th feature and CV the critical value. An active MS corresponds to an

``ON’’ robot command and an inactive one for that of ``OFF’’, as illustrated in Figure 2-4(a).

Figure 2-5 shows an example, in which MAV is used to evaluate the EMG signal of Biceps Brachii. In Figure 2-5, section A indicates the muscle state during relaxation, and section B that during flexion, both of which exhibit some fluctuations. We thus propose a concept of double critical value detection, as illustrated in Figure 2-4(b). In Figure 2-4(b), the state of the muscle MS is determined to be active when the initial value for the feature Fk is larger than the upper critical value CVu, and MS inactive when Fk is smaller than the lower critical value CVl:

k

(2-6)

Figure 2-4 Conceptual diagram for single and double critical value detection: (a) single critical value and (b) double critical value.

The selection of CVu and CVl depends on the following observations. A large CVu implies that the user has to generate a large force to move the robot arm. It may lead to muscle fatigue, in addition to the increase of the crosstalk between muscles. Contrarily, a small CVu results in low tolerance against the noise. Meanwhile, a large CVl may make the robot arm stop its movement earlier than that of the user, while a small CVl leads to the opposite. The approaches of the trial-and-error method and fuzzy system can be utilized for determining CVu and CVl.

Voltage (V)

A B

Time (100 ms)

Figure 2-5 An example of EMG signal evaluation of Biceps Brachi using MAV.

To demonstrate that CVs, CVu and CVl, can be set to be fixed under certain condition in a period of time, we chose a set of fixed CVs empirically and performed the motion of elbow up and down for fifty times continuously. The experimental results in Figure 2-6 show consistent classification, except the twenty-fifth trial (marked in red). The entire process lasted for about 10 minutes. After that, the fatigue of the muscle led to inconsistent classification. It indicates that fixed CVs are appropriate for the proposed system to govern robot motion for a certain period of time, but should not be used when the subject felt fatigued.

Several factors influence the realization of the classifier, including feature selection, number of samples for feature extraction, and choice of the CVs. It is suggested to use

features with smooth waveforms. A larger number of samples may be helpful for feature extraction, at the expense of efficiency and delay.

Classification Output

Time (second)

Figure 2-6 Feasibility evaluation by performing the motion of elbow up and down for fifty times continuously (numbers 0-3 correspond to STOP, UP, DOWN, and ERROR, respectively).

Chapter 3

Experimental Design

The proposed EMG-based upper-limb robot control system is using four sets of electrodes placed on Biceps Brachii (BB), Triceps Brachii (TB), Pectoralis Major (PM), and Teres Minor (TM), as shown in Figure 3-1, to control the robot arm movement of either one DOF or multi-DOFs. For the proposed detection method, the classifier is designed to let the feature extracted from the BB correspond to upper limb flexion, that from the TB for extension, that from PM for internal rotation, that from TM for external rotation, that from the synthesis of BB and PM, for flexion-internal rotation, and that from the synthesis of TB and TM for extension-external rotation. Their muscle states will determine whether it is an up, down, turn-left, turn-right, up-left and down-right movement. Due to some muscle crosstalk or imprecise feature identification, sometimes it may lead to conflict movement decision between the two muscles. Under such circumstances, the classifier will output an error signal.

Therefore, there are eight outputs for the classifier: STOP, UP, DOWN, LEFT, RIGHT, UP-LEFT, DOWN-RIGHT and ERROR.

Table 3-1 summarizes the mapping from EMG to robot movement. When EMG signals from all channels (CH1~4) are determined to be OFF, the classifier outputs 0 as relaxation;

ON for CH1 and OFF for the others, outputs 1 as flexion; ON for CH2 and OFF for the others, outputs 2 as extension; ON for CH3 and OFF for the others, outputs 3 as internal rotation; ON

for CH4 and OFF for the others, outputs 4 as external rotation; simultaneously ON for CH1 &

CH3 and OFF for the others, outputs 5 as flexion plus internal rotation; simultaneously ON for CH2 & CH4 and OFF for the others, outputs 6 as extension plus external rotation; and simultaneously ON for undefined channels, outputs 7 as error detection. Figure 3-2 illustrates the classification outputs corresponding to the robot arm movements.

Biceps Brachii (CH1)

Teres Minor (CH4)

Triceps Brachii (CH2) Pectoralis Major

(CH3)

(a) (b)

Figure 3-1 Electrode locations: (a) Biceps Brachii and Pectoralis Major, and (b) Triceps Brachii and Teres Minor.

Table 3-1 Mapping from EMG to robot movement

Upper Limb Status Classifier

Output Robot Arm

Figure 3-2 Illustrations of classification outputs corresponding to the robot arm movements