Electromyograph Classification of Six Hand-Actions for Twisting Manipulation - IEEE Int. Conf. on Advanced Robotics and MEchatronics

Editor's Notes

• #2: Thank you for introduction. I am from Japan. Today I will talk about electromyograph classification, which is used for prosthetic hands.
• #3: Our motivation of the research is to classify such combined motion of hands. Many of previous researches on electromyograph classification have focus on single motion such as open, close or twisting the wrist. On the other hand, twisting manipulation is necessary in daily activities such as inserting a key and unlocking, and rotating a door-knob to open the door. In this task, we require the combined action that is composed of opening or closing the fingers and twisting the wrist.
• #4: Our contribution of the study is to classify combined actions with a simple framework and devices. We use a commercial band-type sensors array, which is easy for non-professional users to attach it to non-precise position. And conventional signal processing composed of rectification and integrated EMG is adopted. In addition, Support Vector Machine, which is one of conventional classification method is also used. That is, our proposed method is not novel from the viewpoint of signal processing and classification of EMG, but focusing on classification the combined actions with simple measurement system for ease of implement.
• #5: Many of commercial EMG prosthetic hand often have 1 DOF actuator for grabbing something. It is because such simple mechanism is highly robust and makes its control system simple and robust. As for these hands, twisting motion of the wrist is often manually executed by other triggers. On the other hand, high DOF hands with a lot of actuators have been developed mainly in the laboratories. They usually require a lot of sensing data of EMG from each muscle, that is, we need to attach many electrodes on the arm.
• #6: Consequently, we aim to classify the combined actions of the hand by using simple measurement system. we use a commercial EMG sensor and attach it to the arm without precise positioning. In addition, we adopt conventional signal processing and classifier so that they can be easily implemented.
• #7: Our signal processing is not novel but conventionally used for EMG analyses. Rectification contributes to reduce noises and spikes of retrieved raw signal of electromyograph. And we calculate integrated EMG, called as IEMG to recognize character of rectified signal. As for classification, Support Vector Machine and neural network are implemented in our system. The neural network is used to correct EMG affected by upper limb motion. I explain about it later.
• #8: Training of SVM always requires a lot of dataset. For our classifier, the dataset is composed of measured IMEG and motion label of performed action. The former, IEMG is easily calculated with EMG signal retrieved easily by using this band-type sensors array. On the other hand, it is a too heavy task for us to correspond exact motion labels to the IEMG manually. To reduce the hard work for users, we use a hand tracking device, Leap Motion, to automatically recognize posture of hand for training dataset.
• #9: Leap Motion can recognize shape of hand with tracking position of joints and palm of the hand with high accuracy. Under our experimental environment, it can accomplish 100 percent accuracy of recognition for the target six postures. With the result of pre experiment, we regard the recognition by Leap motion as exact motion label.
• #10: This is the overview of our proposed system for classification of EMG signal of six hand actions. The electrodes retrieve EMG signal, and the signal is rectified and smoothed. And then IEMG is calculated. At the same time, the hand tracker, Leap Motion recognize the hand posture as corresponding hand action, and give us a motion label as a part of dataset. The dataset composed of motion label and the corresponding IEMG is used for training of SVM classifier. Finally, SVM classifies test dataset of EMG signal.
• #11: This is a procedure to collect dataset for SVM training. A subject lays down his forearm on the small box and relax. Next, he makes either form of six hand-actions, which is composed of opening or closing, and either of pronation, supination and neutral posture. And then he keeps the posture in 10 seconds for measurement to collect enough number of pairs of data.
• #12: We report our experimental results. The first test is to evaluate accuracy of classification of EMG signal for six hand-actions. The table of the results shows more than 90 percent of accuracy is accomplished by our SVM classifier. Even though the training dataset from three subjects is merged, the classifier is generalized for different individuals. It means some differences between individuals are successfully attenuated.
• #13: These two plots represent the results of real time classification. As for the experienced user, the classification is successfully accomplished with high accuracy and quickness. On the other hand, recognition of EMG signal of non-experienced user sometimes fails. It is because that the non-experienced user was not good at measuring EMG signals during the actions, and then the retrieved signal had large noises that influence classification.
• #14: We investigated interference of some factors. The first is about muscular fatigue. We continued the previous real time classification in 30 minutes, and analyzed its accuracy for each time period of 2 minutes. As a result, the classification accuracy is always higher than 90 percent. If this outlier is neglected, the score is always higher than 93 percent. Consequently, muscular fatigue seldom interferes in our system of EMG classification. By the way, the above results are with the condition that the forearm is laid down on the table. It means that these training and test data may be changed by other posture of forearm. In the daily life, forearm posture is usually changed to adjust it to each task.
• #15: As a result, changing upper limb motion during measurement of EMG interferes in collecting data of EMG. It is natural because some muscles are commonly used to drive a hand and also a forearm. In order to overcome the interference, we modify our classification framework as that in the next slide.
• #16: The band-type sensors array has an IMU to estimate posture of forearm. The data of posture are expressed by quaternion. We introduce a neural network to obtain components of EMG signal for upper limb motion so that we can remove it as noise signal. The input of our neural network is retrieved quaternion, and output is IEMG, which corresponds to upper limb motion. The trained neural network can let us know difference of EMG signal between only hand motion and whole arm motion.
• #17: These are results of classification of EMG with using the neural network for correction. According to the results, our introduced EMG correction using quaternion representing upper limb posture can work successfully and almost higher accuracy can be accomplished. On the other hand, more training data contribute higher accuracy of classification, and effect of the introduced correction is relatively decreased. As the result, enough training data improves the performance of classification even though we do not introduce the process of EMG correction.
• #18: I summarize my talk. We aim to classify six combined actions of hand with using a commercial band-type sensors array. In addition, we investigated interferences of other factors in EMG classification, especially muscular fatigue and upper limb motion. In future works, EMG prosthetic hands with our proposed system will be built. Moreover we’d like to classify other actions in the daily activities with using a simple measurement system.