Publication: Mayoelektrik Protez Elin Yapay Zeka Metotları Kullanılarak Gerçek Zamanlı Olarak Denetlenmesi
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Abstract
Bir kasın kasılması sonucu oluşan mayoelektrik işaretler (EMG), kasılmanın nedeni olan hareketi niteleyen ayırt edici özellikler taşır. Sağlam kas gruplarından alınan EMG işaretleri uygun yöntemlerle işlenerek sınıflandırılabilir, bu kas gruplarına ait hareketler gerçek veya sanal bir yapay eli denetleyecek bir denetim işareti üretilebilir. Literatürdeki çalışmalarda bir elektromekanik eli denetleyecek beş aşamalı bir yapı kullanılmaktadır. Bu aşamalar sırasıyla; işaretin alınması , özellik vektörlerinin çıkarılması, boyut azaltılması, işaretin sınıflandırılması, denetim işaretinin üretilmesi ve elin sürülmesi şeklindedir.Bu tezin amacı, bir elektromekanik elin gerçek zamanlı olarak yapay zeka metotları kullanılarak denetlenmesidir. Bunun için Doğrusal Ağırlıklandırılmış Normalize Radyal Taban İşlevi (DANRTİ) ağları, Çok Katmanlı Yapay Sinir Ağları (ÇKYSA) ve Destek Vektör Makinaları (DVM), ile sınıflandırma işlemi yapılmış ve üretilen denetleme işaretleri ile elektromekanik el sürülmüştür. Çalışmanın ilk aşamasında, Karşı Yayılım Ağı (KYA) , Bulanık C Ortalamaları (BCO), Gustafson-Kessel (GK), kümeleme, Gradyan Azalması (GA), GA-Levenberg-Marquardt (LM) öğrenme algoritmalarından oluşan melez öğrenme yapıları tasarlanmıştır. İkinci aşamasında ise, iki ve dört kanallı EMG işaretleri ile deneysel bir elektromekanik elin gerçek zamanlı olarak denetimi gerçekleştirilmiştir.Melez öğrenme yapılarından KYA, BCO ve GK kümeleme algoritmalarının başarımları GA algoritması ile sınanmıştır. GA algoritmasının başarımının düşük olması nedeniyle GA-LM ` ye geçilmiştir. Deneysel bir elektromekanik elin gerçek zamanlı olarak iki kanallı olarak denetimi için ilk olarak kişinin sağlam kaslarından alınan EMG işaretinin zarfı çıkarılmış ve eşik değerler ile karşılaştırılarak dört farklı el hareketi için PIC devresi ile üretilen Darbe Genişlik Bindirimi (DGB) işaretleri ile denetim gerçekleştirilmiştir. Daha sonra iki kanallı EMG verilerine , Dalgacık Dönüşümü uygulanarak özellik vektörleri çıkarılmış, Doğrusal Diskriminant Analiz (DDA) ile boyut azaltılarak tasarlanan melez ağ yapısı üzerindeki başarımı test edilmiştir. İki kanallı EMG verileri elin dört farklı hareketini belirlemektedir. Yedi farklı hareketi yapabilmek için dört kanallı EMG işaretleri kullanılmıştır. Alınan EMG işaretlerinin dalgacık dönüşümü ile özellikleri vektörleri bulunmuş ve TBA ile boyut azaltılarak DVM , ÇKYSA ve DANRTİ sınıflandırıcı ile sınanmıştır.Eğitim sonucunda üretilen denetleme işareti paralel port aracılığı ile bir PIC devresine verilmiştir. PIC devresi tarafından üretilen DGB işaretleri ile deneysel bir elektro-mekanik elin gerçek zamanlı olarak denetlenmesi sağlanmıştır. Sonuçlardan DVM `nin %94.3, ÇKYSA' nın % 92.9 ve DANRTİ `nin ise %90 başarım sağladığı görülmüştür.
Myoelectric signals (EMG) which occur as a result of muscular contraction have differential features characterizing the movement that is the cause of the contraction. EMG signals acquired from the healthy muscle groups can be processed and classified through appropriate methods. A control signal monitoring the movements belonging to these muscle groups; or monitoring a real or virtual artificial hand can be produced. Studies in the literature use a five-stage structure to control an electromechanical hand. Respectively these stages are as follows (i) acquiring the signals, (ii) identifying feature vectors, (iii) reducing the dimension (iv) classification of the signal, (v) producing control signals and driving artificial hand.This dissertation aims to control an electromechanical hand real-timely using artificial intelligence methods. For the purpose of the study, classification process has been completed using Support Vector Machines (SVM), Multilayer Perceptron (MLP) Networks and Linear Weighted Normalized Radial Basis Function (LWNRBF) Networks, and with the produced control signals, electromechanical hand has been controlled. In the first stage of the study, hybrid learning structures consisting of Counter Propagation Network (CPN), Fuzzy C Means (FCM) and Gustafson-Kessel (GK), grouping Gradient Descent (GD), GD-LM learning algorithms have been designed. In the second stage, using two or four channeled EMG signals, real-time control of an experimental electromechanical hand has been conducted.The success of the hybrid learning structures, which are CPN, FCM and GK grouping algorithms, has been tested by GD algorithm. Due to the low success of GD algorithm, GD-LM has been used. For the real-time two-channeled control of an experimental electromechanical hand, initially EMG signals of a person?s healthy muscles have been identified, threshold values have been compared, and for four different hand movement, PMW signals produced with PIC circuit have been used to perform the control process. Next, to the two-channeled EMG values, wavelet transform has been applied and feature vectors have been determined, size has been reduced via Linear Discriminant Analysis ( LDA) and its success on hybrid network structure has been tested. Two-channeled EMG values determine the four different movements of the hand. In order to be able to perform seven different movements four channeled EMG signals have been used. Wavelet transform of the acquired EMG signals and feature vectors have been identified, and size has been reduced with PCA; then tested with SVM, MLP and LWNRBF classifiers.The control signal produced as the result of the education was transferred to a PIC circuit via a paralel port. An experimental electromechanical hand?s real-time control has been enabled by PWM signals produced by PIC circuit. The results have shown that the success rate of SVM was 94.3 %, that of MLP was 92.9 % and LWNRBF?s was 90 %.
Myoelectric signals (EMG) which occur as a result of muscular contraction have differential features characterizing the movement that is the cause of the contraction. EMG signals acquired from the healthy muscle groups can be processed and classified through appropriate methods. A control signal monitoring the movements belonging to these muscle groups; or monitoring a real or virtual artificial hand can be produced. Studies in the literature use a five-stage structure to control an electromechanical hand. Respectively these stages are as follows (i) acquiring the signals, (ii) identifying feature vectors, (iii) reducing the dimension (iv) classification of the signal, (v) producing control signals and driving artificial hand.This dissertation aims to control an electromechanical hand real-timely using artificial intelligence methods. For the purpose of the study, classification process has been completed using Support Vector Machines (SVM), Multilayer Perceptron (MLP) Networks and Linear Weighted Normalized Radial Basis Function (LWNRBF) Networks, and with the produced control signals, electromechanical hand has been controlled. In the first stage of the study, hybrid learning structures consisting of Counter Propagation Network (CPN), Fuzzy C Means (FCM) and Gustafson-Kessel (GK), grouping Gradient Descent (GD), GD-LM learning algorithms have been designed. In the second stage, using two or four channeled EMG signals, real-time control of an experimental electromechanical hand has been conducted.The success of the hybrid learning structures, which are CPN, FCM and GK grouping algorithms, has been tested by GD algorithm. Due to the low success of GD algorithm, GD-LM has been used. For the real-time two-channeled control of an experimental electromechanical hand, initially EMG signals of a person?s healthy muscles have been identified, threshold values have been compared, and for four different hand movement, PMW signals produced with PIC circuit have been used to perform the control process. Next, to the two-channeled EMG values, wavelet transform has been applied and feature vectors have been determined, size has been reduced via Linear Discriminant Analysis ( LDA) and its success on hybrid network structure has been tested. Two-channeled EMG values determine the four different movements of the hand. In order to be able to perform seven different movements four channeled EMG signals have been used. Wavelet transform of the acquired EMG signals and feature vectors have been identified, and size has been reduced with PCA; then tested with SVM, MLP and LWNRBF classifiers.The control signal produced as the result of the education was transferred to a PIC circuit via a paralel port. An experimental electromechanical hand?s real-time control has been enabled by PWM signals produced by PIC circuit. The results have shown that the success rate of SVM was 94.3 %, that of MLP was 92.9 % and LWNRBF?s was 90 %.
Description
Tez (doktora) -- Ondokuz Mayıs Üniversitesi, 2011
Libra Kayıt No: 74555
Libra Kayıt No: 74555
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
193
