Study On Multi-Motion EMG Control Of Anthropomorphic Prosthetic Hand

The myoelectric hand is a type of anthropomorphic robotic hand that driven by the human electromyography (EMG) signals for amputee rehabilitation. As a multi- discipline system, it integrates the subjects of mechatronics, computer and biology together, trends to develop with more degree-of-freedoms (DOFs) and sensors, higher integration and miniaturization. However, the EMG control of the multifunctional prosthetic hand has been studied deficiently comparing with the hand bodies, which performs with a few motion modes, low accuracy and bad intuitive control feelings. Focus on the multi-DOFs prosthetic hand, this paper presents a new EMG control method aiming at efficiently recognizing more hand motion modes. The contents of this paper includes: the overall structure of the pattern recognition (PR) based EMG control; the hand prehension recognition; the hand gesture and finger motion recognition; and the online learning strategy for the recognition system, et al.Overviews about newly developed multi-DOFs prosthetic hands and typical EMG control methods are given in detail. Then, the general structure of the PR-based EMG control system of the multifunctional prosthetic hand is established, which can be subdivided as the human layer, the biologic signal acquisition layer, the pattern recognition layer, the prosthetic hand layer and the feedback layer. Methods of implementing the critical techniques in each layer are presented. In detail, the placement of the myo-electrodes used for detecting each motion mode is determined by the muscles taking charge of the motion of each finger with the knowledge of the human biological anatomy. The modularized active EMG electrodes are adopted to acquire the surface myoelectric signals. The algorithm verification is performed to get the empirical highest recognition accuracy, and a virtual hand environment is established to demonstrate the EMG control effectiveness and give the amputee an online discipline on his muscle functionality.Four basic preshaping modes of the hand are determined, and the recognition of the preshaping EMG modes is adopted in the hand's grasping strategy when it attempts to hold objects. Although the grasping motion is not actually performed by the human hand, the proper strategy about the prosthetic hand how to grasp the object has been defined. The amplitude change of the EMG signal is reinforced by the Teager-Kaiser Energy (TKE) operator when the muscle contracts, then the accurate preshaping EMG data can be obtained through statistically comparing the signal's statistical values and adding two post-processing approaches. Multifarious methods of the data segmentation, feature extraction and classification are tested and verified on their influence to the recognizability of the PR system. Ultimately, the wavelength of the signal and the support vector machine (SVM) are adopted as the empirical optimum EMG feature and classifier, respectively, to realize the online recognition of the preshaping modes. The experimental results show that the recognition algorithm can reach on both a high accuracy and a low time delay, and the prosthetic hand can grasp objects complying with the predefined prehension modes correctly and swiftly.The recognition of the hand gestures (or finger motions) is with the same importance to the preshaping mode in the EMG control of the prosthetic hand. Based on the reconfiguration of the hand gesture modes, multiple gesture modes are recognized reliably by decoding six channels of EMG signals. The proposed double-decision strategy, which combined the threshold decision with the SVM decisions, can increase the number of the transient samples of the EMG signals, thus to improve the recognition accuracy at the time of the finger motion taking place. A new sample collecting method named"vibrating training"(collect the EMG samples while the muscle is contracting at a fast rhythm) is suggested to improve the completeness of the training samples in each EMG mode. The experiments of a PC-based online multi-mode EMG control of a virtual hand, a DSP-embedded real-time control of a five-fingered prosthetic hand, and an implementation of the EMG control method to an amputee are performed. The results show that the classification accuracy of the EMG modes is relatively high, the recognition speed is fast, and the intuitive feeling of the EMG control is intensive.The stochastic character of the EMG will lead to a accuracy decline of the recognition system. A new multi-class data description method, which trains the positive and negative samples at the same time, is proposed. An new sample updating method named"husking"algorithm is proposed, based on the multi-class hypersphere SVM through weighting the penalty factor by a suggested forgotten factor. The"action sequence"is defined and then adopted for the on-line continual learning of the EMG modes. Experiments about the online learning of the preshaping and finger motion EMG modes show that, the persisting learning of the hand preshaping mode can be regarded as a completive process of the mode description function. For the learning process of the hand motion modes, it cannot achieve a higher recognition accuracy even using more groups of training samples. The"husking"algorithm can efficiently control the increasing number of the support vectors and make a good intragroup discriminating rate. It indicates that the EMG pattern recognition system is capable of working on a long-term stability by fusing the online learning method to it.