Pattern Recognition And Motion Analysis Of SEMG

According to the thinking of brain, the intelligent bionic hand can command the bionichand to complete the appropriate function freely. Among the various control signal sources,the surface electromyogram signal (sEMG) as a kind of physiological signals can bedetected directly from human muscle activities, and the process is noninvasive. The controlprocess of sEMG can be more direct and natural. This paper aims to achieve themulti-functional intelligent bionic hand in real-time. To this as the goal, in view of thecharacteristics of sEMG, this paper does the study on the signal processing and controlachieve of the bionic hand.(1)How to auto-complete sEMG activities detection, and also improve its accuracy isan important prerequisite to achieve real-time and effective control of the bionic hand. Inthis paper, activities detection problem can be equivalent to the edge detection problemin image processing. Taking the advantage of the edge maximum posterior probabilityestimates, a threshold set is proposed to improve the Sobel operator. Meanwhile, accordingto the certain similarity with the endpoint detection of speech, to apply different algorithmswhich were used in speech processing to detect the activities of sEMG. The result of thecontrast experiment shows that these algorithms can achieve the sEMG activities detectionautomatically, and in which the improved Sobel algorithm has the best test results.(2)According to the characteristics that the sEMG are very weak and mixed withdifferent noise easily, this paper proposes a new method that uses the unsupervisedKohonen neural network weights optimized to determine the order of the reconstructionmatrix during the process of the noise reduction in singular value decomposition (SVD)effectively. Butterworth band pass and band stop digital filters are selected to abatement thebaseline drift, the affect of the signal instability and50Hz frequency. Then use the SVD todeal with the signal filtered. To make use of the characteristics of the noise platform aregently and centralized of the singular value spectrum of the signal with noise. Through theprojection on longitudinal axis in spectrum, the Kohonen network optimized is applied toconfirm the boundaries of the noise platform, and then to determine the effective order ofthe reconstruction matrix. Simulation results show that this method achieve the noisereduction of sEMG preferably.(3)Due to the analysis of sEMG mechanism of production, the placement of surfaceelectromyography electrodes are determined. Eight kinds of typical gesture modes are selected through the observation of human hand’s functions which can be able to completein daily life. The activities detection and noise reduction algorithms proposed for signalprocessing are applied to these gesture mode signals. According to compare the differentsignal’s characteristic, the AR model coefficients which is representative and cancharacterize signal fully is selected to put into the unsupervised Kohonen network andconduct the recognition experiments. The experimental results further demonstrate theeffectiveness of the proposed algorithm. And four basic pre-shaping modes of hand aredetermined in view of the possible implementation.(4)In order to improve the accuracy rate of sEMG pattern recognition and onlineidentification ability, a weights optimized and supervised Kohonen network and an on-linesemi-supervised Kohonen network are proposed in this paper. The supervised network isadjusted the network structure by adding an output layer, and then optimize the initialweight. According to analysis the lack of the supervised network, on-line semi-supervisednetwork is built. That network screens the pre-labeled samples by multiple classifiersrecognition results and WilsonTh data clip algorithm, and incorporates those into the labeledsamples to update the classifiers. The experiments of four pre-shaping modes recognitionshow that the two modified neural network classifiers have the higher classification ability.(5)The virtual simulation is conducted to validate the results of pattern recognition andthe sEMG control ability. The proposed signal processing, feature extraction and patternrecognition algorithms are embedded into the intelligent bionic hand. The Adams softwareis used to set up a multi-degree-of-freedom bionic hand. And the MATLAB uses the PIDcontrol algorithm to build the bionic hand controller. The virtual co-simulation based onAdams and MATLAB shows that this study is suitable for the bionic hand with a higherbionic degree.