Research On Myoelectric Control Of Upper Limb Rehabilitation Robot

The sEMG signal is the comprehensive overlay of the electrical activity in the nervous system at the skin surface in space and time. As it is easy to acquire and has no trauma to the body, it is a research focus to realize the upper limb rehabilitation robot myoelectric control utilizing sEMG signal in the field of rehabilitation engineering. The principle of myoelectric control based on pattern recognition is to extract the features that can characterize different motion patterns from muti-channel sEMG signals, and decode the features to the corresponding robot motion control instructions which can drive the upper limb rehabilitation robot and then put the hemiplegic side upper limb into training through the feature pattern classifier. The evaluation indexes of myoelectric control mainly include recognition accuracy and decoding speed.In this paper, the myoelectric control of upper limb rehabilitation robot based on pattern recognition is studied in order to drive the upper limb rehabilitation robot quickly and accurately utilizing the four channel sEMG signals of healthy side upper limb of hemiplegia patients. Problems need to be studied including:(1) the spatial decoupling pretreatment of multi-channel sEMG signals;(2) the features extraction of sEMG signals to characterize different upper limb movement patterns;(3) the classification of upper limb movement patterns.According to sEMG’s weak and vulnerable to interference characters, the paper specifically designed the sEMG acquisition scheme in order to obtain the reliable sEMG signals. On the basis of analyzing the correspondence relation between upper limb movement and muscle, the four-channel sEMG signals acquisition position were placed on the central of the brachioradialis muscle, biceps, triceps and deltoid on subject’s left upper limb, and we also placed the measuring electrodes along the direction of fibers on the muscle belly and reference electrodes on the tendons where the sEMG signals are faint, which reduced the randomness of the signal acquisition and provided the steady control source for the myoelectric control of upper limb rehabilitation robot. In order to eliminate the spatial coupling redundant information among the four-channel sEMG signals, the paper proposed the multi-channel sEMG signals spatial decoupling pretreatment method based on independent component analysis (ICA), and divided the four-channel sEMG signals into four different independent components. For the shortcoming that the order of the independent component after multiple ICA decomposition is not fixed, the paper proposed the matching method between the order of the independent component and the multi-channel sEMG signals and established the clear corresponding relation between the independent component and four-channel sEMG signals based on the correlation before and after ICA decomposition, which made it possible for the independent component to replace the original sEMG signals to become the myoelectric control source. The subsequent result of feature extraction shows that the spatial decoupling makes the feature space theoretically separability decreased slightly, but it significantly increased the speed of feature extraction and reduced the amount of calculation, which shows it is effective for the ICA multi-channel sEMG signals spatial decoupling pretreatment.The sample entropy can be used as a characterization of differernt upper limb patterns of sEMG signal feature. But it involves a large amount of calculation. For this limitation, the method by combining wavelet packet transform with sample entropy was studied to extract the sEMG signal feature in this paper. Extract sample entropy from nine subspaces after five layers wavelet packet decomposition as the features of the motion pattern. For the issue that the feature space dimension is too large, an improved wavelet packet decomposition algorithm was proposed, which made a high-frequency subspace quadratic wavelet transform based on the main wavelet transform, and then extracted sample entropy from four sub-spaces features and constructed joint multi-domain features with integrated EMG values that is TDISaEn. Compared to the feature constructed by the common wavelet packet energy (TDIEner), the results based on the CSI and BP neural network test results show that TDISaEn has a higher separability and recognition accruracy than TDIEner.The network convergence speed is slow and the training time is long using the BP neural network to classify the upper limb motion pattern. In order to overcome this defect, the genetic BP neural network was proposed, which optimize the BP neural network’s connection weights and thresholds with genetic algorithm. Then, get the optimal solution after inheritance for training the BP neural network. The results show that after genetic algorithm optimization, the BP neural network recognition accuracy is not high, but the training time was greatly shortened.The myoelectric control based on pattern recognition was tested utilizing the upper limb rehabilitation robot self-developed. Then integrate the trained genetic BP neural network model into the robot control system and transformed the identified the target movment pattern into the robot motion control commands. The result shows that the upper limb rehabilitation robot performed well and verified the accuracy and stability of myoelectric control.The paper combines sEMG processing and upper limb rehabilitation robot, and the research results are helpful to improve the performance of the upper limb rehabilitation robot. However, there are still some inadequate, and two following problems need to research and improve:(1) The related laws between movement amplitude, speed and sEMG signal;(2) The selection range of subjects sample needs to be expanded, and some clinical factors such as age, gender, degree of hemiplegia should be considered.