Motion Pattern Recognition Of SEMG Signals Based Upper Limb Self-rehabilitation Training

Hemiplegia usually caused by stroke and other brain diseases, most patients are accompanied with Activities of Daily Living ability (ADL) lost to varying degrees. Robot-assisted upper limb self-rehabilitation training can overcome the shortcomings of clinical rehabilitation, which has important clinical and social significances with some recovery mode for the restoration of patients ’ADL ability. It is a feasible way of self-rehabilitation training that use the patient’s healthy upper limb to guide the rehabilitation robot assisting the hemiplegic upper limb executing rehabilitation training, and the motion tracking on the health side of the upper limb is an important aspect of achieving rehabilitation robot control. Surface electromyography (sEMG) signals can directly reflect the human neuromuscular activity with collection convenient, safe, non-invasive, and able to adapt to the special nature of the physiological state of patients with hemiplegia, which is widely used as a source of rehabilitation robot control. sEMG signals based motion pattern recognition are used to achieve the upper limb motion tracking. The research on principles and methods of sEMG signal based motion pattern recognition as well as improving pattern recognition accuracy, speed and ease of use, portability of rehabilitation system has important value for improving the clinical effectiveness of rehabilitation, clinical safety, and is important to improve the signal pattern recognition as well as bio-mechanical engineering theory to solve practical engineering problems.To better achieve sEMG signals based robot assisted upper limb self-rehabilitation of hemiplegic patients, this article focuses on the sEMG signal multi-order feature space construction method; signal channel reduction; support vector machine classification strategies of upper limb motions and rehabilitation rich environment.Multiclass of upper limb rehabilitation motion pattern and multi-channel sEMG signals are analyzed to study the correlation model between them. According to the nervous system plasticity and motor relearning theory, the self-rehabilitation strategy that let the healthy upper limb to assist the hemiplegic upper limb executing rehabilitation trainings is proposed. By analyzing the human upper limb muscle anatomy and their function distribution, and according to the characteristics of upper limb motion in daily life, six shoulder, elbow-related upper limb rehabilitation training motion pattern are identified, the association model between muscles and multi-channel sEMG signals is established. At the same time, the sEMG signal acquisition position is determined. According to the characteristics of the sEMG signals the sEMG signal acquisition system is constructed, a multi-channel sEMG signal acquisition program is developed to achieve the conversion from the healthy upper limb motion pattern to multi-channel sEMG signal. By analyzing sEMG signal acquisition results, the characteristics and laws of sEMG signals is summarized, the goals and methods of sEMG signal based motion pattern recognition are proposed, the conversion from sEMG signals to the motion pattern of the hemiplegic upper limb is achieved, and then, the self-rehabilitation technology system is constructed.From the non-Gaussian analysis of sEMG signals, the multi-order feature space construction method is studied. The sEMG signals is verified to contain non-Gaussian component by skewness and kurtosis. Taking into account that the non-Gaussian component contains rich information of sEMG signal, which is valuable for improving the accuracy of sEMG signal based upper limb motion classification, for the time domain and frequency domain based lower order statistics joint features commonly used can not characterize the non-Gaussian information of signal, the third-order statistics based bispectrum feature extraction method is studied and the bispectrum slice integral feature is extracted. The Principal Component Analysis (PCA) is applied to reduce the dimension of bispectrum slice integral feature space. The multi-order feature space (BisIE) is constructed by combining the bispectrum feature after dimension reduction with the time domain features integrated EMG value. The experimental study is executed comparing with second-order statistics based power spectrum feature space (MMIE) by upper limb motion classification accuracy as index. The results showed that the feature space BisIE combined by bispectrum and time domain features obtained better upper limb motion classification results, which indicate that the BisIE feature space is an effective and providing method to characterize the signal contains non-Gaussian information.Through multivariate selection theory, the channel reduction method of sEMG signal is studied. The original signal channel correlation and BisIE feature space distribution is analyzed. The results show the presence of redundant and invalid information between each channel, indicating that the channel reduction is feasible. The ReliefF algorithm is improved to calculate the motion pattern classification contribution rates of each channel and sorting, then, got the channel effectiveness sequences. The reverse diminishing method is used to reduce the channel with lowest contribution rate. The separability of the combination of the remaining channels is analyzed, and then, the channel reduction has ultimately accomplished. After the signal channel reduction, the number of channels has been reduced while ensuring a higher classification accuracy, which have a significant impact on the classification speed and the real-time, portability, etc. in post-rehabilitation applications and provide theoretical basis for the signal channel reduction.For the limitation of One Versus One Support Vector Machine (OVO-SVM) multi-classification method, the upper limb multi-classification method of support vector machine is studied and a Two-step SVM (TS-SVM) upper limb multi-pattern motion classification strategy is proposed according to the characteristics of upper limb motion. The (OVO-SVM) multi-classification method is improved according to the characteristics of upper limb motion, the six upper limb movement is divided into two levels and the classification decisions are made by two steps. The first step is to distinguish big class in the first layer, the second step is to make classification decision on the result of classification decision in the first step, and then, produce the final classification result. The improved TS-SVM strategy has solved the overelaborate and real-time application problems of OVO-SVM, reduced the number of classifiers and classification decisions required, improved the time efficiency of the classification and maintained a high classification accuracy, and conduced to enhance the real-time and ease of use in the future clinical application.The rehabilitation rich environment is studied applying rehabilitation medicine theory, virtual reality and upper limb rehabilitation robotics, technology. According to the hierarchy of the disableds ’rehabilitation needs, the patients’ ability of activities of daily living is to train as the main target, a rehabilitation program is developed, upper limb self-rehabilitation training environment is constructed, and two rehabilitation modes are designed. After running test, the rehabilitation training environment constructed can provide good support for the rehabilitation of patients, the myoelectric control method studied is precise and efficient.The sEMG signal pattern recognition techniques for rehabilitation robot control and rehabilitation environments are studied. The results have theoretical and clinical implications for carrying out robot-assisted upper limb self-rehabilitation training. There are still some deficiencies in the following areas which need further research and improvement:(1)Study the relationship between range, speed, as well as more freedom complex motion and sEMG signals in order to achieve more natural robot-assisted upper limb self-training.(2)Expand the subjects sample size, add different ages, genders, and clinical status of the patient as a comprehensive study sample.