Motion Pattern Recognition&Kinematic Trajectory Prediction Of Lower Limb Based On EMG-KJA Neuro-Musculo-Skeletal Model

In recent years, the patients with disabilities as a special social group are attractingincreasingly concerns, and how to help them enhance self-care ability and improve thequality of life effectively is not only an important research topic in the field ofrehabilitation medicine, but also a serious challenge for the government of everycountry and all society. EMG controlled prosthetic limb employ the EMG signals ofremaining limb to voluntarily control the human system to achieve the appropriateaction, which is recognized as an ideal solution to compensate the lost function of limb.Its core technology is the neuro musculoskeletal model that can describe therelationship between neural control and musculoskeletal movement. However, up tonow, the researches about neuro musculoskeletal dynamics at home and abroad arestill mainly focused on the hand and upper limb. There are very few studies aimed atlower limb, especially in motion pattern recognition&kinematic trajectory predictionbased on the neuro musculoskeletal dynamics model of lower limb.In this thesis, by using a wireless sEMG collection system from Noraxon and theVICON infrared motion capturing system, the EMG and3-D kinematics data weresynchronously collected from lower lims during four key movement patterns (squatting,standing up, extending knees and walking) from10volunteers. After signalpreprocessing including amplifying and filtering, the features including autoregressive(AR) model coefficients、cepstrum coefficients、singular value and power spectralentropy were extracted from the sEMG signals during pre-action period to classify thekey movement patterns with Support Vector Machines (SVM), Hidden MarkovModel(HMM)，and Artificial NeuralNetwork (ANN). The results show that the SVM isthe best classifier and during the single-feature-input phase, its average classificationaccuracy rate of singular value is88.3%; the best identification rate could achieved91.3%during the multi-feature-input mode combining the time and frequency features.After the movement pattern recognition, the root mean squared (RMS) and medianfrequency (MF) of sEMG signal from single leg (SL) were calculated. The knee jointkinematic trajectory of another leg (AL) was estimated from the SL feature parameterscombing with the feedback angles from AL with the neural musculoskeletal models oflower limb based on EMG-knee joint angle (EMG-KJA) established by ANN and SVM.The results showed that RMS is better than MF for movement prediction; by contrasting the two modeling methods, the ideal prediction way is SVM whose error ismuch lower thanthat of ANN. In order to inprove the degree of real-time and on-line ofprediction, a set of muscle-source channel optimization program was designed for eachaction by removing the unrelated channel one-by-one according to the correlationcoefficient between RMS of sEMG signals and knee joint angle. Tthe predictionresultswere tested to find out the optimal prediction method.The research results in this thesis are not only conductive for the future design ofEMG controlled lower limb prosthesis and the exploration on relevant neurologicaland kinematic mechanism, but also could be employed in motion sensing,human-machine interface, virtual reality, remote control and other related fields.