| As the country with the highest incidence of stroke in the world,one person has a stroke every 12 seconds in China,and the number one cause of death and disability among Chinese people is also stroke.Although the prevention and treatment of stroke in China has been fully implemented and is beginning to bear fruit,the rehabilitation system for stroke patients is still in need of improvement.With the rapid development of information technology,more and more emerging technologies such as Internet of Things and machine learning have been applied to the field of intelligent medicine,opening new space for medical data analysis and processing.In the process of functional recovery of stroke patients,functional electrical stimulation system devices with MEMS sensors are required to be worn to assist in training.In this paper,a series of research works have been conducted around the problems of how to improve the accuracy of raw sensor data,how to avoid the problems of pre-calibration and fixed position placement of sensors,and the recognition of human rehabilitation posture using sensor data:(1)In order to guarantee the accuracy and robustness of the raw data of MEMS sensors and not to affect the accuracy of rehabilitation pose recognition due to noise and other factors,pre-processing of the raw data is required.In this context,this paper starts from the sensor three-axis orthogonal calibration to guarantee the sensor sensitivity;cleans and de-weights the raw data using the quadrature method;and establishes a Kalman filter-based data gain compensation model to obtain stable and noise-free data.Finally,the effectiveness of the proposed data pre-processing method is verified by experiments.(2)In the study of human posture tracking based on sensor data,the need to keep the placement of MEMS sensors fixed is always a problem in the application,and the collected data will generate large errors and drift when the placement is not precise enough and the elastic deformation of the skin occurs.In this context,this paper designs an adaptive angle tracking method based on the upper limb to avoid the problem of precise sensor placement by converting acceleration data into body joint angle variation data,and to avoid data drift due to high signal-to-noise ratio by segmented data processing.Finally,the feasibility and application value of the method are verified by experiments.(3)In this paper,a fully connected neural network model is designed with the goal of upper limb rehabilitation posture recognition.Since the collected rehabilitation data is small sample data,it has a large impact on the training accuracy of the neural network model.The sample data size is increased by 18 times by adding windows to the data.In order to verify the recognition effects of different parameters,the effects of different hidden layer layers,activation functions,loss functions and optimizers on the recognition rate were compared by adjusting the data window size and step size.Finally,a 2-layer hidden layer,Softsign activation function,Categorical Crossentropy loss function and Adamax optimizer with better recognition performance are selected.Finally,the classification effect shown by the confusion matrix was used to synthesize the recognition accuracy and time efficiency of various algorithms to verify the superiority of fully connected neural networks in upper limb rehabilitation posture recognition. |
PDF Full Text Request