Research On Human Upper Limb Rehabilitation Posture Recognition Method Based On MEMS Dat

Stroke is a prevalent disease-causing death and disability among adults in China,with survivors often experiencing severe upper limb functional impairments,resulting in significant loss to the individual and their family.Traditional rehabilitation treatments rely on professional guidance from rehabilitation centers or one-on-one assistance from therapists,which can be costly for stroke patients who require long-term training.Furthermore,clinical rehabilitation treatments,due to time and spatial constraints,often result in low patient motivation and difficulty in maintaining long-term adherence.However,with the rapid development of intelligent technology,smart home-based rehabilitation offers advantages such as eficient medical resource allocation,low usage costs,and freedom from time and space limitations,making it possible for stroke patients to undergo rehabilitation training at home.To achieve effective home-based rehabilitation,accurate monitoring of upper limb rehabilitation training becomes crucial In this context,the main challenges faced by rehabilitation monitoring technology based on MEMS sensor data are how to safely collect posture data and how to extract sensor data features to accurately recognize upper limb rehabilitation postures.This thesis focuses on the research of upper limb rehabilitation posture recognition methods based on MEMS data from aspects of data collection,posture sample space feature enhancement,and multiscale temporal feature capture of rehabilitation posture data.The main contributions are as follows:1.Addressing the lack of a unified data collection scheme and public datasets in current upper limb rehabilitation posture recognition research,this study,with the advice of professional rehabilitation medical staff,designs and collects MEMS data for five classical upper limb rehabilitation postures.A custom posture dataset,RehabLab-412,is constructed after preprocessing and data feature analysis,ensuring subsequent rehabilitation posture recognition research.2.To address the accurate recognition of upper limb rehabilitation pastures using single-node MEMS data,this study proposes a spatial feature-enhanced deep upper limb rehabilitation posture recognition framework,VGG-Rehab.By tracking the spatial displacement and motion speed of upper limb rehabilitation postures through inertial computation and combining an adaptive threshold zerovelocity correction method to constrain the motion space to the true representation of rehabilitation postures,a feature encoding method is proposed for spatial domain transformation of posture samples.This effectively exploits the abstract feature extraction capabilities and local perception properties of deep convolutional units for accurate rehabilitation posture recognition.In the custom rehabilitation posture dataset RehabLab-412,a recognition accuracy of 97.11%is achieved for five upper limb rehabilitation postures.3.In response to the accurate recognition of upper limb rehabilitation postures in MEMS data under various scenarios,this study proposes a lightweight convolutional neural network model,TMCA-Net.A multi-branch,heterogeneous parallel convolutional network structure,TMCA-Block,is designed to capture multi-scale temporal features in complex upper limb rehabilitation postures.By integrating depth-wise separable convolution and channel attention to enhance key feature extraction,the model ensures accurate upper limb rehabilitation posture recognition while reducing the computational scale and resource usage,improving the practicality and versatility of the model.In both the custom RehabLab-412 upper limb rehabilitation posture dataset and the public UCI Smartphone posture dataset,notable posture recognition accuracy is achieved,providing a feasible solution for upper limb rehabilitation monitoring in various scenarios.