Research On Recognition Of Upper Limb Rehabilitation Exercises After Stroke Based On Convolutional Neural Network

Stroke is the leading cause of death and disability among adults in China,which places a great burden on patients,families and society.More than 80% of stroke survivors have motor dysfunction,of which upper limb motor dysfunction is the most common and serious.Traditional rehabilitation exercise is conducted by physical therapists with manual guidance and auxiliary exercise,which has problems such as high manpower consumption and low efficiency.Interactive rehabilitation exercise can mobilize the enthusiasm of patients,improve patient compliance and reduce the burden of medical staff,in which rehabilitation exercise action recognition is very important for interactive rehabilitation exercise.How to effectively extract action characteristics,improve the accuracy of action recognition is facing challenges.In this thesis,the study on the action recognition of stroke upper limb rehabilitation exercise based on convolutional neural network is carried out,from the aspects of multichannel surface electromyography(s EMG),array s EMG and multimodal data fusion.Deep learning methods are proposed for feature extraction and action recognition,and a stroke rehabilitation exercise platform is designed and implemented.The main work of the thesis is as follows:1.For the multi-channel s EMG rehabilitation exercise action recognition,a convolutional recurrent neural network model fused with attention is proposed.The input signal is obtained by blind source separation through non-negative matrix factorization,and the spatiotemporal features of the signal are extracted by using the convolution recurrent neural network to fuse attention.The model automatically assigns feature weights to implement upper limb rehabilitation exercise action recognition.The accuracy on the two sub-sets of Ninapro reach 86.5% and 86.2%,respectively.2.In Order to recognition of hand fine rehabilitation exercise with multi-muscle coordination,a parallel convolutional neural network model based on array s EMG is proposed,and the array s EMG data is reconstructed into multi-frame EMG images and multi-channel EMG sequences.Parallel convolutional neural network and parallel long short-term memory network are used to extract muscle physical space coordination characteristics and activity time-varying characteristics respectively to implement upper limb rehabilitation exercise action recognition,with an accuracy of 96.4%.3.For the recognition of upper limb rehabilitation exercise under the change of muscle strength and limb position,a two-way multi-size convolutional neural network model based on multi-modal data fusion is proposed,which adaptively extracts features of s EMG and acceleration data,and the multi-size convolution block is introduced to increase the model capacity and improve the fitting ability.The recognition of upper limb rehabilitation exercise is implemented with an accuracy of 96.7%.4.Based on technical frameworks such as Spring Boot,Vue.js and My Batis,using Java language and My Sql database,the stroke rehabilitation exercise platform of browser/server architecture is designed and implemented,which has the functions of user management,patient management,exercise program management,exercise record management,exercise template management,exercise project management,system management and other functions to implement the process and information management of stroke clinical rehabilitation exercise.