Dynamic Motor Imagery Recognition Based On EEG

Neuromotor impairment,such as stroke,can cause obstruction of neuromuscular connection.At present,the treatment methods for paralysis can be divided into physical therapy to reconstruct brain muscle connection and brain-computer interface(BCI)to construct brain and electromechanical device.Physical therapy requires a lot of manpower,and about 80% of patients with upper limb paralysis cannot benefit from it.BCI can close the circuit between motor cortex activity and limb mechanical movement through motor image(MI),so as to stimulate the connection between cortex and muscle to some extent.Recent studies on physical therapy have shown that compared with MI,dynamic motor imaging(d MI)has better physiotherapy effect.On the basis of MI,d MI requires the performer to imitate the MI action with the help of assistance,so it can close the circuit between motor cortex and limb muscle movement,directly stimulate and reshape the brain muscle nerve connection.BCI combined with d MI can reduce the cost of physical therapy manpower and improve the effect of physical therapy.In this paper,the BCI scheme based on d MI is formed through scalp electroencephalogram(EEG)acquisition,functional brain network(FBN)construction,feature extraction and pattern recognition.The main research contents and innovations are as follows:(1)A new binarization method for FBN is proposed.Traditional FBN binarization methods,which use empirical threshold to binarize and minimum spanning tree to binarize,are reliable analysis methods in EEG analysis field.Considering two different directions of binarization,the two features of binarization can be fused to enhance the difference between classes of input features.(2)An ELM variant which can balance local features is proposed.Feature fusion occurs when multiple local features are mapped randomly in the feature space layer of ELM.However,due to the random distribution of the mapping basis vector,the fusion proportion of local features in high-dimensional space is unbalanced.In the ideal fusion process,the proportion of local features with large differences is expected to be larger.Therefore,this paper adds fusion variables to ELM feature space layer to improve the fusion effect of ELM.(3)A fast search algorithm with calculation optimization is proposed.In order to find the optimal parameters,this paper uses the leave-one-out cross validation(LOO)optimization algorithm.Singular value decomposition is used to extract the predicted items in the LOO optimization algorithm,which can avoid repeated calculation in grid search,so as to reduce the computational complexity of the original LOO optimization algorithm.