Research On Feature Learning And Pattern Recognition Algorithms For Motor Imagery EEG Signal

Brain-computer interfaces immediately connect the human brain to other artificial intelligence devices.It has been a research hotspot in recent years and is in its infancy.Classification and recognition of EEG motor imaging is one of the cores of brain-computer interface research.However,EEG is susceptible to interference,unstable signals,and large differences between inter-individual,the accuracy of recognition accuracy is not high currently.we study how to reduce the processing time based on the law of physiological signals.After finding an effective segmentation method for the input data,we adopt the custom learning network to learn the MI features from the data,then effectively and accurately identify MI multitasking across time domains and across objects.The main work and innovations of this paper include the following three aspects:Ⅰ.Aiming at the problem that the current machine recognition of EEG topology maps is not good enough to assist in medical rectification,a new EEG energy topology map(ETR)algorithm and ETRCNN model supporting multi-object and multi-task recognition are proposed.The ETR algorithm designs a module of parameterized automatically generated energy topology,which can obtain quantitative and measurable information features to objectively evaluate these physiological signals.In ETRCNN,we embed the ETRJD algorithm,and adopt multiple ETR joint discrimination under the same event to improve the classification and recognition effect through global calibration.The experimental verification have proved that ETRCNN’s classification and discrimination algorithm is 10.11%higher than the classification accuracy of international cutting-edge methods on the same data set.It is also helpful to research EEG traceability which can find out the active regions of brain,carry out spatial positioning and quickly obtain source information.Ⅱ.If the tester samples are not trained in the constructed machine learning model,the classification ability of the model is poor.To against this problem,we proposed a dual-input and dual-attention mechanism network(DHDANet)for motor imagery deep learning recognition.The experimental verification is employed with a dataset of 54 subjects and 21,600 EEG motor imagery collected by Korea University,the model is 2.08%higher than the current international cutting-edge methods.We also performed that DHDANet has a high recall rate and specificity.The recognition rate for poor sample quality is significantly higher than other methods,and the robustness is better.Ⅲ.In view of the fact that the recognition ability is poor in classification and recognition of small samples of EEG motion imagery,the proposed FBCapsNet model integrates the feature learning of DHDANet with the main capsule layer of the capsule network.In addition,the FBCapsNet model introduces a dual classification and recognition mechanism,which first learns the vector weights of feature attributes in different rhythm segments through the capsule network,and then classifies them according to the vector weights of each rhythm segment through Soft Attention.Through experimental verification on the BCI Competition Ⅳ 2b data set containing 9 subjects and 6480 EEG motor imagery,the model has a 2.41%higher classification accuracy than the current international cutting-edge methods.The FBCapsNet model provides a new method for machine learning classification and recognition.